Multi-Scale Temporal Convolutional Networks for Robust Variable-Speed Fault Diagnosis in DC Motor Transmission Systems

Anomaly detection in Direct Current (DC) motors plays a vital role in various industries. However, it remains a challenge to develop systems that are intrinsically robust to unpredicted speed variations and environmental noise. This paper proposes the use of Multi-Scale Temporal Convolutional Network (MSTCN) enhanced with a Convolutional Block Attention Module (CBAM) for feature extraction. The proposed scheme was validated using a proprietary dataset collected with an Arduino Nano 33 BLE IMU. A zero-shot multi-variate stress test involving Fourier-domain resampling and additive Gaussian noise with speed factors and standard deviation levels ranging from 0.5 to 2.0 and 1.5 to 2.0, respectively, was used to perturb the dataset. Under these conditions, it was observed that while the vanilla CNN achieved the lowest latency of 207 microseconds, it failed under high interference, scoring an F1-score of nearly 60 percent at noise level 2.0. The LSTM model reached high accuracy but experienced a computational bottleneck with latency of 1620 microseconds. In contrast, our MSTCN-CBAM model achieved an F1-score of nearly 90 percent at noise level 2.0 and up to 98 percent under moderate noise conditions while providing a 2.6-fold inference speedup (617 microseconds) and a 55 percent reduction in training time compared with the LSTM baseline. These encouraging results highlight the advantages of our MSTCN-CBAM model for robust anomaly detection in the real-world conditions.