Temporal max pooling for robust multi-class fault diagnosis of PMSM-based drive modules at unseen sensor locations

Abstract In vibration-based fault diagnosis of permanent magnet synchronous motor (PMSM) drive modules, sensor relocation can induce distribution shifts that degrade classification performance and increase diagnostic errors. This study evaluates whether a lightweight diagnostic model trained at a single sensor location can retain diagnostic reliability at unseen locations under physically induced disturbances. An in-house triaxial vibration dataset was constructed using five PMSM drive modules representing five diagnostic conditions, four sensor locations, and five disturbance patterns. A compact one-dimensional convolutional neural network–bidirectional long short-term memory (1D-CNN–BiLSTM) backbone requiring 20.20 M multiply-accumulate operations per input segment was evaluated using Raw inference, temporal majority voting, temporal moving average, and temporal max pooling (TMP). Separate speed-specific models using the same architecture, preprocessing, and evaluation protocol were examined at 500, 750, and 1000 rpm with a representative temporal window of 20 predictions. Across the two strictly unseen sensor locations, TMP improved classification accuracy over Raw inference by 6.55–20.05 percentage points, with FAR of 0.00–1.26% and MAR of 0.20–5.06%. These results indicate that max-based temporal aggregation can improve cross-location diagnostic reliability while suppressing false and missed alarms without target-location training data or additional training-time adaptation under the present experimental setting.