A Three-Level Hierarchical Fault Diagnosis Framework for APS Mechatronic Systems with Adaptive Bayesian Root-Cause Analysis

Industry 4.0/5.0 production increasingly relies on tightly integrated mechatronic assets, where closed-loop actuator–plant–sensor (APS) coupling can propagate fault effects across multiple signals and complicate diagnosis. At the same time, safety and cost constraints limit the availability of labeled fault data, while many learning-based fault detection and diagnosis (FDD) methods remain difficult to operationalize because they provide limited interpretability and weak linkage to maintenance-oriented root-cause analysis (RCA). In addition, probabilistic RCA layers are often deployed as static models that do not adapt as assets evolve over time. This paper proposes a three-level diagnostic framework for APS systems that links low-latency monitoring to maintenance-facing RCA within a single traceable pipeline. Level 1 performs binary fault detection from multivariate measurements. Once an alarm is confirmed, Level 2 estimates an interpretable fault descriptor ( F t , L ), where F t denotes the fault type and L ∈ {Sensor, Actuator, Plant} denotes its location. Level 3 activates the corresponding fault-tree branch and ranks physically plausible causes using probability-based Priority Numbers (PNs). To address the limitations of static RCA, PN rankings are updated through human-in-the-loop (HITL) maintenance feedback using a Beta–Bernoulli confirmation model for prior calibration and Laplace/Dirichlet-smoothed likelihood refinement. The framework is demonstrated on a UR5e collaborative manipulator in a hybrid evaluation setting. Nominal data acquired via Real-Time Data Exchange (RTDE) at 125 Hz are used to calibrate and validate a MATLAB/Simulink digital model, while labeled fault data for Levels 1–2 are generated by controlled fault injection in the validated model. Using a consistent offline–online preprocessing pipeline with windowed time–frequency features, multiple lightweight classifiers are benchmarked. Random Forest achieves 99.60% Level 2 accuracy in-distribution and 93.21% on a fully held-out unseen dataset with shifted trajectories and fault parameters, while maintaining near-perfect Level 1 detection (99.95% accuracy, 100% recall). Illustrative RCA cases demonstrate how PN rankings evolve after confirmed maintenance outcomes. Overall, the proposed framework provides an interpretable and deployment-oriented pathway for APS diagnosis by integrating digital-model-based fault learning with adaptive probabilistic RCA.