ABSTRACT Predictive maintenance is increasingly supported by data‐driven methods, which can effectively learn from operational data but may overlook important structural and behavioural knowledge about the monitored system. This paper presents H‐PdM, a hybrid method that combines information derived from system models with machine‐learning‐based analysis of operational data. The two sources provide complementary information that is combined to support fault diagnosis and maintenance decisions. The method is evaluated on a triplex pump digital twin considering different fault mechanisms and their combinations. Several machine learning classifiers are compared, and the final configuration is evaluated under non‐ideal operating conditions, including measurement noise, sensor bias, drift and missing data. The results show that the hybrid approach is more robust than either the data‐driven or model‐driven component used alone. An explainability analysis also shows how system‐derived rules and data‐driven features contribute differently depending on the detected fault. Overall, the results support the use of model and data‐driven information together to achieve more reliable and interpretable predictive maintenance.
ABSTRACT Complex prediction systems often contain multiple fitted modules arranged as a directed acyclic graph (DAG), yet standard feature attribution typically explains one module at a time. We present Hierarchical Attribution Networks (HAN), a post hoc composition rule for tracing non‐negative attribution magnitude through a predictive DAG. For each observation and module, signed local feature attributions are retained, while their absolute magnitudes are normalized into routing coefficients. Terminal magnitude is the absolute deviation of the final model output from its explainer baseline. Flow is propagated upstream, stops at unresolved nodes whose attribution denominator is numerically negligible, and is aggregated over complete paths. This construction conserves routed magnitude by definition; it is not a new Shapley value, a causal decomposition or a claim of global efficiency. We evaluate the rule in a five‐module cascade using 2408 adults from the 2017–2018 National Health and Nutrition Examination Survey. Repeated nested out‐of‐fold prediction yielded weak but permutation‐detectable intermediate models (mean , , and ) and a final diabetes classifier with mean area under the receiver operating characteristic curve of 0.744. The direct age‐to‐risk path was dominant for 91.0% of participants. Synthetic experiments exposed signed‐ratio instability under cancellation and quantified sensitivity to collinearity and graph misspecification. Participant‐level bootstrap refitting, a KernelSHAP comparison and heterogeneous‐model analysis assessed stability and technical portability. HAN provides an auditable descriptive view of how local attribution magnitude is routed across modular predictors, conditional on the supplied graph, models and explainers.
ABSTRACT Owing to the highly networked and cyber‐enabled nature of cyber‐physical systems (CPSs), they are particularly susceptible to various cyber‐attacks and communication anomalies. Such vulnerabilities can lead to critical data loss, which in turn impairs subsequent tasks such as state estimation. Moreover, the above issues may render the prior information of the physical topology unreliable or unavailable in real time. To address these challenges, an improved data‐driven generative adversarial imputation network (GAIN) is proposed to recover CPS data subjected to denial of service (DoS) attacks or missing‐data scenarios. Building upon the original GAIN framework, we improve the generator by incorporating a RASA mechanism, which explicitly models both explicit and implicit token relationships to capture complex data interdependencies. Furthermore, this mechanism enhances the traceability and interpretability of attention allocation, making the proposed approach particularly well‐suited for time‐series data processing. Furthermore, the joint optimization training framework is introduced, through which the synergistic effect of the masked imputation task (MIT) and the observed reconstruction task (ORT), combined with the parallel processing advantage of the self‐attention mechanism, collectively enhances the model's imputation performance, robustness and generalization capability, thereby further strengthening the data recovery capacity of the proposed GAIN with RASA (RASA‐GAIN). Experimental results on multiple datasets, including secure water treatment (SWaT), demonstrate that the proposed RASA‐GAIN achieves an average reduction in RMSE of 53.45% on Dataset A of SWaT, 8.59% on the Letter dataset and 23.25% on the News dataset, compared to the original GAIN. Furthermore, to assess its proficiency in time‐series data recovery, we deploy the proposed RASA‐GAIN to mitigate missing‐data challenges in power systems facing both DoS attacks and random packet losses. The results indicate that RASA‐GAIN outperforms five mainstream imputation models in most scenarios, delivering superior data recovery performance across missing rates ranging from 14.594% to 57.229%.
ABSTRACT The deployment of reinforcement learning (RL) controllers for unmanned aerial vehicle (UAV) swarm navigation has been extensively studied in abstract simulation environments. However, a critical validation gap persists between algorithmic development and operationally representative testing under realistic flight dynamics. This work presents a Software‐In‐The‐Loop (SITL) validation of a federated reinforcement learning (FRL) framework for multi‐UAV swarm navigation, implemented using ArduPilot flight controller stack integrated with the Gazebo Harmonic physics simulator and orchestrated through a Python‐based middleware architecture that connects asynchronous RL policy inference with real‐time MAVLink communication. Federated Proximal Policy Optimization (PPO) framework is tested across four progressively complex operational scenarios: (i) single‐agent point‐to‐point navigation (96.0% success rate, path efficiency ), (ii) five‐UAV leader‐follower formation flight (94.0% success rate, zero inter‐agent collisions, mean formation error m), (iii) dynamic mid‐mission target reassignment requiring 143° heading correction (94.0% success rate, mean redirection time s) and (iv) priority‐based dynamic task allocation, wherein a secondary target with assigned priority emerges mid‐mission, triggering autonomous priority‐proportional swarm partitioning (90.0% dual‐target success rate, mean reallocation time s). Across all scenarios, the framework maintained a near‐zero inter‐agent collision rate (0.0% in Scenarios 2 and 4; a single incident in Scenario 3) and total federation communication overhead at or below 1.50 MB per episode, with sub‐group policy divergence () confirming effective knowledge preservation under split‐swarm operations. The results demonstrate that federated reinforcement learning constitutes a viable control paradigm for UAV swarms operating under realistic flight constraints, establishing quantitative SITL benchmarks for future sim‐to‐real transfer efforts.
ABSTRACT Trustworthy Generative Artificial Intelligence (GenAI) is often treated as a model property addressed through better training data, alignment, evaluation or guardrails. We argue instead that trustworthiness is a lifecycle governance capability that must be engineered, evidenced, monitored and adapted across data acquisition, training and alignment, safe inference, public deployment and post‐deployment use. We develop a lifecycle engineering framework integrating four governance layers – technical, organizational, human‐centred and regulatory – and translate ethical principles and major governance instruments into stage‐specific controls, responsibilities, evidence artefacts, metrics and review triggers. The framework is constructed through a structured conceptual synthesis of peer‐reviewed literature, NIST AI RMF, ISO/IEC 42001, ISO/IEC 23894, ISO/IEC 42005, ALTAI, the EU AI Act and complementary security and agentic‐governance guidance. A source register, seven‐field extraction schema, explicit mapping rules, worked extraction example and weighting sensitivity analysis make the construction logic inspectable. A public‐facing retrieval‐augmented generation case and a contrastive agentic‐system stress test illustrate operationalization through measurable indicators, responsible actors, review triggers and capability boundaries without being presented as independent empirical validation. We further address hallucination, prompt injection, privacy leakage, overreliance, unsafe tool use, memory, planners and multi‐agent workflows. The perspective positions trustworthy GenAI as an evidence‐driven, adaptive socio‐technical capability rather than a one‐time model or compliance claim.
