Research On Anomalous Acoustic Detection Of Industrial Equipment Based On Deep Generative Models

With the era of intelligent manufacturing,the modern industrial production process is developing towards intelligence,efficiency,and integration.To guarantee the safe and stable performance of the industrial production process,it is very important to carry out effective anomaly detection on industrial equipment.Among many anomaly detection signal collection methods,acoustic signals have the advantages of simple installation of acquisition equipment,low cost,and non-contact measurement.Therefore,abnormal acoustic signal detection of industrial equipment has attracted extensive attention from researchers.In particular,the successful application of deep learning in anomaly detection in recent years has further promoted the development of abnormal acoustic signal detection technology.However,related research still faces many difficulties and challenges,such as the scarcity of abnormal acoustic signal data,the inconsistent distribution of acoustic signal data in the source domain and the target domain,etc.In response to these pain points,this paper will focus on the research of abnormal acoustic signals of industrial equipment based on the deep generation model.The specific research contents are as follows:(1)To address the scarcity issue of abnormal sound signal data in industrial equipment,a Stacked Auto Encoders(SAE)-based abnormal sound signal detection method is proposed.This method first extracts log mel-frequency spectrum features from normal acoustic signals.Second,it uses SAE to learn the probability distribution of features of normal acoustic signals.Finally,it employs the reconstruction error to judge whether the acoustic signal is abnormal.In addition,considering that the log-mel spectrum may filter out the relevant information of abnormal samples,this paper further proposes an abnormal sound detection method based on feature fusion and stacked autoencoder(FMT-SAE).Specifically,The FMT-SAT first uses a convolutional neural network to extract the time-domain features of the acoustic signal to supplement the abnormal information that cannot be obtained from the log-mel spectrum.Second,it fuses the time-domain and the log-mel spectrum features in series.Finally,it utilizes SAE to learn the probability distribution of fusion features,and then judges whether the acoustic signal is abnormal by reconstructing the error.The experimental results on the MIMII dataset show that,compared with the baseline method and the current classical method,the proposed SDAN and SAAN can effectively detect abnormal sound signals,demonstrating that the effectiveness of the proposed method.In addition,the ablation experiments showed that the feature fusion strategy obtained different degrees of improvement compared with log mel-band energies or time features only,with the AVGTotal AUC increased by 2.13% and 9.62% respectively,and AVGTotalp AUC increased by 1.28% and 4.67% respectively.(2)To address the distribution inconsistency problem of acoustic signal data in the source domain and target domain,a Siamese Domain Adversarial Network(SDAN)-based method is proposed for detecting abnormal acoustic signals of industrial equipment.Specifically,two generators with the same structure and shared parameters are used as twin networks to process data in two different domains.Second,through an adversarial learning strategy,the data distributions of the source domain and the target domain are aligned,so that the generator can effectively fit the probability distribution of normal acoustic signals in the training process.Finally,the abnormal samples are discriminated by the reconstruction error.In addition,in order to better identify abnormal acoustic signals in different industrial equipment,this paper proposes and introduces a frequency-based Attention mechanism into the SDAN,and further proposes a Siam-Attention Adversarial Network(SAAN).The experimental results on the MIMII DUE public data set show that,compared with the baseline AE and Ganomaly methods,the proposed SAAN can effectively detect abnormal acoustic signals in the scene where the data distribution of the source domain and the target domain are inconsistent,demonstrating that the effectiveness of the proposed method.In addition,the comparative experiments of SAAN and SDAN showed that SAAN could improve the performance of abnormal acoustic signal detection,the HMTotal AUC and HMTotalp AUC increased by 1.43% and 1.02% respectively.