Design And Implementation Of Machine Abnormal Sound Detection System Based On Unsupervised Learning

With the development and progress of industrial technology,more and more types of machines are used in production and people’s daily life.Machines are closely related to our lives.Therefore,regular maintenance and inspection of machinery and equipment are carried out to detect abnormalities in advance and repair them to ensure that machinery and equipment are normal operation becomes a necessary task.When the machine is running,there is often a regular sound,but when an abnormality occurs,the sound will change,so the sound can be used as a standard to judge whether the machine is operating normally.Based on the characteristics of machine operating sounds,this paper designs and implements an unsupervised machine abnormal sound detection system.The main research and work contents are as follows:（1）The occurrence of machine anomalies is unpredictable.Deliberate damage to the machine can only obtain limited types of abnormal sounds,and the cost is high,so it is difficult to collect abnormal sounds.In response to this problem,this paper proposes to use a network model based on unsupervised learning to train sound features.The training process only uses the sound of the machine running normally,so as to avoid the problem of lack of abnormal sound.In MFCC feature extraction,a filter bank is used to simulate auditory features from an angle,while the extraction method of modulated spectral features uses a gammatone filter bank and an 8-channel modulation filter bank to simulate the filtering in the cochlear structure and the human auditory system,respectively.Structure,extract acoustic features from two perspectives to make the acoustic features more obvious,and the training target of the neural network is more clear.Through experimental comparison,it is confirmed that the characteristics of the sound are represented by the modulation spectral features.（2）There are various types of machine sounds,and a network model that classifies and trains multiple sound characteristics will not perform better than a model trained for one sound,so the abnormal sound detection system trains a model for each type of sound.It is also because there are many types of machine sounds and it is difficult to collect sounds.It is unrealistic to train a voice model in advance.Different models need to be trained for different types of machine sounds.With the increase of audio data,the model needs to be updated in time,and the system model needs to be updated.To expand,design an abnormal sound detection system based on B/S（browser/server）architecture based on the above requirements.The main function of the system is to train models for different types of sounds to detect whether the sound is abnormal.In order to make the system application more convenient,a small program is designed to be used on the mobile phone.The small program can record and save the voice file to the directory or record and detect it in time.The current mainstream neural network framework is written in Python,and it is more convenient to use Python to write neural networks than Java,so use Python to write network models and complete the processing,training and prediction of sound data.Socket communication is used to realize the communication between the web terminal system and the network training and speech prediction module written in Python language.Such a system design can meet the maintenance and inspection of various machines such as machines in the workshop,machines deployed in public places,and enterprise or private custom machines.Organize,train and use.To sum up,this paper designs and completes an unsupervised abnormal sound detection system,which can train a neural network model to detect the sound of the machine running to judge whether the machine is operating normally,meet the needs of regular machine maintenance and inspection,and prevent machine operation failures..This system provides an interface for network model training and abnormal sound detection,and completes a basic abnormal sound detection system,which can be quickly deployed and applied.