The Research And Detection Of Acoustic Anomaly Events Based On Deep Dictionary Learning

Today,when mechanized production has become the mainstream,the smooth operation of the mechanical system is the fundamental to ensure the efficiency and quality of mechanical production,so the monitoring of faults in the production process is very important.We can achieve this aim with mechanical acoustic detection.Facing the feature of training dataset that there are only normal acoustic audio and no abnormal acoustic audio,we use the dictionary learning model to capture the most essential features of audio to obtain their perfect performance on the unsupervised one class classifier.The main work of this paper is summarized as follows:First,extract inherent physical characteristics in mechanical operation,and use them as features for subsequent analysis.In view of the characteristics that the normal sound and abnormal sound of mechanical operation are very similar,we propose an acoustic anomaly event detection system base on dictionary learning.Through dictionary learning and sparse representation,the essential features of each machine are obtained,a feature domain is established for each machine,and its sparse representation coefficients on this domain are obtained.In this way,we can expand the distance between normal and abnormal running sounds for better classification prediction results.We use the dataset in the DCASE2020 T2 task to conduct experiments,and use AUC as the evaluation index to compare the results of the acoustic anomaly detection system based on dictionary learning and the Baseline system given by the T2 task.The result proves the effectiveness of the acoustic anomaly event detection based on dictionary learning in some machine types.Secondly,in response to the problem of poor performance on some machine types in the above point,we want to further expand the distance between the normal and abnormal data features in the dataset,and use the currently widely used deep learning concept to combine it with the dictionary learning fusion,an acoustic anomaly event detection system based on deep dictionary learning is proposed.In order to show the effectiveness of the system compared with the shallow dictionary learning model,we conduct experiments under the same conditions.The experiments show that the model based on deep dictionary learning has stronger feature representation ability than the shallow dictionary learning model,and further improves the AUC results.The effectiveness of the system is verified.