Weakly Supervised Detection Of Machine Anomalous Sounds Based On Auto-Encoder

As an important part of security monitoring,audio monitoring makes up for many shortcomings of video monitoring,and is extremely important for intelligent monitoring and security protection.Among them,abnormal sound detection is the core technology of audio monitoring,that is,identifying abnormal sounds in security monitoring to determine whether an abnormal event has occurred.The traditional abnormal sound detection method mainly extracts appropriate features and then classifies them through a classifier.In terms of feature extraction,classic audio features and manual features are often too specific to fully express audio characteristics,leading to deviations in modeling results;traditional abnormal sound detection models(support vector machine,gaussian mixture model,hidden markov models,etc.)lack of modeling ability and inability to model normal audio characteristics well,resulting in defects in the model.In recent years,deep neural networks have performed better and better in various audio tasks.This paper uses deep neural networks to study abnormal sound detection.The main work is summarized as follows:(1)In industrial applications,judging whether the machine is abnormal through the sound of the machine is an effective method for machine status monitoring.A common challenge for machine anomalous sounds detection is the diversity of malfunctioning sounds and the scarcity of malfunctioning sounds samples.To deal with this problem,an unsupervised machine abnormal sound detection method based on the deep autoencoder network is proposed.The structure of the autoencoder is changed and adjusted,and it is compared with the traditional autoencoder.Autoencoder networks are compared in terms of detection capabilities.The experimental results show that the optimized autoencoder network has good performance and achieved good results in the DCASE Challenge 2020 Task 2 competition.(2)Constructed a weakly supervised machine abnormal sound detection algorithm based on variational autoencoder.In view of the unsatisfactory effect of the method proposed in(1)in abnormal sound detection,when only normal audio data of the machine is available,after audio feature extraction,the variational autoencoder network proposed in this paper can amplify the difference between the machine's normal sound features and other sound features in the audio domain,and at the same time create a classification boundary for the normal sound features,any sound with a different distribution from the normal sound features will be regarded as abnormal sound.The DCASE2020 Challenge Task2 development dataset is used in experiments and AUC is selected as the evaluation index.The unsupervised machine abnormal sound detection method based on the VAE was systematically compared with the Task2 baseline system and other state-of-the-art system.The experimental results show that compared with the above three methods,the AUC value of the proposed method is significantly improved,which verifies the superiority of the proposed method in machine abnormal sound detection task.