Research On Machine Abnormal Sound Detection Method Based On Self Supervised Model

Sound information is one of the most important information carriers in daily life.Using sound information to monitor the environment or various machines can make up for many deficiencies of video surveillance.Among them,abnormal sound detection is the most important technology in audio monitoring.Abnormal sound detection is actually a binary problem,that is to judge whether the monitored sound contains abnormal sound.Generally speaking,the detection method can be divided into two steps: feature extraction and model training.There are audio feature extraction and manual feature extraction.In terms of models,traditional abnormal sound detection models such as support vector machine,Gaussian mixture model or hidden Markov model have the disadvantages of insufficient modeling ability and unable to do a good representation of audio information.At present,with the development of deep neural networks,it is more and more common to use deep neural networks to achieve various audio tasks.In this paper,deep neural network is used to study the task of machine abnormal sound detection.The main work is summarized as follows:In this paper,an unsupervised model LSTM-AE-GMM,which uses only normal machine sound training,is presented under the condition that abnormal machine sound is not easy to collect.The model uses Auto Encoder(AE)as the backbone network and Long Short-term Memory(LSTM)to improve the original AE so that it can notice the time series information in the sound.At the same time,Gaussian Mixture Model(GMM)and Variational Auto Encoder(VAE)are used to generate artificial data to compensate for the shortage of training data.Compared with the AE model attached to the dataset,the LSTM-AE-GMM proposed in this paper increases the AUC value by 0.154 on average on six machines: toycar,toyconveyor,fan,pump,slider and valve,and increases the p AUC value by 0.146 on average.In this paper,an auxiliary task is added to LSTM-AE-GMM to change it to a self-supervised model.The AUC value on the six machines is increased by 0.0236 on average,and the p AUC value is increased by 0.0389 on average.Inspired by convolution,this paper also designs a self-supervised model based on Conform to detect machine abnormal sound.The result of AUC in toycar,fan,pump and valve is better than LSTM-AE-GMM with supplementary tasks.At the same time,this paper uses a variety of data enhancement methods to further improve the accuracy of the model.