GOL-SFSTS based few-shot learning mechanical anomaly detection using multi-channel audio signal

Abstract

Anomaly detection has become an essential component of mechanical equipment preventive maintenance. When encountered in complicated industrial and military applications, audio signals can be captured in a non-contact manner, preserving significant engineering potential. Common systems that operate on single-channel raw sound inputs have limited anti-noise capabilities and few-shot learning (FSL) robustness. This study proposed a competitive FSL multi-channel anomaly detection framework that uses the global optimization layer-based spallation few-shot training strategy (GOL-SFSTS) on audio Mel-spectrogram samples. The transform on 8-channel audio Mel Spectrogram increases the model's feature capture ability for multi-channel signals greatly. A 'hot-plug' discrepancy distance operator GOL is used to improve the model's anti-noise capacity when dealing with real-world operating situations. SFSTS is a growable learning approach designed to make the model more suitable for later classes and to increase FSL compatibility. We conduct ablation comparisons with other cutting-edge techniques, and FSL classification results on MIMII datasets show that our framework outperforms ResNet34 in full, 30-shot, 15-shot, and 5-shot by 9.6 %, 10.4 %, 9.8 %, and 5 %, respectively. The results of four trials indicate that the designed GOL-SFSTS framework is effective for anomaly detection applications. This paper's impressive achievements have opened up new opportunities for FSL multi-channel audio mechanical anomaly detection.