Online OCF Diagnosis for DTP-PMSM via Semi-Supervised Learning and Compression-based Deep Learning Network with OOD Detection

Dual three-phase permanent magnet synchronous motor (DTP-PMSM) systems are widely used owing to their high power density and reliability. However, their performance can be significantly degraded by open-circuit faults (OCFs), which are highly common faults in such systems. Although deep learning based diagnosis has achieved high accuracy, most methods are supervised and rely on labeled data. The acquisition of such data requires manual labeling, incurs labor costs, and may pose a risk of irreversible damage. A semi-supervised learning (SSL) fault diagnosis method, time series match (TS-Match), is proposed in the paper, which leverages unlabeled data through consistency regularization between weakly and strongly augmented current signals and introduces a two-stage distance-based out-of-distribution (OOD) strategy for unknown faults. On the experimental dataset and operating conditions considered in this study, both offline and online experiments show that our method achieves 100% diagnosis accuracy, compared with traditional methods, while also achieving 100% OOD detection accuracy for unknown faults. Furthermore, model compression, including pruning, fine-tuning, and quantization, enables the deployment of trained deep learning models on DSP boards for real-time online diagnosis, reducing the inference time from 3.72 ms to 3.02 ms, an 18.82% reduction, without requiring additional hardware resources.