Fault Diagnosis in an Induction‐Motor‐Driven Rotating Machinery Test Rig Under Unseen Operating Loads Using Multimodal Fusion and Multiagent Deep Reinforcement Learning

ABSTRACT Conventional fault‐diagnosis methods, largely grounded in one‐dimensional signal analysis and linear modeling, often fail to detect complex nonlinear faults in rotating machinery operating under variable loads. Data‐driven approaches also face challenges in extracting discriminative information from high‐dimensional multisensor measurements and in achieving robust generalization with limited fault data. To address these issues, this study proposes a reinforcement‐learning (RL)‐based framework for fault diagnosis in an induction‐motor‐driven rotating machinery test rig using multimodal feature fusion and multiagent learning. The framework redesigns the environment, agents, reward function, and state representation to support diagnosis from heterogeneous sensor modalities. Modality‐specific member agents extract features from vibration and current measurements, while a leader agent fuses these features into a unified representation supplied to the environment for fault classification. A diagnosis‐dependent reward, defined from the correctness of the predicted health state, iteratively updates the agent policies and fusion process. Experiments on a laboratory shaft‐bearing system driven by a three‐phase induction motor show a validation Macro‐F1 of 99.21% and a test accuracy of 97.82% under an unseen 4‐Nm load. For fair comparison, the recent RL‐based baseline was reimplemented and evaluated on the same data set, identical 0/2 Nm training‐validation split, identical 4 Nm test split, same preprocessing, and same hardware. Under this matched protocol, the proposed method achieved an absolute accuracy gain of 5.20 percentage points, equivalent to a 5.61% relative improvement, and reduced inference latency by 34.8%. These results demonstrate the effectiveness of the proposed framework for accurate fault diagnosis in the investigated rotating machinery setup.