Intelligently Fault Diagnosis Of Machine Based On Wavelet Packet Decomposition And Deep Learning

The vibration signal is usually used for fault diagnosis of mechanical equipment.The traditional method is to reduce the noise of the signal and then analyze the time-frequency to extract signal features and match with known fault features to determine whether mechanical equipment is in fault.With the development of big data,massive vibration signals make human fault identification an impossible task.Fault diagnosis must be realized by artificial intelligence instead of people.How to combine advanced signal denoising,feature extraction technology and excellent machine learning algorithm is the focus of this paper.This paper will study the vibration signal of rolling bearing.With the research of signal noise reduction,signal feature and recognition model of machine learning,eventually realize the purpose of intelligently diagnosing and recognizing the single fault of rolling bearing.First,in terms of noise reduction technology.According to the disadvantage of traditional SVD singular value decomposition that the effective rank of singular value must be determined by human experience,this paper proposes an improved singular value decomposition noise reduction method with K-means clustering algorithm and some decision rules.As a result,the intelligent selection of effective rank order of singular values is realized.It successfully reduces the effect of human experience in the original method.And the consequence of experiment confirms that the method has good performance in noise reduction.Secondly,in the aspect of signal feature extraction,introduces the difference between time-frequency analysis and frequency-domain analysis from the difference and superiority of wavelet transform compared with Fourier transform,And a wavelet packet decomposition and reconstruction method based on wavelet transform is used.By resolving the vibration signal into different wave channels to calculate the energy of the resolved signal separately.And then get the energy characteristics of the vibration signal,combined with the time-domain characteristics of the signal,as the result of the feature extraction of the signal.Finally,in the aspect of intelligent recognition model,the current mainstream classification and recognition algorithms are studied.The traditional machine learning model and deep learning are explored respectively.All of the traditional machine learning method have achieved good results.Convolutional Neural Network(CNN)and Long Short Time Memory network(LSTM)are selected as the main methods of deep learning.Because deep learning is an end-to-end learning method,we can directly take the vibration signal as the input of the model.Finally,the recognition effect of CNN is very satisfactory.When the sample size is sufficient,CNN obtains the highest recognition accuracy,and more importantly,it saves the steps of artificial feature extraction and further realizes the intelligence.