Research On Abnormal Sound Recognition Technology Of Machine Fault Based On Deep Learning

With the development of the "Industry 4.0" strategy and the "Made in China 2025" plan,the transformation and upgrading of traditional machines has become the focus of industrial construction,and machine fault diagnosis as an important technology in the industrial field has gradually become the focus of attention.Early machine fault diagnosis is the manual diagnosis of machine faults by fault diagnosis experts or engineers through their own experience,but the manual experience of experts and engineers requires long-term accumulation,which is costly and the fault diagnosis efficiency is low.With the intelligent development of machinery and equipment and the continuous increase in the types of machine faults,the number of fault diagnosis experts and engineers has been unable to meet the needs of modern industry.In addition,there are fewer abnormal data for machine faults,and traditional machine fault diagnosis methods have limited learning of data characteristics,which brings great challenges to machine fault diagnosis with high-quality requirements.This paper analyzes,researches and compares machine fault diagnosis technology at home and abroad,and focuses on research on fault diagnosis technology and related theories based on deep learning.In view of the current lack of abnormal data in machine fault diagnosis,and the difficulty in determining the boundary between normal and abnormal,unsupervised learning based on convolutional autoencoders and supervised learning methods based on residual networks are designed and implemented,and the design is designed through simulation experiments.The method has been verified,and the specific research contents are as follows:(1)The current mainstream fault diagnosis technology is a fault diagnosis method based on an automatic encoder.In this method,the model is generally trained by learning normal sound data,so that the reconstruction error of the model on the normal sound data is small,while the reconstruction error of the unseen fault data is large,and the reconstruction error value is used to determine whether malfunction.The ability of autoencoders to extract data features is relatively weak.The convolutional layer can provide better data features for encoding and decoding in the autoencoder,so that the reconstruction error of normal data is reduced,and the difference between normal data and faulty data is increased.The reconstruction error distance between.Therefore,by designing a convolutional autoencoder and applying it to the recognition of abnormal sound of machine faults,this paper proposes a method of identifying abnormal sound of machine faults based on the convolutional autoencoder.On the evaluation data set,the AUC and p AUC values of the autoencoder-based baseline system are increased by 0.17% and 2.15%,respectively.(2)Aiming at the problem of fuzzy decision-making between normal and abnormal in the unsupervised learning method of anomaly detection,supervised learning can better describe the boundary between normal and abnormal data,so the supervised learning method can get better recognition results.In this paper,by re-dividing the data set,the unsupervised learning task becomes a binary classification task of supervised learning,and a supervised learning based residual network-based fault abnormal sound recognition method is designed and implemented,so that the model can learn better The data feature boundary of the data,which in turn improves the recognition rate of the model.In the evaluation data set,the AUC and p AUC values of the baseline system based on the autoencoder are increased by 8.76%and 16.53%,respectively.(3)Use Py Charm to build a simulation experiment platform to verify the designed method of machine fault abnormal sound recognition based on convolutional autoencoder and residual network,and compare it with the baseline system,Gaussian mixture model and convolutional neural network model The results are compared and analyzed,and then the residual network model is further verified by the expanded data set.The experimental results show that the method of machine fault abnormal sound recognition based on residual network designed in this paper has better recognition effect and generalization performance.