Consistency-Aware Weakly Supervised Anomaly Sensing for Large-Scale Expressway ETC Gantry Transactions Authors: Yijia Li, Haiyan Jiang, Xiaoxue Xu, Vladimir Zyryanov Published: 2026 ABSTRACT Electronic toll collection (ETC) gantries generate transaction records, yet existing anomaly-detection approaches often depend on manual labels or external information, limiting multiview inconsistency ranking under restricted supervision. We propose consistency-aware weakly supervised anomaly sensing for ETC (CAWS-ETC), combining monetary, temporal, structural-pattern, and contextual-semantic evidence and transferring high-confidence references to Light Gradient Boosting Machine (LightGBM). Evaluation used 14,510,847 transactions from 2525 gantries. On joint-transfer benchmarks, CAWS-ETC achieved area under the precision–recall curve (AUPRC) values of 0.9354 for rule-aligned interventions and 0.9018 for rule-orthogonal challenges, versus 0.7609 and 0.7377 for Isolation Forest. A blinded audit of 1200 unmodified transactions by two independent reviewers yielded 777 determinate labels; CAWS-ETC achieved a sampling-weighted AUPRC of 0.7946, while Isolation Forest showed higher ranking point estimates on this temporal-only subset. Post hoc attribution showed that direct rule-created and high-confidence probabilistic labels were identical after the 0.90/0.10 selection, and matched LightGBM models produced essentially identical rankings. Thus, capability beyond direct rule activation arose primarily from discriminative transfer rather than measurable gains from probabilistic aggregation or posterior-confidence weighting. Because all experiments used one day from one provincial network, multi-day, seasonal, and cross-region generalizability remain unverified. 1. Introduction Electronic toll collection (ETC) systems have evolved from automated payment facilities into distributed roadside sensing infrastructures. Each gantry passage can generate structured information on transaction time, toll calculation, vehicle category, transaction medium, and related operational states. When such records are collected across an expressway network, they can support traffic state estimation, travel time analysis, vehicle arrival prediction, infrastructure positioning, and other intelligent transportation applications [1,2,3]. The analytical value of these applications, however, depends on the integrity of the underlying transaction records. Inaccurate timestamps, inconsistent fee components, missing arrays, conflicting vehicle codes, or unusual combinations of transaction attributes can propagate into downstream estimates and operational decisions. Existing ETC data studies have addressed several aspects of this reliability problem. Missing records have been reconstructed using temporal, spatial, and multi-attribute correlations [4]. Dynamic time-warping and supervised tree-based models have also been applied to detect predefined abnormal ETC patterns [5,6]. These studies demonstrate that transaction abnormalities can be identified when the expected route structure, external capture data, or representative abnormal classes are available. Nevertheless, operational ETC databases often do not contain manually adjudicated anomaly labels, and access to supplementary topology, trajectory, image, or plate recognition data may be restricted. Under such conditions, a record can be inconsistent without belonging to a previously defined failure class, while an unusual field value can still be legitimate within a particular vehicle, medium, or transaction context. From a machine learning perspective, ETC transaction auditing is an anomaly-detection problem. Anomaly detection aims to identify observations that deviate from an expected data-generating pattern, but the definition of abnormality is inherently dependent on the application, available supervision, and structure of the data [7,8]. Existing methods include isolation-based models, one-class learning, reconstruction models, representation learning, and self-supervised transformation methods [7,8,9,10,11,12]. Although these approaches have substantially expanded the available methodological repertoire, comparative studies have shown that no single detector consistently dominates across anomaly types, supervision regimes, and data corruptions [10,11]. This limitation is particularly relevant to ETC transactions, in which deterministic accounting relations coexist with contextual temporal deviations, categorical conflicts, recurring missingness, and rare but potentially legitimate operating states. Weak supervision provides a possible bridge between rigid business rules and fully supervised anomaly classification. Instead of assigning definitive labels to individual records, domain knowledge can be represented through labeling functions that emit noisy normal or anomaly evidence and abstain when their conditions are not satisfied. A probabilistic model can then estimate the reliability of these sources and combine their outputs into weak training labels [13]. This paradigm has been increasingly considered in domains where expert knowledge is available but manually verified labels are scarce or costly [14]. However, directly applying weak rules is insufficient when the objective is to rank subtle deviations that do not cross a predetermined threshold. The weak evidence must be transferred to a discriminative model that can learn continuous feature effects and interactions while remaining computationally feasible for large transaction tables. In this study, a transaction anomaly is defined as a record that violates a deterministic field relationship, exhibits a statistically implausible deviation from comparable transactions, or contains an unusual combination of transaction attributes. This definition concerns transaction data consistency rather than confirmed fraud, toll evasion, equipment failure, or legal wrongdoing. On this basis, this study develops CAWS-ETC, a consistency-aware weakly supervised anomaly-sensing framework for ETC gantry transactions. The framework represents each transaction through monetary, temporal, structural-pattern, and contextual–semantic consistency views. Abstaining weak sources are integrated through a probabilistic latent-class model, after which high-confidence weak labels are transferred to a weighted Light Gradient Boosting Machine (LightGBM) discriminator. LightGBM was selected because tree boosting can efficiently model nonlinear interactions, missing values, heterogeneous variables, and large tabular datasets [15]. The resulting output is treated as an anomaly-ranking score rather than a calibrated probability of a specific operational cause. The main contributions of this study are summarized as follows:We formulate ETC anomaly sensing as a four-view transaction consistency problem combining monetary, temporal, structural-pattern, and contextual–semantic evidence, while excluding direct identifiers and treating the output as an anomaly-ranking signal rather than a causal diagnosis.We develop CAWS-ETC, which integrates abstaining programmatic consistency sources and transfers high-confidence references to LightGBM. A matched direct-rule LightGBM comparator is included to distinguish discriminative transfer from any incremental contribution of probabilistic aggregation.We establish a leakage-resistant multi-evidence validation protocol combining held-out time and gantry evaluation, rule-aligned and rule-orthogonal challenges, blinded adjudication of unmodified transactions by two independent reviewers, and operational proxies excluded from model construction.We evaluate robustness, interpretability, and scalability through ablation, sensitivity analysis, tree-contribution analysis, low-prevalence testing, and privacy-preserving gantry–time aggregation. On the controlled benchmarks, CAWS-ETC achieved joint-transfer area under the precision–recall curve (AUPRC) values of 0.9354 for rule-aligned interventions and 0.9018 for rule-orthogonal challenges, compared with 0.7609 and 0.7377, respectively, for Isolation Forest. We formulate ETC anomaly sensing as a four-view transaction consistency problem combining monetary, temporal, structural-pattern, and contextual–semantic evidence, while excluding direct identifiers and treating the output as an anomaly-ranking signal rather than a causal diagnosis. We formulate ETC anomaly sensing as a four-view transaction consistency problem combining monetary, temporal, structural-pattern, and contextual–semantic evidence, while excluding direct identifiers and treating the output as an anomaly-ranking signal rather than a causal diagnosis. We develop CAWS-ETC, which integrates abstaining programmatic consistency sources and transfers high-confidence references to LightGBM. A matched direct-rule LightGBM comparator is included to distinguish discriminative transfer from any incremental contribution of probabilistic aggregation. We develop CAWS-ETC, which integrates abstaining programmatic consistency sources and transfers high-confidence references to LightGBM. A matched direct-rule LightGBM comparator is included to distinguish discriminative transfer from any incremental contribution of probabilistic aggregation. We establish a leakage-resistant multi-evidence validation protocol combining held-out time and gantry evaluation, rule-aligned and rule-orthogonal challenges, blinded adjudication of unmodified transactions by two independent reviewers, and operational proxies excluded from model construction. We establish a leakage-resistant multi-evidence validation protocol combining held-out time and gantry evaluation, rule-aligned and rule-orthogonal challenges, blinded adjudication of unmodified transactions by two independent reviewers, and operational proxies excluded from model construction. We evaluate robustness, interpretability, and scalability through ablation, sensitivity analysis, tree-contribution analysis, low-prevalence testing, and privacy-preserving gantry–time aggregation. On the controlled benchmarks, CAWS-ETC achieved joint-transfer area under the precision–recall curve (AUPRC) values of 0.9354 for rule-aligned interventions and 0.9018 for rule-orthogonal challenges, compared with 0.7609 and 0.7377, respectively, for Isolation Forest. We evaluate robustness, interpretability, and scalability through ablation, sensitivity analysis, tree-contribution analysis, low-prevalence testing, and privacy-preserving gantry–time aggregation. On the controlled benchmarks, CAWS-ETC achieved joint-transfer area under the precision–recall curve (AUPRC) values of 0.9354 for rule-aligned interventions and 0.9018 for rule-orthogonal challenges, compared with 0.7609 and 0.7377, respectively, for Isolation Forest. The experiments indicate that CAWS-ETC maintained similar performance when evaluated on later within-day periods, held-out gantries, and their joint combination. Its AUPRC ranged from 0.9354 to 0.9404 on rule-aligned interventions and from 0.9014 to 0.9023 on the rule-orthogonal challenge. The blinded manual audit provided the principal real-record validation: CAWS-ETC achieved a sampling-weighted AUPRC of 0.7946 against 777 determinate manual labels, while Isolation Forest produced higher ranking point estimates on this temporal-only manually adjudicated subset. The subsequent post hoc supervision-source attribution analysis refined the interpretation of the rule-orthogonal results. After the original 0.90/0.10 posterior selection, high-confidence probabilistic labels were exactly equivalent to direct rule-created labels in training eligibility and binary class. Under matched records and model capacity, direct-rule LightGBM and the matched confidence-weighted LightGBM produced essentially identical rankings. These results indicate that the ability to rank deviations beyond direct rule activation arose primarily from discriminative transfer of programmatic consistency supervision, whereas no measurable incremental ranking benefit from probabilistic aggregation or posterior-confidence weighting was identified in the present dataset. The remainder of this paper is organized as follows. Section 2 reviews ETC transaction analysis, anomaly-detection methods, weak supervision, and the methodological gap addressed in this study. Section 3 describes the data preparation, multi-view consistency features, probabilistic weak-label learning, discriminative ranking model, and validation protocol. Section 4 presents the data-audit, detection, transferability, robustness, interpretability, and computational results. Section 5 discusses the methodological and operational implications and limitations. Section 6 concludes the paper and outlines directions for further research. 2. Literature Review 2.1. ETC Gantry Transaction Data and Anomaly Detection ETC infrastructure records a sequence of discrete vehicle–gantry transaction events and therefore provides a different type of sensing information from continuously sampled global positioning system (GPS), loop detector, radar, or video data. The records are typically sparse in physical space but rich in transaction context, including passage time, vehicle category, transaction medium, fee calculation, and route-related identifiers. Consequently, previous ETC research can be grouped into three broad categories: traffic-state and travel time inference, transaction data restoration, and abnormal-pattern detection. This functional expansion has transformed ETC data from a toll settlement resource into a source of information for broader intelligent transportation applications [1]. The first category uses ETC passage events to infer traffic conditions between gantries. Zou et al. [2] combined ETC records with spatial–temporal modeling to estimate expressway speeds, while Lai et al. [3] used ETC observations to estimate downstream vehicle arrival times. Related studies have reconstructed vehicle travel processes, classified traffic states, and inferred section-level conditions from successive gantry observations. These applications demonstrate the value of ETC data when a consistent relationship exists among passage time, spatial location, and the sequence of observations. Their reliability nevertheless depends on valid timestamps, correctly associated gantries, and complete transaction records. A record-level inconsistency can therefore affect both the transaction itself and any traffic variable derived from it. The second category focuses on missing and incomplete ETC information. Zou et al. [4] developed a missing-data restoration model using multi-attribute features, illustrating that ETC fields contain exploitable dependencies across time, space, vehicle attributes, and transaction context. Restoration methods are appropriate when the missing component and its expected counterpart can be clearly defined. They do not, however, fully address records that are syntactically complete but internally inconsistent. Examples include agreement failures between total and component fees, validly formatted but implausible timestamps, disagreements between vehicle code fields, or combinations that are rare only under a specific transaction context. Such conditions require an anomaly formulation broader than missing-value imputation. The third category detects predefined abnormal ETC patterns. Guo et al. [5] used an improved dynamic time-warping (DTW) procedure to compare ETC transaction sequences and identify abnormal data rapidly. Zou et al. [6] integrated transfer learning and Extreme Gradient Boosting (XGBoost) for ETC transaction anomaly detection, using road network topology and capture information to derive abnormal patterns. These studies show that domain structure can substantially improve detection compared with generic unsupervised modeling. Their target, however, is generally determined through predefined abnormal trajectories, supplementary sensing, or labeled source tasks. Such assumptions are not always satisfied in restricted operational datasets that contain only transaction table attributes and no manually verified anomaly classes. A further difficulty is that an ETC field cannot always be classified as normal or anomalous in isolation. A missing fee array may be legitimate for one transaction medium but suspicious for another. A disagreement between recorded and billing vehicle types may reflect an operational correction, an unusual vehicle class, or a data inconsistency. Similarly, a long elapsed time can indicate a valid extended trip, a timestamp placeholder, or an abnormal transaction sequence. Therefore, a high-level anomaly-sensing framework should distinguish deterministic contradictions from context-dependent deviations and should allow a knowledge source to abstain when its prerequisites are absent. The ETC literature has consequently established three important foundations for the present study. First, ETC records contain useful temporal, monetary, vehicle, and operational dependencies. Second, domain-informed models are more meaningful than treating all fields as independent numerical variables. Third, data quality must be evaluated before ETC records are used for downstream traffic analysis. Nevertheless, existing studies provide limited guidance on how heterogeneous consistency relations can be combined when manual labels and supplementary sensing modalities are unavailable. CAWS-ETC is designed for this setting by integrating monetary, temporal, structural-pattern, and contextual-semantic consistency evidence directly from transaction-table data, without requiring manually adjudicated labels or supplementary sensing modalities during model development. 2.2. Weak Supervision and Consistency-Aware Anomaly Learning Anomaly-detection methods are commonly categorized according to their representation of normality and the amount of supervision available. Isolation-based approaches separate unusual records through short partition paths; one-class methods estimate a boundary around normal observations; reconstruction methods identify records that cannot be accurately reproduced; and deep representation methods learn latent structures in which anomalous samples are expected to remain distinguishable [7,8,9,10]. Self-supervised approaches further construct surrogate transformations or pretext tasks to learn anomaly-sensitive representations without manually supplied class labels [12]. These methods have achieved strong results in many domains, but their performance remains dependent on the anomaly mechanism, contamination level, feature representation, and evaluation protocol. For structured ETC tables, unsupervised methods offer an attractive baseline because they do not require anomaly labels. Isolation Forest, for example, is computationally efficient and can process mixed tabular representations after suitable feature construction [16]. However, an unsupervised model has no direct knowledge that a recorded fee should equal the sum of its components, that one timestamp should not occur after another, or that the meaning of a missing field depends on the transaction medium. It may therefore assign high scores to statistically rare but valid records while overlooking contradictions that are operationally important but numerically small. Benchmark studies further indicate that increased supervision, even when limited, can materially change anomaly-detection performance and that algorithm rankings vary across anomaly types [11]. Conventional supervised learning can incorporate operational knowledge through labeled examples, but manually adjudicating ETC anomalies is expensive and conceptually difficult. A reviewer must often consult multiple fields, fee-calculation logic, temporal relations, and operational context before deciding whether a record is genuinely inconsistent. Labels can also be incomplete because only recognized failure mechanisms are documented. Treating system-generated status codes as definitive ground truth introduces another risk: the classifier may reproduce the logic of those codes rather than learn a broader transaction-consistency construct. Weak supervision addresses this setting by encoding multiple imperfect sources of knowledge instead of requiring one authoritative label. Ratner et al. [13] formalized labeling functions as programmatic sources that can express heuristics and abstain outside their domains of competence. Their outputs can be combined through a generative model that estimates source accuracy and dependence without direct access to ground-truth labels. More recent reviews have shown that weak supervision is particularly relevant to industrial monitoring and predictive maintenance problems, where domain rules and historical operating knowledge are available but verified fault labels remain sparse [14]. Weak supervision nevertheless introduces several methodological requirements. First, weak sources should express distinct and interpretable relationships rather than duplicate the same heuristic under different names. Second, normal evidence must be conservative: satisfaction of one field relation does not imply that the entire transaction is normal. Third, source correlations can bias posterior probabilities when conditional independence is only approximate. Fourth, a probabilistic weak-label model can combine rule outputs but cannot rank records that share the same weak-source pattern. A secondary discriminative model is therefore required when the objective includes continuous ranking and generalization to subthreshold deviations. Whether probabilistic aggregation and posterior-confidence weighting provide an incremental benefit once the same discriminator is trained directly on rule-created labels is therefore an empirical component-attribution question rather than an assumption of the proposed framework. Gradient-boosted decision trees are appropriate for this transfer because they model nonlinear thresholds and interactions while remaining efficient on large tabular data. LightGBM improves the computational efficiency of gradient boosting through optimized sampling and feature handling and has been widely adopted for classification and ranking tasks [15]. Its native handling of missing numerical values and its compatibility with sample weights are useful when high-confidence weak labels have unequal reliability. Tree-based models can also be interpreted through additive contribution methods. Lundberg et al. [17] showed that tree explanations can connect local predictions to global feature effects, enabling an audit of whether the model relies on plausible consistency attributes rather than excluded identifiers or operational outcomes. The present study differs from conventional weak-label classification in two respects. First, the weak sources are organized around four consistency views rather than around alternative annotators of one predefined anomaly class. Second, the final output is deliberately interpreted as a ranking score. The probabilistic model transfers domain evidence to the discriminator, but the score is not treated as a calibrated probability of fraud or equipment failure. This distinction is essential because weak labels inherit the scope and limitations of the rules from which they are constructed. 2.3. Research Gap and Methodological Positioning The literature reveals three unresolved gaps. The first concerns the target construct. Existing ETC studies commonly focus on missing records, trajectory or path abnormalities, toll evasion, or named operational failure types. These are important tasks, but they do not cover the broader situation in which a transaction exhibits internal or contextual inconsistency without a confirmed causal label. A method designed for this setting must rank heterogeneous inconsistency evidence while avoiding unsupported conclusions about why the record occurred. The second gap concerns the supervision mechanism. Generic unsupervised detectors do not incorporate known monetary, temporal, and semantic relationships, whereas supervised ETC models require abnormal examples or additional data sources. Hard rules incorporate domain knowledge but produce discontinuous binary decisions and cannot directly rank subthreshold or multiview deviations. A probabilistic weak-supervision layer can integrate multiple sources and represent source-level uncertainty, but its incremental value over a discriminative model trained directly on rule-created labels is not guaranteed. The unresolved methodological problem therefore has two parts: whether programmatic consistency supervision can be transferred into a continuous ranking model and whether probabilistic aggregation provides additional ranking value beyond this discriminative transfer. The third gap concerns validation under label scarcity. Anomaly-detection evaluation is sensitive to class imbalance, metric selection, synthetic design, and the definition of an anomalous event [18,19,20]. The precision–recall curve is more informative than the receiver operating characteristic curve when the positive class is rare because it directly reflects the trade-off between identifying anomalies and controlling false alerts [18]. Studies of time-series anomaly detection have also shown that favorable scoring protocols can overestimate performance and alter method rankings [19,20]. Therefore, a single synthetic test or operational status field is insufficient to establish credible performance. CAWS-ETC is positioned to address these gaps through a layered validation design. Rule-aligned interventions verify that the implemented pipeline recovers known consistency mechanisms, but they are interpreted only as positive controls. Held-out gantries and later within-day periods evaluate transfer without mixing contextual reference statistics across partitions. System-recorded transaction outcomes are excluded from all training stages and used only to measure independent association and high-score enrichment. Most importantly, rule-orthogonal challenges keep the primary weak-source state unchanged and test whether the discriminative model has learned continuous and multiview information beyond direct rule activation. AUPRC is emphasized for ranking under imbalance, while the area under the receiver operating characteristic curve (AUROC), precision, recall, F1-score, low-prevalence stress tests, ablation, sensitivity, and tree contributions provide complementary evidence. Accordingly, the methodological position of this study is neither purely rule-based, fully supervised, nor conventionally unsupervised. It is a domain-informed weak-supervision approach developed under restricted information and without manually adjudicated labels during model construction. Manual labels were introduced only after the original CAWS-ETC development and pre-specified evaluation pipeline had been frozen. They were used for independent evaluation of unmodified held-out transactions and were later reused only as an evaluation set for the post hoc supervision-source attribution analysis. Controlled interventions assess mechanism recovery, manual adjudication assesses observable real-record inconsistencies, operational proxies assess association rather than ground-truth accuracy, and the final score expresses consistency-based anomaly evidence rather than causal certainty. Recent studies published in 2026 further illustrate both the progress and the remaining methodological gap. Taheri Sarteshnizi et al. [21] systematically examined time-series anomaly detection methods on real-world traffic-volume data and introduced a visually inspected anomaly-labeled dataset, demonstrating the importance of domain-adapted anomaly definitions and rigorous real-data evaluation; however, their setting focuses on traffic-volume time series rather than heterogeneous ETC transaction tables and domain-specific monetary, structural, and semantic consistency relations. Wang et al. [22] extended weakly supervised tabular anomaly detection to cell-level localization, improving the interpretability of anomalous samples under limited supervision, although the method still assumes access to coarse sample-level anomaly labels. These recent developments reinforce the need for anomaly-ranking methods that can exploit domain knowledge when manually verified labels and supplementary sensing information are unavailable. Taken together, the principal advantage of CAWS-ETC in the setting considered here is its ability to combine transaction-table-only domain consistency knowledge with continuous anomaly ranking and leakage-resistant multi-evidence validation, without requiring manual labels during model development or supplementary sensing modalities. This advantage is methodological and setting-specific rather than a claim of universal detector superiority. To provide a structured comparison of the recent literature and make the methodological positioning more transparent, Table 1 summarizes representative studies published from 2021 to 2026 in terms of their main focus, principal strengths, and limitations relevant to the present study. 3. Materials and Methods The CAWS-ETC workflow is organized into five linked stages, as summarized in Figure 1. Stage 1 performs local integrity auditing, privacy-preserving preprocessing, and spatial–temporal partitioning before contextual statistics are estimated. Stage 2 constructs the 46 monetary, temporal, structural-pattern, and contextual–semantic features and the associated candidate weak sources. Stage 3 integrates the selected weak sources through the probabilistic latent-class model and transfers the resulting high-confidence references to the weighted LightGBM discriminator. Stage 4 evaluates the frozen model using within-dataset temporal, spatial, and joint holdouts, rule-aligned and rule-orthogonal challenges, blinded manual adjudication, and independent operational proxies. Stage 5 examines robustness, interpretability, and privacy-preserving gantry–time aggregation. The post hoc supervision-source attribution analysis is shown separately because it does not modify the originally fitted CAWS-ETC model. Figure 1 provides a stage-level overview, while Algorithm 1 summarizes the corresponding procedural sequence and explicitly separates model-development steps from evaluation-only procedures; Section 3.1, Section 3.2, Section 3.3 and Section 3.4 provide the full technical definitions. For easier workflow tracing, Figure 1 also identifies the principal output passed forward from each stage, while the corresponding parameters and technical definitions are detailed in Section 3.1, Section 3.2, Section 3.3 and Section 3.4. Algorithm 1. CAWS-ETC development, scoring, and evaluation workflowInput: Raw ETC transaction table D; fixed random seed 20,221,213.Output: Transaction-level CAWS-ETC anomaly-ranking scores, gantry–time anomaly burdens, and evaluation results.Stage 1: Data preparation and privacy preservation1.Audit the source schema, duplicate identifiers, timestamp validity, monetary-component consistency, and component-array integrity.2.Isolate direct identifiers, pseudonymize gantry identifiers, and exclude operational-status proxy fields from model development.3.Partition gantries and observation periods into training, validation, and test roles before estimating any context-dependent reference statistics.Stage 2: Four-view consistency representation4.Estimate contextual medians, median absolute deviations, categorical frequencies, and missingness-pattern frequencies using train_core only, and freeze these references.5.Construct the 46 monetary, temporal, structural-pattern, and contextual–semantic model-input features.6.Generate candidate weak sources and retain the nine active anomaly-oriented sources and the joint normal anchor.Stage 3: Weak supervision and discriminative ranking7.Fit the probabilistic latent-class model on the deterministic 500,000-record train_core sample and compute weak-label posterior probabilities.8.Select high-confidence anomaly references with p ≥ 0.90 and normal references with p ≤ 0.10; exclude intermediate-probability records from discriminator training.9.Select 180,000 anomaly and 420,000 normal training records, assign confidence and class weights, and fit the LightGBM discriminator using validation_joint for early stopping.10.Freeze the fitted discriminator and generate continuous CAWS-ETC anomaly-ranking scores for the required datasets.Stage 4: Independent evaluation and attribution11.Evaluate the frozen model using within-dataset temporal, spatial, and joint holdouts, rule-aligned positive controls, rule-orthogonal challenges, blinded manual adjudication, and independent operational proxies.12.Keep manual labels and operational proxies evaluation-only and exclude them from fitting, early stopping, and threshold selection.13.Conduct the post hoc supervision-source attribution analysis using direct-rule LightGBM and matched confidence-weighted LightGBM without modifying the originally fitted CAWS-ETC model.Stage 5: Robustness, interpretation, and operational aggregation14.Perform feature-family ablation, threshold/sample-size/seed sensitivity analyses, and tree-contribution interpretation.15.Aggregate transaction scores into privacy-preserving, pseudonymized gantry–time anomaly burdens for operational prioritization. 1.Audit the source schema, duplicate identifiers, timestamp validity, monetary-component consistency, and component-array integrity. Audit the source schema, duplicate identifiers, timestamp validity, monetary-component consistency, and component-array integrity. 2.Isolate direct identifiers, pseudonymize gantry identifiers, and exclude operational-status proxy fields from model development. Isolate direct identifiers, pseudonymize gantry identifiers, and exclude operational-status proxy fields from model development. 3.Partition gantries and observation periods into training, validation, and test roles before estimating any context-dependent reference statistics. Partition gantries and observation periods into training, validation, and test roles before estimating any context-dependent reference statistics. 4.Estimate contextual medians, median absolute deviations, categorical frequencies, and missingness-pattern frequencies using train_core only, and freeze these references. Estimate contextual medians, median absolute deviations, categorical frequencies, and missingness-pattern frequencies using train_core only, and freeze these references. 5.Construct the 46 monetary, temporal, structural-pattern, and contextual–semantic model-input features. Construct the 46 monetary, temporal, structural-pattern, and contextual–semantic model-input features. 6.Generate candidate weak sources and retain the nine active anomaly-oriented sources and the joint normal anchor. Generate candidate weak sources and retain the nine active anomaly-oriented sources and the joint normal anchor. 7.Fit the probabilistic latent-class model on the deterministic 500,000-record train_core sample and compute weak-label posterior probabilities. Fit the probabilistic latent-class model on the deterministic 500,000-record train_core sample and compute weak-label posterior probabilities. 8.Select high-confidence anomaly references with p ≥ 0.90 and normal references with p ≤ 0.10; exclude intermediate-probability records from discriminator training. Select high-confidence anomaly references with p ≥ 0.90 and normal references with p ≤ 0.10; exclude intermediate-probability records from discriminator training. 9.Select 180,000 anomaly and 420,000 normal training records, assign confidence and class weights, and fit the LightGBM discriminator using validation_joint for early stopping. Select 180,000 anomaly and 420,000 normal training records, assign confidence and class weights, and fit the LightGBM discriminator using validation_joint for early stopping. 10.Freeze the fitted discriminator and generate continuous CAWS-ETC anomaly-ranking scores for the required datasets. Freeze the fitted discriminator and generate continuous CAWS-ETC anomaly-ranking scores for the required datasets. 11.Evaluate the frozen model using within-dataset temporal, spatial, and joint holdouts, rule-aligned positive controls, rule-orthogonal challenges, blinded manual adjudication, and independent operational proxies. Evaluate the frozen model using within-dataset temporal, spatial, and joint holdouts, rule-aligned positive controls, rule-orthogonal challenges, blinded manual adjudication, and independent operational proxies. 12.Keep manual labels and operational proxies evaluation-only and exclude them from fitting, early stopping, and threshold selection. Keep manual labels and operational proxies evaluation-only and exclude them from fitting, early stopping, and threshold selection. 13.Conduct the post hoc supervision-source attribution analysis using direct-rule LightGBM and matched confidence-weighted LightGBM without modifying the originally fitted CAWS-ETC model. Conduct the post hoc supervision-source attribution analysis using direct-rule LightGBM and matched confidence-weighted LightGBM without modifying the originally fitted CAWS-ETC model. 14.Perform feature-family ablation, threshold/sample-size/seed sensitivity analyses, and tree-contribution interpretation. Perform feature-family ablation, threshold/sample-size/seed sensitivity analyses, and tree-contribution interpretation. 15.Aggregate transaction scores into privacy-preserving, pseudonymized gantry–time anomaly burdens for operational prioritization. Aggregate transaction scores into privacy-preserving, pseudonymized gantry–time anomaly burdens for operational prioritization. 3.1. Data Preparation and Problem Formulation The empirical data used in this study consist of electronic toll collection (ETC) gantry transaction records collected across the Shandong expressway network on 13 December 2022. The original semicolon-delimited file contained 14,510,847 transaction records from 2525 gantries over a complete 24 h observation period. Each row represents a vehicle–gantry transaction event rather than a continuously sampled vehicle trajectory. The 24 source fields describe transaction and passage identifiers, gantry attributes, transaction and entry times, payable and discounted fees, toll-interval components, transaction media, vehicle type and class, previous-gantry information, and operational status. During the development of the originally fitted CAWS-ETC and the four pre-specified comparator methods, no road-network topology, image data, vehicle speed measurements, manually verified anomaly labels, or records from other dates were available for feature construction, weak-label generation, model fitting, early stopping, or threshold selection. Manual adjudication was introduced only after these original model-development and threshold-selection procedures had been frozen. The resulting manual labels were subsequently reused only as an evaluation set for the post hoc supervision-source attribution models described in Section 3.4 and were not used for their fitting, model-capacity selection, or tuning. The 24 source fields were grouped according to their analytical roles, including transaction identifiers, gantry information, temporal attributes, monetary and component-array attributes, vehicle and transaction context, and operational-status indicators. Table 2 summarizes the representative fields in each group and specifies whether they were retained as model inputs, used only for local integrity checks, or reserved exclusively for independent evaluation. A local two-pass quality control procedure was applied before feature construction. The source file was first scanned sequentially to verify the expected field structure without loading the complete dataset into memory. The observed header exactly matched the 24-field schema, and no malformed rows, duplicated column names, or rows containing surplus values were detected. A subsequent columnar audit confirmed 14,510,847 distinct transaction identifiers and the same number of distinct row hashes; thus, neither duplicated transaction identifiers nor fully duplicated records were present. The transaction timestamp covered the entire observation day, from 00:00:00 to 23:59:59, and all three timestamp fields were syntactically parseable. The raw records were then converted into a compressed Parquet representation and subjected to refined integrity checks. Parseability alone was not treated as evidence of temporal validity. Entry and previous-gantry timestamps were considered plausible only when they were no earlier than 1 January 2000, did not exceed the observation period, and did not occur after the corresponding transaction timestamp. Under these criteria, 197,650 entry timestamps and 2,599,253 previous-gantry timestamps preceded the plausibility boundary; 438 and 706 timestamps, respectively, exceeded the observation period; and 475 entry timestamps and 986 previous-gantry timestamps occurred after the corresponding transaction time. These records were retained and represented through validity indicators rather than deleted, because invalid or placeholder timestamps constitute potentially informative evidence of inconsistency. Monetary component strings were separated using the vertical-bar delimiter and converted into component counts and sums. To distinguish the total amounts stored in each transaction from their component-level values, three monetary residuals were calculated for transaction i:(1)ripay=Fipay−∑k=1Kipayfikpay(2)ridiscount=Fidiscount−∑k=1Kidiscountfikdiscount(3)rinet=(Fipay−Fidiscount)−∑k=1Kinetfiknet where Fipay and Fidiscount denote the recorded total payable and discount amounts for transaction i, respectively; fikpay, fikdiscount, and fiknet denote the k-th payable, discount, and net-fee component; and Kipay, Kidiscount, and Kinet are the corresponding numbers of component values. A residual equal to zero indicates exact agreement between the recorded total and its component representation. The refined audit identified 78,409 payable-total mismatches, 8620 discount-total mismatches, and 78,203 net-total mismatches. No inconsistency was found among the lengths of the toll-interval and monetary component arrays. Missing component arrays, negative monetary values, and differences between the recorded and billing vehicle types were not removed or automatically labeled as anomalies. Their frequencies varied substantially across transaction media and operational states, indicating that they should be treated as context-dependent evidence rather than universal errors. The principal integrity checks, their operational definitions, and the observed numbers of affected records are summarized in Table 3. The table distinguishes deterministic inconsistencies from context-dependent conditions that were retained as analytical attributes rather than being automatically removed. Direct identifiers were isolated from the analytical workflow. Transaction identifiers were used only for local duplicate checks, after which raw vehicle plates, transaction and passage identifiers, raw gantry identifiers, and gantry hexadecimal codes were excluded from the model-development file. Gantry identifiers were replaced with stable pseudonyms generated using a locally stored random salt, and only the non-identifying plate color suffix was retained. The random salt and private mapping table remained on the local computer. The fields trade_result, fee_calc_result, and special_type were also excluded from feature-reference estimation and weak-label generation. They were reserved as independent operational proxies for later evaluation, thereby preventing system-recorded outcomes from leaking into the anomaly score. Missing numerical attributes were retained rather than deterministically imputed, and categorical levels absent from the training reference were mapped to a dedicated unknown category during model application. After privacy-preserving preprocessing, transaction i was represented by a multi-view analytical vector:(4)xi=[xiM,xiT,xiS,xiC] where xiM, xiT, xiS, and xiC contain monetary, temporal, structural, and contextual–semantic attributes, respectively. The complete vector contained 46 model-input variables. A transaction anomaly is defined as a record that violates a deterministic field relationship, exhibits a statistically implausible deviation from comparable transactions, or contains an unusual combination of transaction attributes. Let yi∈{0, 1} denote the unobserved anomaly state of transaction i, where 1 represents an anomalous transaction and 0 represents a normal transaction. Because manually adjudicated labels were unavailable during model development, yi could not be observed for feature construction, weak supervision, discriminator training, early stopping, or threshold selection. A manually adjudicated subset was subsequently constructed only after the original CAWS-ETC development and pre-specified evaluation procedures had been frozen. These labels were used for independent real-record evaluation and were later reused only as an evaluation set for the post hoc supervision-source attribution analysis described in Section 3.4. The analytical objective was therefore to construct a continuous consistency-based ranking score, whose construction is described in Section 3.2 and Section 3.3. The score was designed to express the strength of anomaly evidence rather than the calibrated probability of fraud, toll evasion, equipment failure, or legal wrongdoing. 3.2. Multi-View Consistency Modeling and Weak Label Generation Each transaction was represented through four complementary consistency views. The monetary view comprised residuals between recorded totals and component sums, component availability, parsing status, fee magnitudes, and the discount ratio. The temporal view comprised timestamp-validity indicators, entry-to-transaction and previous-gantry-to-transaction durations, and their contextual deviations. The structural-pattern view described the availability, number, and alignment of the toll-interval and monetary component arrays, together with recurring missingness and validity patterns. The contextual-semantic view described transaction medium, original and billing vehicle types, vehicle class, plate color, and the empirical frequencies of vehicle-code combinations. These four views produced 46 model-input variables; the operational proxy fields defined in Section 3.1 were excluded from the feature set. For records satisfying the timestamp-validity conditions described in Section 3.1, two elapsed-time attributes were calculated:(5)∆tientry=titrans−tientry(6)∆tiprevious=titrans−tiprevious where titrans is the transaction timestamp, tientry is the recorded expressway-entry timestamp, and tiprevious is the previous-gantry timestamp. Accordingly, Δtientry represents the elapsed time from entry to the current transaction, whereas Δtiprevious represents the elapsed time from the previous gantry to the current transaction. When either timestamp failed the validity conditions, the corresponding duration was retained as missing rather than being assigned an artificial numerical value. The data were partitioned before any context-dependent reference statistic was estimated. Gantries were stratified by their coarse G/S prefix and reproducibly assigned to training, validation, and test groups at nominal proportions of 70%, 15%, and 15%, using a fixed random seed of 20,221,213. The resulting groups contained 1768, 379, and 378 gantries, respectively. The observation day was divided into a training period from 00:00 to 17:59, a validation period from 18:00 to 20:59, and a test period from 21:00 to 23:59. The intersection of training gantries and the training period formed the core training subset (train_core), which contained 8,079,105 records. All contextual medians, median absolute deviations, category frequencies, and missingness-pattern frequencies were estimated exclusively from this subset and were frozen before application to any validation or test record. The Cartesian combinations of the spatial and temporal partitions defined the analytical roles used in the experiments. The joint validation subset (validation_joint) comprised validation gantries during the validation period. The temporal-transfer test subset (test_temporal) comprised training gantries during the test period; the spatial-transfer test subset (test_spatial) comprised test gantries during the training period; and the joint-transfer test subset (test_joint) comprised test gantries during the test period. The remaining combinations were retained as auxiliary or secondary validation records. Table 4 reports the spatial group, temporal period, analytical purpose, and record count of each role. Because the distributions of fee amounts and elapsed times were highly skewed, contextual deviations were quantified using a robust median-based distance. For a continuous attribute zi belonging to context ci, the distance was calculated as:(7)diMAD(z)=|zi−z~ci|1.4826MADci(z)+1 where zi is the observed value of the selected continuous attribute; ci denotes the context assigned to transaction i; z~ci is the median of the attribute among train_core records in context ci; and MADci(z)=medianl:cl=ci|zl−z~ci| is the corresponding median absolute deviation (MAD). The factor 1.4826 gives a normal-consistent scaling of the median absolute deviation, whereas the additive constant 1 follows the implemented feature-construction procedure and prevents numerical instability when the empirical MAD is zero. All contextual statistics were estimated exclusively from train_core and were applied to the validation and test partitions without re-estimation. Each labeling function mapped the available transaction attributes to one of three possible outputs:(8)λj(xi)∈{−1,0,1}, j=1,2,….,J where j indexes the labeling functions, J is the total number of candidate functions, −1 denotes abstention, 0 denotes normal evidence, and 1 denotes anomaly evidence. Abstention was used whenever a function’s required inputs were unavailable or its empirical context was insufficient. Deterministic anomaly functions identified payable or net-total discrepancies, discount-total discrepancies, non-missing arrays that failed parsing, invalid entry timestamps, and conflicts between an available previous-gantry field and its timestamp. Contextual anomaly functions identified entry or previous-gantry duration deviations of at least 8 robust units, rare disagreements between original and billing vehicle types, rare out-of-dictionary code pairs, and uncommon missingness patterns containing additional suspicious evidence. Rare vehicle pairs and rare patterns were defined by at most 20 training observations. Conservative normal functions represented exact monetary consistency, plausible and non-extreme temporal relations, vehicle-type pairs observed at least 1000 times, and missingness patterns observed at least 10,000 times in train_core. The candidate matrix initially contained ten anomaly-oriented and four normal-oriented functions. The monetary-array parsing function was retained as a deterministic integrity safeguard but produced no votes because every non-missing monetary array in the data was numerically parseable. Because this source had zero empirical coverage, it was excluded from probabilistic label-model fitting and consequently supplied no parser-specific weak supervision to the CAWS-ETC discriminator. Zero empirical coverage did not, however, invalidate the logical definition of the parsing check. The same predefined condition was therefore retained as a deterministic integrity safeguard for the rule-based positive-control comparators used in the controlled-intervention evaluation. Pairwise diagnostics further showed that the family-specific normal functions were highly overlapping and could conflict with legitimate anomaly evidence from another view. Because normality within one view did not establish global transaction normality, the four family-specific normal functions were combined into a conservative joint normal anchor:(9)λiJN={0,λi,MN=λi,TN=λi,SN=λi,CN=0−1, otherwise where λi,MN, λi,TN, λi,SN, and λi,CN denote the monetary, temporal, structural-pattern, and contextual–semantic normal functions for transaction i, respectively. The joint anchor λiJN emitted a normal vote only when all four normal conditions were simultaneously satisfied; otherwise, it abstained. The probabilistic label model consequently used nine active anomaly-oriented sources and one joint normal anchor. Table 5 summarizes the ten weak sources entered into the probabilistic label model. Full definitions and diagnostics for the complete candidate set are provided in the separate Supplementary Material Table S1. 3.3. Probabilistic Weak-Label Learning and Anomaly Scoring The nine active anomaly-oriented sources and the joint normal anchor defined in Section 3.2 were integrated through a regularized binary latent-class model. Let λi=(λi1,…,λiJ) denote the selected weak-source outputs for transaction i. For labeling function j, define θjlc=P(λij=l∣yi=c), where l∈{−1, 0, 1} is the observed weak vote and c∈{0, 1} is the latent class. Under the conditional-independence assumption, the observed-data log-likelihood is:(10)L(π,Θ)=∑i∈SLMlog[π∏j=1Jθj,λij,1+(1−π)∏j=1Jθj,λij,0] where SLM is the deterministic 500,000-record train_core sample used to fit the label model; π=P(yi=1) is the latent anomaly-class prevalence; and Θ is the collection of source-specific categorical probabilities. Expectation–maximization was used to estimate π and Θ. Orientation-preserving Dirichlet pseudo-counts were added during the parameter updates to avoid zero-probability estimates and latent-class switching. Given the fitted prevalence and source-specific conditional probabilities, the posterior weak-label probability for transaction i was obtained as:(11)p~i=P(yi=1|λi)=π∏j=1Jθj,λij,1π∏j=1Jθj,λij,1+(1−π)∏j=1Jθj,λij,0 where p~i is the probabilistic weak-label score generated by the latent-class model. It represents the posterior probability under the fitted weak-source model and should not be interpreted as a manually verified probability of a real operational anomaly. Three initializations were fitted, and the solution with the largest observed-data likelihood was retained after verifying that every anomaly-oriented function was more likely to vote under the latent anomaly state and that the joint normal anchor was more likely to vote under the latent normal state. All three fits converged after 58 iterations; the retained model yielded an estimated latent prevalence of 0.1852. This quantity was used only as a label-model parameter and was not interpreted as the true prevalence of anomalous ETC transactions. The fitted model was applied to all 14,510,847 records. Posterior probabilities greater than or equal to 0.90 and less than or equal to 0.10 were designated as high-confidence anomaly and normal references, respectively. These original thresholds provided 682,628 anomaly candidates and 5,970,006 normal candidates in train_core; therefore, the pre-specified relaxed thresholds were not invoked. The remaining intermediate-probability records were excluded from discriminator training rather than being forced into either class. The posterior probabilities were converted into high-confidence pseudo-labels for discriminator training:(12)y^i={1,p~i≥τA0,p~i≤τN∅,τN