CTAP-Net: A Multi-Head Self-Attention-Based Multi-Task Framework for BLDC Motor Fault Diagnosis and RUL Prediction

Abstract

The significance of predictive maintenance in Brushless DC (BLDC) motors has increased rapidly because of the continued need for higher reliability, operational efficiency, downtime reduction, and ongoing health assessment. Existing methods tend to consider fault diagnosis and Remaining Useful Life (RUL) assessment as independent tasks which limits their capabilities to use similar representation for different diagnostic and forecasting objectives. Herein, we propose a Multi-Head Self attention based Prognostic Network (CTAP-Net) for simultaneous fault severity assessment and RUL forecasting through multi-task learning. CTAP-Net employs a combination of stacked Long Short-Term Memory (LSTM) networks and multi-head self attention mechanism for capturing temporal degradation patterns developed based on multivariate motors’ signals such as current, temperature, speed, and torque. In addition, there is the implementation of a cross-task feature fusion mechanism that encourages shared learning of representations in both the prognostic and diagnostic branches. Results of the experimental evaluation show that CTAP-Net achieves accuracy of 89.1750% in fault severity classification and Macro-F1 of 0.890904. Regarding RUL estimation, this model yields RMSE equal to 0.042591, MAE equals 0.032429, and $R^{2}$ equal to 0.952374. To avoid temporal data leakage, the data splitting was performed before any construction of the sliding window without randomization. These experimental results prove the efficiency of the CTAP-Net architecture in combined modeling of diagnostic and prognostic tasks.