Self-supervised Detection Of Anomalous Sounds For Machine Condition Monitoring

With the gradual advancement of Industry 4.0,machine anomaly detection has attracted widespread attention.Among them,the machine anomalous sound detection technology aims to use the sound emitted by the machine to perform machine anomaly detection.At present,most of the research on machine anomalous sound detection technology adopts the method of deep learning.The biggest difficulty faced by the deep learning method is that it is difficult to collect the anomalous sound produced by the machine during operation,but it is easy to collect enough machine normal sounds.This makes it impossible to use traditional supervised learning methods for machine anomalous sound detection.In this thesis,a self-supervised learning method is designed to solve the problem that the machine anomalous sound detection model has no anomalous data during the training process,including three aspects:improvement of the network structure,enhancement of the convergence of the latent distribution and optimization of the anomaly score calculation rules.Firstly,the original fully connected network structure is improved by using residual convolutional network,and auxiliary tasks are designed to assist model training.The specific method is to add auxiliary branches to the latent feature layer of the model,and use the auxiliary labels contained in the normal voice of the machine to assist the training of the machine anomalous detection model.This not only ensures the training of the original anomalous sound detection task,but also enables the model to extract more accurate feature expressions.In addition to the auxiliary training of the model,the training combined with a variety of data augmentation methods makes the distribution of latent features more convergent and enhances the anomaly detection ability of the model.At the same time,the original anomaly score calculation rule is optimized by using Gaussian mixture model.Finally,the residual convolutional network using the self-supervised learning method proposed in this thesis obtained an AUC improvement of 0.104 and a pAUC improvement of 0.146 on the development dataset compared with the baseline system;AUC improvement of 0.063 and 0.131 on the evaluation dataset increased pAUC.In order to verify the detection effect of the self-supervised learning method on real machine sound,the effect of the self-supervised learning method was verified by using the measured transformer sound data and aero-engine sound data.And considering the application scenario of machine anomalous sound detection,a machine anomalous sound detection system is developed on the basis of the above self-supervised learning method.The system uses Raspberry Pi as the core board and is equipped with a touch screen to realize the human-computer interaction function,and the system has two detection modes,offline and online,which can meet different detection requirements.The off-line detection and on-line detection modes of the machine anomalous sound detection system are respectively verified using the measured transformer sound data and aero-engine sound data.The experimental results show that the self-supervised learning method can well detect machine sounds in real situations,and the detection system can distinguish normal sounds and anomalous sounds well,and realize the detection of anomalous sounds.