Hybrid Adaptive Modeling using Neural Networks Trained with Nonlinear Dynamics Based Features
Zihan Liu, Prashant N. Kambali, C. Nataraj · 2025
Read Hybrid Adaptive Modeling using Neural Networks Trained with Nonlinear Dynamics Based Features by Zihan Liu, Prashant N. Kambali, C. Nataraj (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2501-11835.txt.
Introduction
Zihan Liu, Prashant N Kambali, and C Nataraj build hybrid models that marry physics equations with neural networks fed by dynamics-based features. Their 2025 paper starts from a lifecycle problem. Physics models are sharp at design time but drift as real systems age. Pure data models hunger for data and fail when conditions change.
Their approach extracts information from nonlinear dynamics analysis, using perturbation methods, and embeds it as features for the learning side. Fresh measurements then update the model as the system runs. The aim is a model that keeps the physics model's feel for parameters while gaining the data model's fit to the present state. Such adaptive twins serve design, control, and fault diagnosis for complex machines.
Catalog record
| CREATOR | Zihan Liu, Prashant N. Kambali, C. Nataraj |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 54 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2501-11835
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Hybrid Adaptive Modeling using Neural Networks Trained with Nonlinear Dynamics Based Features Zihan Liu, Prashant N. Kambali, C. Nataraj Abstract Accurate models are essential for design, performance prediction, control, and diagnostics in complex engineering systems. Physics-based models excel during the design phase but often become outdated during system deployment due to changing operational conditions, unknown interactions, excitations, and parametric drift. While data-based models can capture the current state of complex systems, they face significant challenges, including excessive data dependence, limited generalizability to changing conditions, and inability to predict parametric dependence. This has led to combining physics and data in modeling, termed physics-infused machine learning, often using numerical simulations from physics-based models. This paper introduces a novel approach that departs from standard techniques by uncovering information from nonlinear dynamical modeling and embedding it in data-based models. The goal is to create a hybrid adaptive modeling framework that integrates data-based modeling with newly measured data and analytical nonlinear dynamical models for enhanced accuracy, parametric dependence, and improved generalizability. By explicitly incorporating nonlinear dynamic phenomena through perturbation methods, the predictive capabilities are more realistic and insightful compared to knowledge obtained from brute-force numerical simulations. In particular, perturbation methods are utilized to derive asymptotic solutions which are parameterized to generate frequency responses. Frequency responses provide comprehensive insights into dynamics and nonlinearity which are quantified and extracted as high-quality features. A machine-learning model, trained by these features, tracks parameter variations and updates the mismatched model. The results demonstrate that this adaptive modeling method outperforms numerical gray box models in prediction accuracy and computational efficiency. † † graphicalabstract: † † highlights: The proposed method combines physics-based modeling and data-based modeling to model complex nonlinear systems and adapt to parameter changes. The physical model is parameterized to synthesize data for model training in the absence of real data. Features extracted from nonlinear phenomena, captured by analytical solutions, are highly effective for tracking parameter variations that have an impact on related nonlinear phenomena. Feature ranking and selection techniques are employed to form an appropriate feature set for model training. The nonlinear dynamics-based features enhance the generalization and robustness of the proposed model compared to widely used gray box modeling in parameter identification. 1 Introduction The evolution of systems from localized, regulated entities into interconnected, intricate large networks has inevitably introduced challenges to accurate modeling. As systems grow in complexity, the task of modeling becomes increasingly error-prone. To address this, the integration of self-adaptation capabilities, adjusting their behavior to operational condition changes, uncertainty, nonlinearity, and various application environments to maintain their functionality and performance [ 1 , 2 ] , has emerged as a viable solution. An adaptive system can automatically modify itself in response to changes in its operating environment and model drift. For achieving adaptation, one of the fundamental system properties is self-awareness [ 3 ] , which refers to the ability to monitor resources, state and behavior. Hence, the newly measured data is critical to indicate state changes in the system. More importantly, it is essential to clearly identify the type and impact of these changes to determine whether adaptation is necessary. These decisions are grounded in prior knowledge of the system: models, rules, policies, goals and utility [ 4 ] . The following question is ‘What kind of change is needed?’ Parameter adaptation is a straightforward approach. In a well-developed physics-based model, unknown parameters or real-time varied parameters (e.g., excess loads, external disturbance, changes in ambient conditions) are identified and updated by incorporating the measured dynamic data. However, for complex realistic systems with nonlinear changes and effects that exceed the established model’s scope, additional components or environmental modifications may be required to prevent performance degradation. In these cases, data-based modeling is frequently employed in adaptive systems due to its flexibility and ability to learn nonlinear patterns. Krupitzer et al. [ 5 , 6 ] study engineering approaches of self-adaptive systems with notable references to specific applications in control [ 7 , 8 ] , biomedical [ 9 ] , mechanical [ 10 , 11 ] , and transportation [ 12 , 13 ] . In such adaptive systems, the accuracy and robustness of the adaptive model play a crucial role. The decline in system performance can be attributed to disparities within the model. Factors like wear, fatigue, and varying operating circumstances such as coupling with other mechanical components can lead to model drift and mismatch. As a result, developing an adaptive modeling technique for tracking system fluctuations is of great significance. It is necessary to note that system variations with time are usually orders of magnitude larger than the system dynamic response time. Physics-based modeling (PBM) is a traditional modeling approach that has acquired reasonable verisimilitude with real behaviors by employing traditional mathematical models and computer codes [ 14 ] . In general, PBM is based on physics or well-established rules, which are capable of qualitatively capturing some observable effects and providing useful insights into underlying causalities. The limitations of PBM derive from the fact that many realistic physical systems are getting increasingly complicated, beyond what can be adequately modeled with the physics we know. Recently, as machine learning, sensor technology, computer hardware, and big data continue to progress, the system’s uncertainties and unpredictability become more discernible in the acquired data. These have inspired a novel approach to model physical systems, namely, Data-based Modeling (DBM) [ 15 ] . Typically, the DBM approach uses observed data to learn a pattern that is used to predict system response for future inputs. Because DBM is heavily dependent on the quality and quantity of collected data, as well as the type of the system, effective, accurate and sufficient data acquisition is critical for the applications of the DBM approach. An intrinsic shortcoming of DBM is that it cannot be easily parameterized, or extended to situations that the model has not been exposed to. System deterioration, environmental changes, and operation condition changes, all of which are typical in practice, can jeopardize the validity and generalizability of the DBM. Additionally, data-based models are limited in revealing the physical causality underlying observed dynamics. As a result, they are incompetent to predict the parametric dependence of various nonlinear phenomena. Adaptive models rely on prior knowledge to construct a physical model meanwhile utilizing newly measured data from altered systems. This inherent characteristic suggests the combination of PBM and DBM. Dong et al. [ 16 ] proposed a nonlinear adaptive line-of-sight guidance algorithm in a light autonomous underwater vehicle system (LAUV) to update the varying attack angle for adaptation to perturbations. Asad et al. [ 17 ] developed a Hybrid Genetic Algorithm by combining Genetic Algorithms and traditional optimization methods to reduce the mismatches between the model and the real system for an air-conditioning set-up. Ni et al. [ 18 ] proposed an adaptive state-space model for predicting the remaining useful life of a planetary gearbox. In order to model both slow dynamics and long-term dynamics of power systems, Stajcar et al. [ 19 ] introduced an adaptive modeling framework for dynamics of multiple time frames. Sands [ 20 ] proposed a methodology using a DC motor with a minimum-phase mathematical model controlled by a self-tuning regulator without a model pole cancellation. Elkhodary et