Research On Mechanical Sound Anomaly Detection Method Based On Auto-encoder

With the rapid development of industrial unmanned,intelligence and digitalization,it is crucially important to monitor and diagnose the operation status of mechanical equipment in factories.The traditional manual inspection means have problems such as inefficient,Time-consuming,and labor-intensive.Therefore,the condition identification and fault diagnosis methods of mechanical equipment based on vibration,acoustic emission and other sensing means are widely studied and applied.Compared to vibration-based sensing which requires contact deployment on the device to be measured,acoustic sensing uses microphones to capture sound to determine the status of the target under test,which has the advantages of lower equipment cost,non-contact detection,and real-time online monitoring.A mechanical acoustic anomaly detection method based on Dual Channel Self-supervised Auto-Encoder（DCSS-AE）is proposed for the problems of difficult extraction of mechanical anomalous acoustic features and low recognition rate of timing anomalies in industrial complex environments.Then,a Mask Auto-Encoder（MAE）based mechanical sound anomaly detection is proposed,considering that the features learned by the training method of auto-encoder through reconstructed data cannot effectively distinguish between normal/anomalous sounds.The main contents are as follows:（1）In response to the inability of traditional deep self-encoders to effectively extract the time-frequency information of sound signals,a two-channel encoder composed of a bidirectional recurrent network and a fully connected network is designed to compress the normal signal and achieve the deep extraction and fusion of the time-series information and frequency domain information of normal sound.For the traditional deep self-encoder cannot effectively extract the time-frequency information of sound signals,a two-channel encoder composed of a bidirectional recurrent network and a fully connected network is designed to compress the normal signal,realize the deep extraction and fusion of the time-series information and frequency domain information of normal sound,and design a self-supervised classifier to guide the training of the above two-channel self-encoder to improve the effectiveness of its extracted time-frequency features for anomaly detection tasks.The experiments were validated using four mechanical sound data of fan,pump,valve,and slide on the MIMII data set,and the average AUC detection result was 0.810,which improved 11.8% compared with the baseline system result of DCASE2020 Challenge Task2,especially 19.2% on the non-stationary data set.（2）Considering the limited capability of the self-encoder model trained by data reconstruction for mechanical normal/abnormal detection,a mask-based self-encoder method for mechanical sound abnormality detection is proposed using a one-dimensional convolutional neural network as the model backbone.For the sound signal with two dimensions of time and frequency domain,the data are processed with frame-level masking and frequency domain masking as the input of the network,and the network is trained with the reconstruction loss of the input data and the generation loss of the data masking part as the objective function.The method increases the difficulty of network training and can effectively utilize the ability of convolutional neural network feature extraction,which in turn enables the model to learn to generate information about normal samples being masked and improves the ability of the network to distinguish normal/abnormal samples.The method is still experimentally validated on the MIMII data set,and the average AUC detection result of 0.870 is obtained,which verifies the effectiveness of the method on the mechanical sound anomaly detection task.