Physics-consistent constraint probabilistic modeling and prospective risk assessment method for intelligent decision-making in flight test points Authors: Tianchang Liu, Chen Zhao, Ling Pu Published: 2026 ABSTRACT Addressing the real-time monitoring challenges posed by high dynamics, strong nonlinearity, and observational uncertainty in flight test envelope boundary missions, traditional threshold methods and pure data-driven models struggle to balance identification accuracy, physical consistency, and predictive warning capability. To this end, this paper proposes a CPSSMF aimed at providing interpretable and calibratable unified probabilistic evidence for safety-critical decision-making. Based on sequential Bayesian inference, the framework achieves closed-loop monitoring through three coupled modules: (1) PSA module, which injects physical constraints into posterior distribution shaping through endogenous potential functions, combined with output-side safety masking and multi-source anomaly detection to suppress non-physical state jumps; (2) HPRF module, which performs multi-step rolling prediction based on corrected posteriors, outputting trend risks within a fixed time horizon to achieve quantitative trade-off between lead time and false alarm rate; (3) EGUAQ module, which dynamically evaluates data value using posterior entropy to support active verification. Evaluation results based on simulated flight stall test data demonstrate that CPSSMF maintains high phase identification accuracy (Acc 0.969, F1 0.785) while achieving an average warning lead time of approximately 3.34 s under the same false alarm constraints, significantly outperforming baseline methods. Furthermore, under sensor contradiction injection and out-of-distribution disturbance scenarios, the method exhibits excellent robustness and stability. This study establishes an interpretable analysis chain integrating situational awareness, risk quantification, and anomaly diagnosis, effectively enhancing the engineering applicability of flight test safety monitoring. Introduction Flight testing runs throughout the entire process of aircraft development verification, airworthiness certification, and performance and stability quality assessment, serving as a key source of evidence connecting design assumptions, simulation inferences, and real flight environments1,2. Compared to conventional airline operations, test points are often located near envelope boundaries and accompanied by high dynamic maneuvers and strong nonlinear coupling, presenting typical characteristics of short test windows, long operational chains, strong handling constraints, and rigid risk boundaries3,4, thus imposing safety-critical requirements on online monitoring and decision support5. As flight test missions gradually evolve from pure verification to an emphasis on both verification and exploration, test organization methods and risk management paradigms are evolving synchronously6,7. Relying solely on planned execution and post-event review can no longer cover the real-time situational awareness and rapid response requirements during high-intensity test point execution8,9. Engineering practice requires systems to reliably identify the current phase of test points and provide quantitative characterization of short-term future risk trends within sampling-level time scales based on multi-channel measurements, supporting operational decisions such as advancing, maintaining, repeating, retreating, or terminating10–12. The core of such requirements is not merely pursuing marginal improvements in phase classification accuracy, but demanding that online outputs simultaneously possess trustworthiness, temporal continuity, interpretability, and calibrability13,14. This involves suppressing high-frequency oscillations in phase identification that cause alarm jitter and handling interference, while avoiding overconfident erroneous suggestions when sensor anomalies15, channel conflicts, and out-of-distribution conditions occur. The system should present the cost trade-offs among lead time, false alarms, and missed detections in a reviewable manner, enabling threshold calibration, risk budget al.location, and hierarchical strategies for test point types to have consistent evidence basis and traceable analysis chains16–20. Regarding real-time monitoring and warning around test points, related research and engineering practices have broadly formed three types of technical routes, each with their respective boundary conditions21–23. The first is threshold and rule-driven methods, with typical implementations including key parameter thresholds, combination criteria, envelope protection logic, and control law protection24. The advantages of this type of method lie in transparent mechanisms, strong interpretability, ease of review and certification, and easy integration into existing test procedures, thus maintaining a dominant position for a long time25. However, under high dynamic and strong noise conditions, threshold triggering often exhibits lag and oscillation, is more sensitive to single-channel anomalies, and discrete boundary triggering in state transition regions amplifies phase distortion, making it difficult to naturally form trend risk quantification for fixed prediction horizons, and even more difficult to systematically provide disposal lead time26,27. The second is data-driven time series modeling methods, such as models based on recurrent networks, temporal convolution, and attention mechanisms and their variants, which can learn complex correlation structures from multi-channel time series and typically achieve high performance in offline identification or relatively stable distribution conditions28–31, but still face consistency and credibility issues in safety-critical high dynamic flight missions32. One issue is the lack of explicit physical consistency and process constraints, making it easy to learn statistically correlated but physically infeasible mappings, producing mechanistically distorted outputs under boundary conditions33. The second is insufficient robustness and uncertainty characterization for out-of-distribution inputs, faults, and drift, potentially leading to overconfidence and passing erroneous judgments to the disposal layer with high confidence34,35. The third is unclear analysis paths, making it difficult to provide analyzable explanations of trigger causes, evidence sources, and confidence levels for effectiveness analysis and safety management chains, thereby constraining engineering implementation and continuous iteration36. The third is probabilistic graphical models and Bayesian filtering frameworks, such as hidden Markov models, switching state space models, and particle filters. This type of method naturally fits online sequential inference and can express uncertainty with probabilistic semantics, with inference chains that are more traceable and convenient for collaboration with decision rules37,38. However, if relying solely on statistical emission and transition while lacking task-relevant physical feasibility constraints, posterior probabilities may accumulate mass on infeasible interpretations and produce multi-modal competition and phase jumps, thereby weakening alarm stability and risk prediction reliability39. If empirical correction is further overlaid at the output end, it easily forms frequent pullbacks and temporal discontinuities, making it difficult for the system to provide verifiable risk evidence and affecting human-machine trust maintenance. Overall, the key bottleneck does not come from insufficient model stacking, but from the lack of a methodological framework that can unify phase identification, credible gating, prospective risk, and data closed-loop within the same online probabilistic semantics and analysis chain, enabling the system to output stable, interpretable, and calibratable structured situational results under conditions of enhanced noise, conflicting observations, and out-of-distribution perturbations40–42. Meanwhile, recent advances in graph-based domain adaptation techniques offer a highly attractive avenue for further reducing the need for target domain data, for instance, hierarchical graph structures that capture local-to-global topological relationships have demonstrated strong capabilities in semi-supervised domain adaptation43. Furthermore, leveraging Graph Convolutional Networks to drive multi-layer perceptron generalization via cross-domain knowledge distillation44 provides a robust framework for transferring complex, multi-channel sensory correlations from a well-tested aircraft platform to a newly developed prototype. Based on the aforementioned background, this paper targets safety-critical and process-constrained test point scenarios, with the research problem being to construct a unified online framework that can stably estimate test point phases at sampling-level time scales while synchronously expressing uncertainty, output calibrated trend risks and quantifiable disposal lead times within future fixed time domains, and provide interpretable anomaly evidence and trustworthy gating strategies when observational contradictions, sensor drift, or out-of-distribution disturbances occur, while simultaneously driving uncertainty-based data value assessment and verification acquisition to form a safe and controllable closed-loop improvement path. The main innovations of this paper are: (1) Unified probabilistic analysis chain output: centered on stage posterior distributions, unifying stage identification, uncertainty expression, risk quantification, and anomaly diagnosis, avoiding stage distortion and semantic fragmentation caused by hard decisions, supporting analyzable and traceable online decision-making in safety-critical scenarios; (2) Online inference and interpretable analysis with physical consistency constraints: injecting physical and feasibility constraints in an endogenous manner into sequential inference, combined with output-end safety shielding to form interpretable corrective evidence, improving temporal consistency and robustness against channel contradictions and out-of-distribution perturbations; (3) Calibratable forward-looking risk analysis: proposing fixed-horizon windowed trend risk assessment to achieve systematic characterization and engineering calibration of lead-time versus false alarm/missed detection trade-offs; simultaneously driving active review and data acquisition through uncertainty indicators such as entropy, constructing an executable continuous iterative closed loop under safety gating. Model and methods Overall model: constrained probability state space intelligent monitoring framework Constrained Probability State Space Monitoring Framework CPSSMF for Online Intelligent Decision-Making of Flight Test Points. The overall objective is to employ sequential Bayesian posterior as a unified semantic carrier under envelope boundaries and high dynamic conditions to form analyzable probabilistic evidence supporting safety-critical handling. The framework integrates three coupled functionalities sharing the same posterior representation and collaborating within the same inference pipeline. First, PSA incorporates physical consistency constraints and multi-source anomaly evidence into trusted gating, implementing robust corrections to the posterior and discrete outputs through constrained potential functions and feasible domain projection, thereby suppressing the probability mass accumulation of infeasible states and reducing high-confidence misjudgments caused by observational conflicts and out-of-distribution disturbances. Second, HPRF uses multi-step rolling propagation to map current uncertainties into prospective trend risks within a fixed prediction time horizon of 3 to 8 s, achieving calibrated probabilistic risk quantification and lead characterization. Third, EGUAQ measures posterior sharpness and cognitive uncertainty using information entropy, generating reevaluation and acquisition priorities under PSA gating constraints to enhance data utilization efficiency and support closed-loop iteration. The system input consists of the observation vector x_k∈R^d at time k (comprising multi-channel information such as angle of attack, pitch angle, calibrated airspeed, normal acceleration, control inputs, derived features, etc.), model parameters, and sets of rule thresholds. The system output is the situational vector Y_k, which includes stage posterior distributions, discrete stage conclusions, fixed time-domain risk indicators, anomaly diagnosis evidence, uncertainty measures, and handling suggestions. This facilitates engineering processes such as reevaluation, calibration, and analysis with regard to decision-making bases, trigger reasons, confidence levels, and predictive lead times. The framework is demonstrated in Fig. 1. Fig. 1System architecture diagram. System architecture diagram. In probabilistic modeling, this study abstracts the experimental point process as a state space generative process of discrete hidden state sequences \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{z}_{k}\in\:\mathcal{S}$$\end{document} and continuous observation sequences \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{k}$$\end{document}, employs Gaussian hidden Markov models to characterize phase transition dynamics and conditional observation distributions, and implements online posterior recursion and approximate inference through particle filtering, thereby obtaining explicit representation of phase uncertainty under real-time computational constraints. To address the physically inconsistent posterior mass accumulation problem caused by the contradiction between high dynamic maneuvering and multi-sensor channels, this study injects physical consistency constraints into the inference loop in an endogenous manner: through constraint potential functions to perform probabilistic shaping and feasible domain biasing of the posterior, making states that do not satisfy necessary physical conditions receive consistent probabilistic penalties, mechanistically suppressing phase jumps and high-confidence misjudgments, avoiding frequent corrections and unattributable output fluctuations caused by relying solely on output-side rule-based corrections. The overall generative and inference process can be given by a set of compact formal expressions, as shown in Eq. 1.1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{l}\begin{array}{l}\left\{\begin{array}{ll}&\:{\mathrm{z}}_{1}\sim\:\mathrm{Cat}\left({\uppi\:}\right),\:\:{\mathrm{z}}_{\mathrm{k}}\mid\:{\mathrm{z}}_{\mathrm{k}-1}\sim\:\mathrm{Cat}\left({\mathrm{A}}_{{\mathrm{z}}_{\mathrm{k}-1},:}\right)\\\:&\:{\mathrm{x}}_{\mathrm{k}}\mid\:{\mathrm{z}}_{\mathrm{k}}=\mathrm{s}\sim\:\mathcal{N}\left({{\upmu\:}}_{\mathrm{s}},{{\Sigma\:}}_{\mathrm{s}}\right),\\\:&\:{\uppsi\:}\left(\mathrm{s},{\mathrm{x}}_{\mathrm{k}}\right)\in\:\left(\mathrm{0,1}\right]\hspace{0.25em}\\\:&\:\mathrm{p}\left({\mathrm{z}}_{\mathrm{k}}\mid\:{\mathrm{x}}_{1:\mathrm{k}}\right)=\frac{1}{{{\upeta\:}}_{\mathrm{k}}}\left[{\uppsi\:}\left({\mathrm{z}}_{\mathrm{k}},{\mathrm{x}}_{\mathrm{k}}\right)\hspace{0.17em}\mathrm{p}\left({\mathrm{x}}_{\mathrm{k}}\mid\:{\mathrm{z}}_{\mathrm{k}}\right)\sum\limits_{{\mathrm{z}}_{\mathrm{k}-1}\in\:\mathcal{S}}\mathrm{p}\left({\mathrm{z}}_{\mathrm{k}}\mid\:{\mathrm{z}}_{\mathrm{k}-1}\right)\hspace{0.17em}\mathrm{p}\left({\mathrm{z}}_{\mathrm{k}-1}\mid\:{\mathrm{x}}_{1:\mathrm{k}-1}\right)\right]\end{array}\right.\end{array}\end{array}$$\end{document} Where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\pi\:$$\end{document} represents the initial distribution, A is the state transition matrix, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left({\mu\:}_{s},{\varSigma\:}_{s}\right)$$\end{document} are the Gaussian emission parameters for each stage, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\eta\:}_{k}$$\end{document} is the normalization constant. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\psi\:$$\end{document} is the physical potential function, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:p\left({x}_{k}\mid\:{z}_{k}\right)$$\end{document} is the observation likelihood. By incorporating necessary physical conditions as potential functions into Bayesian recursion, this is equivalent to imposing controllable prior-type suppression terms on infeasible solutions, causing the posterior mass to move away from infeasible regions at the inference source, mechanistically improving temporal consistency and feasible domain consistency. This property is particularly critical for subsequent forward-looking risk assessment based on posterior rolling extrapolation. Compared to analytical filtering, particle methods are more easily integrated with GPU tensorization and convenient for extension to more complex emission distributions, mixture models, or nonlinear observation mappings, thus balancing engineering deployment with model scalability. In unified output interface design, adopts the principle of stage posterior driving all sub-modules to avoid semantic inconsistency and non-attributable contributions caused by the mutual independence of stage identification, alarm logic, and anomaly detection. The system outputs structured situation vectors at each moment as shown in Eq. 2.2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{ll}{Y}_{k}=\left({p}_{k},{\widehat{z}}_{k},{z}_{k}^{\star\:},{R}_{k},{H}_{k},{A}_{k},{\mathrm{A}\mathrm{l}\mathrm{e}\mathrm{r}\mathrm{t}}_{k},{\mathrm{A}\mathrm{d}\mathrm{v}\mathrm{i}\mathrm{c}\mathrm{e}}_{k}\right)\end{array}$$\end{document} Where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{p}_{k}=p\left({z}_{k}\mid\:{x}_{1:k}\right)$$\end{document} is the posterior distribution of the stage, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\widehat{z}}_{k}=\mathrm{a}\mathrm{r}\mathrm{g}{\mathrm{m}\mathrm{a}\mathrm{x}}_{s\in\:\mathcal{S}}{p}_{k}\left(s\right)$$\end{document} is the unconstrained most likely stage, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{z}_{k}^{\star\:}$$\end{document} is the stage conclusion after constraint projection by PSA’s Safety Shield; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{R}_{k}$$\end{document} is the fixed-horizon forward-looking trend risk indicator output by HPRF, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{H}_{k}$$\end{document} is the posterior entropy uncertainty measure aggregated by EGUAQ, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{A}_{k}$$\end{document} is the anomaly diagnosis evidence provided by PSA, and finally \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{A}\mathrm{l}\mathrm{e}\mathrm{r}\mathrm{t}}_{k}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{A}\mathrm{d}\mathrm{v}\mathrm{i}\mathrm{c}\mathrm{e}}_{k}$$\end{document} are generated through a safety-first fusion strategy. The key advantage of this interface lies in that \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{p}_{k}$$\end{document} serves as a unified semantic foundation making risk prediction, trusted gating, and active review accountable for the same uncertainty source, while correction results and anomaly evidence are explicitly incorporated into the output structure, enabling the system to possess interpretability and traceability in engineering effectiveness analysis, threshold calibration, and safety demonstration, avoiding the difficulty of explaining triggering mechanisms and applicable domain boundaries when only outputting single-value labels or black-box scores. The three sub-modules form gated coupling through a unified probabilistic interface rather than parallel stacking. PSA is responsible for reliability assurance and physical consistency safeguarding, injecting necessary physical conditions endogenously into sequential inference, shaping the posterior through constrained potential functions, and implementing feasible domain projection and safety masking at the output end; at the same time, it integrates residual deviations, low likelihoods, and abnormal evidence such as rule sustained trigger rates to delineate the trustworthy operating domain, explicitly exposing mismatches and triggering conservative gating when observational conflicts, drifts, or out-of-distribution perturbations occur, thereby suppressing high-confidence misjudgment under low-reliability conditions. HPRF further maps posterior uncertainty into trend risks within predetermined 3 to 8-second prediction windows, utilizes transition structures for multi-step rolling propagation, constructs probability mass metrics for concerned events, elevating risk assessment from instantaneous threshold detection to calibrated forward-looking evaluation, and supports lead time and false alarm/miss calibration trade-offs. EGUAQ measures uncertainty using information entropy, generating review and acquisition priorities to enhance data efficiency; it automatically suppresses exploration and prioritizes safety handling when risks escalate or anomalies become prominent. The framework adheres to the principles of training-deployment isolation and reproducibility, performing only recursive and gated operations online, adopting conservative updates for missing information and recording abnormal evidence, ensuring stable and analyzable output. Submodel A: physical consistency safety analysis In formal representation, PSA can be viewed as introducing feasibility constraint potential functions in HMM recursion and combining normalization with configurable penalty constraint posterior update mechanisms. Let the discrete hidden state set be \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{S}$$\end{document}, observation vector be \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{k}\in\:{\mathbb{R}}^{d}$$\end{document}, physical feasible set be \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{F}\subseteq\:\mathcal{S}\times\:{\mathbb{R}}^{d}$$\end{document}, which consists of necessary criteria including envelope boundaries, phase entry and exit conditions, channel consistency, and process constraints, and is represented by constraint families \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{g}_{m}\left(z,x\right)\le\:0,\hspace{0.33em}m=1,\dots\:,M$$\end{document}, then the unnormalized posterior is shown in Eq. 3, and the normalized posterior is obtained through Eq. 4, where the potential function is defined as shown in Eq. 5.3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{c}{\stackrel{\sim}{p}}_{k}\left(z\right)=\left[p\left({x}_{k}\mid\:z\right)\hspace{0.17em}\psi\:\left(z,{x}_{k}\right)\right]\sum\limits_{{z}^{{\prime\:}}\in\:\mathcal{S}}{A}_{{z}^{{\prime\:}},z}\hspace{0.17em}{p}_{k-1}\left({z}^{{\prime\:}}\right)\end{array}$$\end{document}4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{c}{p}_{k}\left(z\right)=\frac{{\stackrel{\sim}{p}}_{k}\left(z\right)}{\sum\:_{s\in\:\mathcal{S}}{\stackrel{\sim}{p}}_{k}\left(s\right)}\end{array}$$\end{document}5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{c}\psi\:\left(z,{x}_{k}\right)=\prod\limits_{m=1}^{M}\mathrm{e}\mathrm{x}\mathrm{p}\left(-{\lambda\:}_{m}\hspace{0.17em}{\left[{g}_{m}\left(z,{x}_{k}\right)\right]}_{+}\right),\:\:{\left[u\right]}_{+}=max\left(u,0\right)\end{array}$$\end{document} Among them, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\lambda\:}_{m}$$\end{document} adjusts the penalty intensity of soft constraints. In our implementation, the constraint set \:F explicitly encompasses three core physical rules reflecting aerodynamic and flight mechanics boundaries: (1) Global Sensor Integrity Constraint (Hard Constraint): \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{g}_{1}\left({x}_{k}\right)=\mathbb{I}(\mathrm{AoA}>25.0\wedge\:\mathrm{Pitch}<5.0\wedge\:{N}_{z}<1.5)$$\end{document}. This rule identifies sensor malfunctions (e.g., stuck AoA vanes) independent of the state. If triggered, the state is forced back to ‘Trim’, overriding any data-driven HMM output. (2) Kinematic Energy Constraint (Soft Constraint): \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{g}_{2}(z,{x}_{k})=\mathbb{I}(z=\mathrm{'Deep\_Stall'}\wedge\:\mathrm{Pitch}<10.0\wedge\:\mathrm{AoA}<20.0)$$\end{document}. A state cannot be ‘Deep Stall’ if the aircraft is at a low angle of attack and low pitch, which violates energy conservation. The penalty parameter \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\lambda\:}_{2}$$\end{document} is set to a high value (e.g., \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{10}^{6}$$\end{document}) to effectively zero out the probability mass for this state. (3) Aerodynamic Recovery Logic (Soft Constraint):\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{g}_{3}(z,{x}_{k})=\mathbb{I}(z\in\:\{\mathrm{'Full\_Stall'},\mathrm{'Deep\_Stall'}\}\wedge\:\mathrm{Pitch}<-10.0\wedge\:{N}_{z}>0.8)$$\end{document}. If the pitch is significantly nose-down with a recovering load factor, the aircraft is physically in a ‘Recovery’ or ‘Dive’ phase, contradicting a sustained stall state. These rules parameterize the logical safety boundaries. The sensitivity to constraint design lies in the scaling of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\lambda\:}_{m}$$\end{document}; excessively low values fail to correct erroneous HMM inferences during sensor noise, while excessively high values (acting as hard constraints) may cause state chattering if the thresholds (e.g., Pitch < −10.0) are set too rigidly near transition boundaries. By integrating these rules into the potential function, the inference seamlessly balances statistical data fitting with strict aerodynamic principles. The unconstrained most likely stage is obtained from the posterior distribution, as shown in Eq. 6. Subsequently, the corrected stage is obtained through constrained discrete choice, as shown in Eq. 7.6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{c}{\widehat{z}}_{k}=arg\underset{z\in\:\mathcal{S}}{\mathrm{m}\mathrm{a}\mathrm{x}}{p}_{k}\left(z\right)\end{array}$$\end{document}7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{c}{z}_{k}^{\star\:}={\varPi\:}_{\mathcal{F}}\left({\widehat{z}}_{k},{x}_{k}\right)=arg\underset{z\in\:\mathcal{S}}{\mathrm{m}\mathrm{i}\mathrm{n}}\left(d\left(z,{\widehat{z}}_{k}\right)+\sum\limits_{m=1}^{M}{w}_{m}\hspace{0.17em}\mathbb{I}\left[{g}_{m}\left(z,{x}_{k}\right)>0\right]\right)\end{array}$$\end{document} Where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathbb{I}\left(\cdot\:\right)$$\end{document} is an indicator function that takes the value 1 when the condition in parentheses holds and 0 otherwise, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:d\left(z,{\widehat{z}}_{k}\right)$$\end{document} characterizes the principle of minimal modification and can be used to limit the search range for bias correction, for example, allowing only migration between adjacent stages to suppress large jumps across stages, w_m reflects the engineering importance and risk weights of different constraints. To ensure the traceability of the bias correction process, the system synchronously outputs the bias correction evidence set as shown in Eq. 8.8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm{Reason}_{k}=\left\{m| {g}_{m}\left({z}_{k}^{\star},{x}_{k}\right)\le\:0\wedge\:{g}_{m}\left({\hat{z}}_{k},{x}_{k}\right)> 0\right\}$$\end{document} This set provides the constrained entries leading to corrections, transforming remediation from relying on unverifiable rules replacement to an interpretable, recordable, and replayable constraint consistency process. On a trustworthy gating mechanism, PSA elevates anomaly detection from single-threshold alerts to evidence-layer modeling focused on decision reliability, used to define the trustworthy operating domain of the system and provide actionable criteria for subsequent risk prediction and proactive verification. When anomaly evidence indicates structural mismatch between observations and the generative model, the system should perform confidence degradation and policy conservatism, including lowering confidence in discrete-stage conclusions, prioritizing safety constraints, and suppressing exploratory prompts, thereby avoiding misleading suggestions under boundary conditions or sensor drift. To achieve explainable and reviewable gating, PSA organizes anomaly information into multi-source statistics relevant to the current reference stage and forms an anomaly vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{A}_{k}$$\end{document}. This vector consists of three complementary signals covering three typical failure modes: point anomalies, out-of-distribution anomalies, and long-term inconsistencies. The first is stage-conditioned standardized residual statistics, using \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{r}_{k,j}=\left({x}_{k,j}-{\mu\:}_{{s}_{k},j}\right)/{\sigma\:}_{{s}_{k},j}$$\end{document} to construct cross-channel maximum deviation \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{a}_{k}^{\left(Z\right)}={\mathrm{m}\mathrm{a}\mathrm{x}}_{j}\left|{r}_{k,j}\right|$$\end{document}, exhibiting high sensitivity to single-point spikes, abrupt changes, and local faults. The second is overall low-likelihood evidence, measured by the marginal log evidence \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{l}_{k}=\mathrm{l}\mathrm{o}\mathrm{g}\sum\:_{z\in\:\mathcal{S}}p\left({x}_{k}|z\right)\hspace{0.17em}{p}_{k|k-1}\left(z\right)$$\end{document} of the predictive prior and observational likelihood, which is more sensitive to multidimensional structural disruption or regime shifts. The third is the sustained intervention rate of safety analysis, quantifying the long-term inconsistency risk implied by the frequent occurrence of corrections within a sliding window using \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{q}_{k}=\frac{1}{W}\sum\:_{i=k-W+1}^{k}1\left[{z}_{i}^{\star\:}\ne\:{\widehat{z}}_{i}\right]$$\end{document}, thereby capturing cumulative deviations and rule conflicts driven by constraints. These pieces of information jointly constitute \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{A}_{k}$$\end{document}, which is mapped via the gating function to downstream policy selection signals, enabling the system to provide both the source and analysis chain of confidence degradation alongside the output risk level, thus enhancing operational transparency and traceability in analysis. Submodel B: horizon probabilistic risk forecasting In terms of input-output relationships, HPRF relies solely on the stage posterior distributions shaped by PSA constraints and their transition structures, without directly performing black-box risk regression on raw observations, thereby maintaining semantic consistency and interpretability. The random variables manipulated by risk assessment are completely consistent with stage identification, and risk increases can be traced back to the increase in probability mass of concern stages, which can be further corresponded with abnormal evidence and credible gating results output by PSA. More importantly, PSA has suppressed physically inconsistent interpretations and enhanced posterior temporal consistency within the inference loop, making the initial distributions of rolling predictions smoother and more feasible, thereby reducing spurious risk spikes caused by infeasible posterior accumulation from the source. Formally, let the discrete stage set be \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{S}$$\end{document}, the concern event set be \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{E}\subset\:\mathcal{S}$$\end{document}, the current stage posterior be \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{p}_{k}$$\end{document}, and the transition matrix be \:A. HPRF obtains future stage distribution sequences through multi-step rolling prediction, as shown in Eq. 9. At each prediction step, the probability mass on the attention set can be used as the event occurrence probability for that step using Eq. 10.9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{c}{p}_{k+\tau\:\mid\:k}={p}_{k}{A}^{\tau\:},\:\:\tau\:=1,2,\dots\:,H\end{array}$$\end{document}10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{c}{\rho\:}_{k+\tau\:\mid\:k}=\sum\limits_{z\in\:\mathcal{E}}\left[{p}_{k+\tau\:\mid\:k}\right]\left(z\right)\end{array}$$\end{document} Subsequently, within a fixed time-domain window, future risks are aggregated to obtain a trend risk indicator that is more robust to noise and more sensitive to lead times. Taking the discrete step interval \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left[{\tau\:}_{1},{\tau\:}_{2}\right]$$\end{document} corresponding to 3 to 8 s as an example, a window average form is adopted to suppress alert jitter and introduce engineering-applicable state machine behavior, as shown in Eq. 11.11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{c}{R}_{k}=\frac{1}{{\tau\:}_{2}-{\tau\:}_{1}+1}\sum\limits_{\tau\:={\tau\:}_{1}}^{{\tau\:}_{2}}{\rho\:}_{k+\tau\:\mid\:k}\end{array}$$\end{document} HPRF employs threshold hysteresis at the triggering level and incorporates PSA’s credible gating signals into the decision logic to constrain release behavior under low credibility conditions, as shown in Eq. 12.12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{c}\begin{array}{r}{\mathrm{a}\mathrm{l}\mathrm{e}\mathrm{r}\mathrm{t}}_{k}=\left\{\begin{array}{ll}1,&\:{R}_{k}\ge\:{T}_{\uparrow\:}\hspace{0.25em}\mathrm{a}\mathrm{nd}\hspace{0.25em}{\mathrm{g}\mathrm{a}\mathrm{t}\mathrm{e}}_{k}=\mathrm{t}\mathrm{r}\mathrm{u}\mathrm{s}\mathrm{t}\mathrm{e}\mathrm{d}\\\:0,&\:{R}_{k}\le\:{T}_{\downarrow\:}\hspace{0.25em}\mathrm{o}\mathrm{r}\hspace{0.25em}{\mathrm{g}\mathrm{a}\mathrm{t}\mathrm{e}}_{k}=\mathrm{u}\mathrm{n}\mathrm{t}\mathrm{r}\mathrm{u}\mathrm{s}\mathrm{t}\mathrm{e}\mathrm{d}\\\:{\mathrm{a}\mathrm{l}\mathrm{e}\mathrm{r}\mathrm{t}}_{k-1},&\:\mathrm{o}\mathrm{thers}\end{array}\right.\end{array}\end{array}$$\end{document} The key to this link lies in the transformation of risk criteria from instantaneous peak values of single-step event probabilities to windowed aggregation of probability mass for events of concern within fixed future time domains, making it less sensitive to short-term noise, transient posterior fluctuations, and phase transition spikes. The hysteresis mechanism collaborates with credibility gating to reduce frequent switching between activation and deactivation, and prevents unsafe false deactivations when anomalous evidence is significant or credibility degrades, making the alarm rhythm better match the time scales of flight test online decision-making and human-machine collaboration. The adoption of 3 to 8 s prediction windows has engineering rationality, with approximately 3 s covering the minimum reaction magnitude from trend identification to completing one conservative manipulation or procedural action, and approximately 8 s covering the critical period during which high-dynamic test points evolve from boundary approach to unacceptable states, thereby aligning risk definition with handling procedures and avoiding insufficient lead time caused by overly short windows or credibility degradation caused by overly long windows. Windowed characterization transforms early warning from instantaneous boundary violation detection to probability assessment of entering concern sets within future windows, making threshold calibration semantically clear and explicitly presenting the trade-offs between lead time, false alarms, and missed detections; thresholds can be calibrated through performance curve scanning under constraints of missed detection or false alarm budgets in a reviewable manner. Submodel C: entropy-driven uncertainty quantification and active learning prompting (EGUAQ) In terms of computational workflow, EGUAQ takes the phase posterior \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{p}_{k}=p\left({z}_{k}\mid\:{x}_{1:k}\right)$$\end{document} output by PSA as input, first constructs moment-level uncertainty measures, then obtains segment-level scores through temporal aggregation, and finally generates executable prompting, ranking, and acquisition decisions under the constraints of confidence gating and trend risk. The primary uncertainty indicator employs posterior entropy with clear information-theoretic definition and applicability to discrete multi-class posteriors to characterize the dispersion degree of posterior distributions, supplemented by peak confidence and posterior temporal variation measures to distinguish between inherent categorical confusion of the model and short-term fluctuations caused by observational contradictions, phase transitions, or inference instability. The corresponding moment-level probability distribution characteristics are shown in Eq. 13.13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{c}\left\{\begin{array}{c}{H}_{k}=-\sum\limits_{z\in\:\mathcal{S}}{p}_{k}\left(z\right)log{p}_{k}\left(z\right)\\\:{C}_{k}=\underset{z\in\:\mathcal{S}}{\mathrm{m}\mathrm{a}\mathrm{x}}{p}_{k}\left(z\right)\\\:{D}_{k}=\frac{1}{2}\sum\limits_{z\in\:\mathcal{S}}\left|{p}_{k}\left(z\right)-{p}_{k-1}\left(z\right)\right|\end{array}\right.\end{array}$$\end{document} Where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{H}_{k}$$\end{document} reflects the overall uncertainty scale, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{C}_{k}$$\end{document} provides the confidence upper bound of the most likely stage, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{D}_{k}$$\end{document} characterizes the magnitude of posterior distribution drift between adjacent time points and can serve as a temporal consistency proxy. To match the work granularity of manual review and semi-automatic annotation, EGUAQ performs segment aggregation over a sliding window of length \:W and forms a unified score, as shown in Eq. 14.14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{c}{U}_{k}=\alpha\:\cdot\:\frac{1}{W}\sum\limits_{i=k-W+1}^{k}{H}_{i}+\beta\:\cdot\:\frac{1}{W}\sum\limits_{i=k-W+1}^{k}{D}_{i}+\gamma\:\cdot\:\left(1-\frac{1}{W}\sum\limits_{i=k-W+1}^{k}{C}_{i}\right)\end{array}$$\end{document} Among them, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\alpha\:,\beta\:,\gamma\:$$\end{document} are scalable weighting coefficients used to balance three types of signals: diffusiveness, instability, and peak insufficiency. Based on this score, EGUAQ uses the credible gated output \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{g}\mathrm{a}\mathrm{t}\mathrm{e}}_{k}$$\end{document} of PSA and the trend risk \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{R}_{k}$$\end{document} of HPRF to jointly constrain the query triggering conditions, ensuring that active learning is executed only in contexts that are credible and risk-controlled. The design benefits of EGUAQ are mainly threefold: first, all calculations are based on real-time posterior and existing module outputs, with no need to add new analytical black boxes, thereby enhancing reproducibility; second, it constructs interpretable uncertainty using entropy, peak confidence, and posterior changes, distinguishing between stable confusion and instability caused by competition during conflicting observation phases, thereby favoring samples that promote the convergence and calibration of decision boundaries; third, it constrains active learning with risk and credible gating, avoiding excessive prompting in high-risk or low-credibility states, meeting the safety-critical conservatism and reducing data interference. Results and discussion Real-time situational awareness and predictive monitoring results Using simulated flight test point sequences as evaluation carriers, the problem is formalized as a joint task of phase recognition and prospective assessment. The sequences consist of multi-channel observations \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{k}$$\end{document} with a sampling interval of 0.1 s, while phase annotations \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{z}_{k}\in\:S$$\end{document} are provided to evaluate phase recognition performance. Three random seeds are employed to measure statistical stability, with 20 sequences trained under each seed and evaluated on 10 independent test sequences. Real-time situational awareness and predictive monitoring results are shown in Fig. 2, where (a) shows the current attitude of the test aircraft, (b) presents the heatmap results of current state and predicted states 3 to 8 s ahead, and (c) provides a comparative analysis of interpretability and safety (Fig. 2). Fig. 2Experimental analysis diagram. Experimental analysis diagram. From the three result windows after model execution, it can be concluded that this method does not directly map multi-source observations to a single hard decision label, but rather implements an end-to-end, analyzable stall situation awareness and decision support chain. The system takes multi-channel flight observations as input, and based on feature engineering and probabilistic statistical modeling, outputs online posterior probability distributions of fine-grained stages, and performs multi-step probability propagation under explicit Markov transition mechanism constraints, thereby forming forward-looking state risk distributions within a fixed prediction horizon. Meanwhile, the system introduces distribution consistency analysis based on historical statistical envelopes and safety gating with physical consistency constraints to perform evidential verification and anomaly suppression on inference and prediction results, ultimately presenting three types of information synchronously on the same time axis: attitude semantic presentation, probabilistic prediction visualization, and consistency verification curves. The design goal is not to perform static classification at a specific moment, but to simultaneously satisfy system-level requirements such as stage separability, output stability, forward-looking risk assessment, and result interpretability in stall and recovery tasks, enabling online monitoring, post-event review, and strategy updates to share the same probabilistic semantic interface and form a closed loop. Test point key parameters are shown in Fig. 3, which presents the key flight parameter time series of a certain test segment and its projections in the aerodynamic characteristics and dynamics phase space, used to characterize the state evolution and potential instability symptoms near the test point. The time series displays the synchronized changes of angle of attack, pitch angle, calibrated airspeed, altitude and normal load factor. It can be observed that the angle of attack gradually increases during this process, while the airspeed shows a monotonic upward trend, and the altitude continuously decreases over time and rapidly decays to low value range in the later stage, indicating that the aircraft enters a significant energy and altitude loss process. The lower left figure shows the lift curve, where CL grows approximately linearly with AoA in the low to medium angle of attack range, conforming to the characteristics of the linear lift segment; when the angle of attack approaches the high angle of attack end, the curve exhibits inflection accompanied by slight decline, indicating that lift growth tends to saturate and initial stall or nonlinear aerodynamic effects appear. The lower right figure shows the phase plane trajectory of angle of attack and pitch angular velocity, where the trajectory smoothly migrates from the low angle of attack region to the high angle of attack region, and exhibits local loops and enhanced fluctuations at the high angle of attack end, reflecting weakened pitch dynamic damping and enhanced state coupling at this stage, with dynamic characteristics of entering limit cycle oscillation or near the instability boundary. Overall, this figure provides visual evidence of risk evolution in the test segment from three complementary perspectives of time domain, aerodynamic static characteristics and dynamics phase space, providing support for subsequent stage identification, forward-looking risk assessment and interpretation of anomaly gating results. Fig. 3Key parameter analysis chart. Key parameter analysis chart. Attitude analysis provides physical semantic grounding and readability support for model outputs, with multiple experimental results shown in Fig. 4. The window provides attitude parameters such as Pitch, Roll, Yaw and three-dimensional attitude illustrations at each sampling moment, enabling posterior probabilities of phases to be mapped to intuitive spatial maneuvering states. Stall development and recovery typically involve strong nonlinear and highly dynamic coupling characteristics, such as rapid pitch angle diving, cross-coupling between roll and yaw, and the progressive and time-delayed nature of attitude recovery during recovery processes. Relying solely on multiple numerical curves often makes it difficult to form consistent situational judgments within limited time. Three-dimensional attitude illustrations transform abstract dynamic changes into directly verifiable geometric relationships, enabling analysts to cross-check attitude evolution and phase determination consistency on the same time axis, thereby reducing the risk of conclusions being uninterpretable or disconnected from flight mechanisms. The contribution of this window is primarily reflected in information organization and verifiability aspects, converging attitude information that originally required cross-chart linkage inference into a single semantic interface, and providing a low-cost physical consistency verification entry point for the reasonableness of posterior probabilities. Fig. 4Analysis of attitude change. Analysis of attitude change. Prediction analysis comprehensively demonstrates the key capabilities of CPSSMF in probabilistic state estimation and short-term forward prediction, with multiple experimental results shown in Fig. 5. The upper half of the window displays the current stage posterior distribution, while the lower half shows the probability heatmap within the prediction interval of 3 to 8 s into the future. Therefore, the system outputs a set of probability mass functions rather than single-point labels at any given moment. Concentrated posterior indicates being within a stage with high confidence, while dispersed posterior suggests being at stage boundaries, increased observation noise, or enhanced multi-stage competition. More importantly, the system extends traditional state identification to predictive monitoring by performing multi-step rolling propagation of the current posterior through learned transition matrices, providing stage reachability and probability migration trends within future windows, thereby offering clear time margins for alarm preparation, monitoring intensification, and conservative strategy switching. Compared to threshold triggering or point-wise independent classifiers, this mechanism explicitly encodes stage persistence, evolution paths, and reachability constraints as structural priors, making prediction results semantically closer to the real dynamic evolution logic of flight processes, and to some extent avoiding state jumps driven by instantaneous noise. Meanwhile, online inference through particle filtering maintains estimation stability under conditions of observation distribution fluctuations, short-term anomalies, or local data loss through importance sampling, resampling, and probability fusion, mechanistically suppressing state jumps and enhancing temporal consistency. Fig. 5Stage prediction analysis. Stage prediction analysis. The multiple experimental results of the analysis chain design for enhanced interpretability and improved security are shown in Fig. 6. By overlaying multivariate time series with historical statistical envelopes, real-time observations are compared to historical baselines with a focus on distribution consistency. Long-term proximity to or deviations beyond the envelope for key flight parameters typically correspond to more extreme maneuvering states, domain shifts in operating conditions, sensor drift, or model mismatches—sources of risk. In such cases, even if the posterior stage has not yet clearly switched, the model can still signal risk accumulation based on statistical evidence and trigger more conservative strategies. The core value of this module lies in elevating model outputs from being utilizable to evidence-backed. The system not only reports the probable stage and future risk trends but also provides traceable evidence of rising risk, detailing which variables have deviated from the historical statistical baseline and how. Complementing the statistical analysis, physical consistency constraints and anomaly detection logic jointly form safety gating mechanisms. If model outputs are incompatible with fundamental flight dynamics or observed likelihoods consistently remain low, the system can suppress error propagation and mark anomalous segments, preventing misinterpretation of observation mismatches as normal stage transitions. Compared to strategies that rely solely on single-variable thresholds, this framework extends validation from single-point decision-making to multivariate consistency analysis, providing an engineering pathway for traceable analytical trajectories. Based on the current description, it can at least be considered that the system covers multiple critical curves and provides references with confidence interval significance in envelope forms, enabling statistically meaningful and reviewable gating criteria. Fig. 6Interpretability analysis. Interpretability analysis. Joint analysis of the three windows reveals that different time slices correspond to the evolutionary stages of stall tasks from normal flight, deceleration approach, buffeting development, entering stall to recovery. The attitude window provides spatial semantics and reflects dynamic characteristics during stage transitions, the prediction window exhibits posterior migration and future probability trends during boundary and risk escalation phases, and the validation window reveals risk accumulation processes through deviation trajectories from historical envelopes, thereby capturing precursor signals before stage transitions are fully completed. This joint presentation embodies the consistency of mechanism, statistics, and evidence. The Markov transition structure provides mechanistic constraints for stage persistence and primary pathways, Gaussian emission or equivalent statistical observation models establish probabilistic mapping between multi-source features and stages, and prediction and validation visualizations transform inference results into readable analytical chains while supporting review and attribution. The system can therefore output stable fine-grained stage posteriors at the identification level, output 3–8 s forward-looking risk probabilities at the decision level, and output interpretable evidence based on statistical consistency and physical constraints at the analysis level, thereby better meeting the requirements for verifiability, traceability, and robustness in safety-critical scenarios. Baseline scenarios and ablation analysis To ensure a rigorous evaluation of the predictive warning capabilities, we first explicitly define the evaluation protocol and calibration methodology: (1) Target Warning Event: The critical event is defined as the aircraft entering an unrecoverable ‘Deep Stall’ state; (2) Lead Time: Defined as the temporal difference (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varDelta\:t$$\end{document}) between the first timestamp where the windowed trend risk \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{R}_{k}$$\end{document} continuously exceeds the calibrated threshold and the actual physical onset of the target event; (3) Prospective Miss Rate vs. Detection Miss Rate: It is crucial to distinguish that the reported Miss Rate evaluates prospective early warning, rather than concurrent detection. In safety-critical flight test engineering, threshold calibration must strictly suppress False Alarms, which can lead to pilot fatigue and ignored warnings. Consequently, a conservative threshold is selected, intentionally trading a higher prospective miss rate for a robust, low false-alarm operational envelope. Table 1 presents a comparison of various methods on stage identification and event-level predictive warning tasks under the baseline scenario. HMM demonstrates strong stage identifiability when the distribution matching assumption holds, achieving Acc 0.994 and F1 0.848. The AoA threshold method also attains high Acc and F1, but both methods exhibit Miss 1.000 in event-level warning with unavailable or no LeadTime output, indicating that relying solely on stage discrimination or single-variable threshold triggering cannot support probabilistic risk measurement within a fixed prediction horizon. In contrast, the ablation model incorporating particle filtering and risk mean criteria already possesses effective predictive capability, with an average LeadTime of approximately 3.3 s and Miss reduced to 0.571. Furthermore, CPSSMF maintains high stage identification accuracy while achieving superior warning quality and comprehensive trade-offs, with Acc of 0.969 and F1 of 0.785. Under the same Miss level, it consistently controls FalseAlarm at 0.250, significantly outperforming ShieldOnly’s 0.812, demonstrating that the synergy between physics-consistent gating and penalty term regularization can effectively suppress false triggers caused by observation noise and boundary uncertainty. Additionally, under the same false alarm and miss constraints, CPSSMF still achieves the highest average LeadTime of 3.336 s, further improving disposal time margin compared to PenaltyOnly’s 3.319 s. Overall, CPSSMF achieves superior systematic balance among identifiability, predictability, and calibrability, better aligning with the requirements for stable warning and engineering deployability in safety-critical flight test scenarios. It is worth noting that we deliberately selected exact HMM and physical threshold logic as baselines rather than modern deep sequential models (e.g., Transformers or TCNs). In safety-critical flight test monitoring, the primary bottleneck is not achieving marginal gains in classification accuracy, but rather ensuring strict physical consistency, interpretability, and robust uncertainty quantification under extreme edge-of-envelope conditions. Deep sequential models, while powerful in data-rich environments, often operate as “black boxes” lacking rigorous probabilistic semantics and endogenous physical bounds, making them prone to silent failures during sensor anomalies or out-of-distribution events. Our ablation study precisely demonstrates that the proposed CPSSMF framework, by integrating probabilistic inference with physical safety shields, provides a transparent and calibrated risk assessment chain that pure data-driven models inherently lack. Table 1Baseline scenario and ablation analysis.MethodsWarning StrategyAccF1MissFalseAlarmLeadTime(s)Baseline_Exact_HMM_FilterN/A0.9940.8481.0000.000N/ABaseline_AoA_ThresholdAoA > = 251.0000.8611.0000.000N/AAblation _ShieldOnlyRiskMean(3–8 s) > = 0.20.9270.7330.5710.8123.303Ablation _PenaltyOnlyRiskMean(3–8 s) > = 0.20.9690.7850.5710.2503.319CPSSMFRiskMean(3–8 s) > = 0.20.9690.7850.5710.2503.336 Table 1Baseline scenario and ablation analysis.MethodsWarning StrategyAccF1MissFalseAlarmLeadTime(s)Baseline_Exact_HMM_FilterN/A0.9940.8481.0000.000N/ABaseline_AoA_ThresholdAoA > = 251.0000.8611.0000.000N/AAblation _ShieldOnlyRiskMean(3–8 s) > = 0.20.9270.7330.5710.8123.303Ablation _PenaltyOnlyRiskMean(3–8 s) > = 0.20.9690.7850.5710.2503.319CPSSMFRiskMean(3–8 s) > = 0.20.9690.7850.5710.2503.336 Baseline scenario and ablation analysis. The sensitivity analysis results of risk threshold scanning are shown in Table 2, presenting a clear cost-benefit tradeoff overall. As the threshold decreases, warning triggers become more aggressive, leading to increased average LeadTime, but FalseAlarm rises significantly synchronously, resulting in decreased effective warning ratio and weakened system availability and mission load controllability. Under current simulation conditions and event definitions, the 0.2 threshold strategy achieves a better operational balance between lead time and false alarm constraints, maintaining FalseAlarm at a relatively low level while still preserving an average lead time on the order of approximately 3 s, thus being closer to a deployable engineering operating point. Table 2Threshold analysis.Warning Strategy(CPSSMF)MissFalseAlarmLeadTime Mean (s)LeadTime Standard Deviation (s)RiskMean(3–8 s) ≥ 0.10.3571.0004.0702.007RiskMean(3–8 s) ≥ 0.20.5710.2503.3361.958RiskMean(3–8 s) ≥ 0.30.5710.2502.2850.579 Table 2Threshold analysis.Warning Strategy(CPSSMF)MissFalseAlarmLeadTime Mean (s)LeadTime Standard Deviation (s)RiskMean(3–8 s) ≥ 0.10.3571.0004.0702.007RiskMean(3–8 s) ≥ 0.20.5710.2503.3361.958RiskMean(3–8 s) ≥ 0.30.5710.2502.2850.579 Threshold analysis. Furthermore, stratified evaluation of the test sequence based on dynamic intensity reveals that under the same threshold, the average LeadTime for HighDyn conditions is approximately 2.78 s, while for Nominal conditions it reaches 6.11 s. This reflects faster state transitions and steeper phase boundaries under high dynamic conditions, resulting in a contraction of the effective predictable time domain within a fixed prediction window, thereby limiting the accumulation of foresight risk and the available lead time. This result aligns with the theoretical expectations of sequential inference concerning process noise and state transition rates. It also indicates that adopting condition-aware threshold tuning or stratified calibration can achieve more consistent control of missed detections and false alarms across different test point types, providing a quantitative basis for the subsequent design of adaptive warning strategies tailored to mission profiles. Disturbance robustness evaluation To evaluate the resilience of CPSSMF, two types of synthetic disturbance mechanisms are constructed. While these cannot fully encapsulate the complexity of real-world distribution shifts, they are designed to mathematically represent typical flight test fault typologies, such as sensor degradation and conflicting channel data. The first is Contradiction injection disturbance: randomly injects high-intensity Buffet signals in the low angle-of-attack region, simulating sensor electrical noise or localized structural vibrations unrelated to stall. The second is Calibration deviation disturbance: applies fixed bias to angle-of-attack and introduces proportional scaling errors to calibrated airspeed, mimicking pitot-static system blockage or vane misalignment. Table 4 presents representative results under contradiction injection conditions. Table 4Perturbation robustness evaluation results.MeathodWarning StrategyAccMacro-F1MissFalseAlarmLeadTime(s)Baseline_Exact_HMM_FilterN/A0.9410.7791.0000.000N/ABaseline_AoA_ThresholdAoA > = 251.0000.8611.0000.000N/AAblation_ShieldOnlyRiskMean(3–8 s) > = 0.20.9260.7310.5710.8129.441Ablation_PenaltyOnlyRiskMean(3–8 s) > = 0.20.9680.7830.5710.2503.319CPSSMFRiskMean(3–8 s) > = 0.20.9680.7830.5710.2503.336 Table 4Perturbation robustness evaluation results.MeathodWarning StrategyAccMacro-F1MissFalseAlarmLeadTime(s)Baseline_Exact_HMM_FilterN/A0.9410.7791.0000.000N/ABaseline_AoA_ThresholdAoA > = 251.0000.8611.0000.000N/AAblation_ShieldOnlyRiskMean(3–8 s) > = 0.20.9260.7310.5710.8129.441Ablation_PenaltyOnlyRiskMean(3–8 s) > = 0.20.9680.7830.5710.2503.319CPSSMFRiskMean(3–8 s) > = 0.20.9680.7830.5710.2503.336 Perturbation robustness evaluation results. Compared to unperturbed data, the accuracy and F1 of the probabilistic baseline significantly decreased, indicating that recognizers relying on statistical consistency and observational distribution fitting are highly sensitive to observational contradictions and out-of-distribution samples, with their posterior distributions being more prone to systematic bias that further deteriorates the stability and usability of event-level early warning. In contrast, CPSSMF’s stage recognition performance remained essentially unchanged, with accuracy around 0.968 and average F1 around 0.783, while the false alarm level and average lead time of early warning were close to those in unperturbed scenarios, demonstrating that feasibility shaping can constrain the perturbation of physically inconsistent observations on posterior inference within a controllable range, prevent infeasible states from occupying excessive probability mass, and suppress cascading amplification of risk time series. Ablation results further revealed that ShieldOnly exhibited obvious conservative tendencies in this scenario, manifested as high false alarms and abnormally increased lead time, reflecting that when only projecting discrete outputs onto feasible domains without in-loop stabilization of particle weight updates and posterior morphology, risk measures are easily amplified by diffuse and unstable state distributions. Relatively speaking, in-loop penalty terms are more critical in suppressing infeasible state probability mass accumulation, improving numerical stability of sequential importance sampling, and smoothing risk time axes, while the Shield mechanism’s main value tends more toward providing interpretable correction paths under extreme contradiction conditions, thereby enhancing system diagnosability and safety argument support. Meanwhile, the results demonstrate that feasibility shaping can constrain the perturbation of physically inconsistent observations on posterior inference within a controllable range. In additon, regarding calibration stability, Table 4 shows that by using the exact same calibrated risk threshold (0.2) established in the baseline scenario, CPSSMF maintains completely stable Miss (0.571) and False Alarm (0.250) rates even under severe sensor perturbations. This cross-scenario stability highlights the robustness of combining probabilistic inference with physical shields, proving that the calibration is not overfitted to a specific noise profile. Furthermore, regarding calibration stability, Table 4 shows that by using the exact same calibrated risk threshold established in the baseline scenario, CPSSMF maintains completely stable Miss and False Alarm rates even under severe sensor perturbations. This cross-scenario stability highlights the resilience of combining probabilistic inference with physical shields under controlled perturbation conditions. In the same time, we acknowledge that these synthetic tests are not a substitute for validation against real-world out-of-distribution anomalies. Therefore, our claims regarding robustness are deliberately qualified, referring only to resilience against the specific modeled fault typologies examined in this study. From a decision-making semantic perspective, the key output of this framework is not a single-stage label, but rather a structured situation vector that jointly characterizes the stage posterior distribution, forward-looking risks within a fixed prediction window, uncertainty levels, and anomaly evidence, directly supporting hierarchical decision-making closed loops. When physical consistency constraints are not triggered and both risk and uncertainty are low, the system tends to provide advancement or maintenance recommendations to complete criterion sampling and improve task efficiency; when risk is triggered or anomaly evidence is detected, the strategy automatically switches to safety-first mode, suppressing exploratory prompts and suggesting protective or termination actions; when risk remains low but uncertainty is high with no anomalies, the system outputs minimal intervention review and collection recommendations, transforming model uncertainty into manageable data closed-loop assets. This design provides explicit threshold tuning channels for the trade-offs between lead time, false alarms, and missed detections, and achieves conservative degradation under out-of-distribution or mismatched conditions through anomaly analysis chains. Conclusion This research addresses the safety-critical requirements for real-time decision-making at flight test points by proposing a unified framework that integrates probabilistic modeling with physical consistency constraints and fixed time-horizon forward-looking risk assessment, organizing decision information into two analytical chains to meet the semantic completeness and traceability requirements of engineering operations. The first analytical chain centers on trusted gating, posterior shaping, and output projection, introducing physical consistency constraints to suppress posterior probability mass drift caused by observation noise, boundary uncertainties, and contradictory evidence, outputting interpretable stage posterior distributions and anomaly indicators to avoid brittle triggering and false dismissals that rely solely on hard thresholds or empirical rules, while providing reviewable probabilistic evidence for operational actions including advancement, maintenance, rollback, and termination. The second analytical chain focuses on trend risk quantification within fixed prediction windows and entropy-driven uncertainty prompting, elevating early warning from instantaneous boundary violation detection to probability mass assessment of entering concern sets within future windows, making the trade-offs between lead time, false alarms, and missed detections explicitly parametrizable and supporting threshold tuning and risk budget configuration, while using stage posterior entropy to characterize epistemic uncertainty and information gain for triggering minimal-intervention reviews and data collection and guiding priorities for data replay and retraining, thereby transforming model untrustworthiness into actionable data closed-loop improvement pathways. Under unified simulation evaluation protocols and baseline and ablation controls, results demonstrate that this framework maintains competitive stage recognition capabilities under undisturbed conditions, while its key advantages are concentrated in the stability of risk outputs and robustness under boundary conditions. When contradictory injection perturbations are introduced, consistency constraints can effectively suppress posterior drift triggered by conflicting observations, thereby maintaining stable recognition and early warning behavior at unified operating points and providing trusted probabilistic inputs for fixed time-horizon risk prediction; simultaneously, fixed window risk measures reveal that lead time has structural upper limits under highly dynamic operating conditions, suggesting that engineering deployment should adopt layered thresholds and calibrated diagnostics, supported by robust statistical reports for safety argumentation. Overall, this research establishes an implementation approach with attributable mechanisms, quantifiable metrics, interpretable outputs, and decoupled training and online inference, with subsequent validation of its systematic utility planned on larger-scale data and richer scenario combinations. While the current evaluation based on high-fidelity procedural simulation demonstrates the theoretical soundness and robustness of the CPSSMF, we acknowledge a critical gap between simulated environments and real-world operational deployment. Real flight test environments involve unmodeled uncertainties, such as aeroelastic structural vibrations, complex atmospheric disturbances (e.g., wind shear and gusts), and unpredictable sensor degradation profiles that are difficult to fully emulate. Furthermore, the relatively small dataset size utilized in this study highlights a core advantage of the framework: because physical invariants and feasibility constraints are explicitly injected into the potential functions (PSA module), the model does not require massive datasets to implicitly “re-learn” physics, yielding high sample efficiency. By bridging Sim-to-Real gap in the furture, we aim to transition this methodological framework into a certified engineering tool for flight test safety monitoring.