Research And Application Of Deep Self-Learning Fault Diagnosis System

As machinery equipment become more and more large,complicated,high-speed and automatic,due to the increase of correlation among the equipment,it increases the possibility of failure and the difficulty of maintenance.Therefore,the establishment of stable and reliable mechanical equipment health monitoring and diagnosis system has a very important significance.Signal processing and fault feature extraction are the key factor of the establishment of intelligent health monitoring and fault diagnosis system.However,due to the complexity of signal processing methods,the difficulty and uncertainty of artificial selection feature are increased.How to establish a model that can automatically extract the fault feature of the original signal and effectively diagnose the mechanical failure is the focus of this paper.Based on the theory of convolutional neural networks(CNN)to construct the depth network,this paper presents a new deep learning model-multi scale convolutional neural networks(MSCNN),which is more suitable for one-dimensional signal.The main contents of this paper include:Firstly,this paper deduces the forward transfer structure and inverse iterative algorithm of MSCNN.Then,the influence of the activation function and the cost function to the network is analyzed in detail,and the complexity of the algorithm is analyzed.Secondly,the MSCNN is successfully applied in the fault diagnosis of rolling bearing and planetary gearbox.Its advantages are as follows: 1)MSCNN can automatically extract the fault feature of the signal and make a diagnosis.2)In the situation that the traditional signal processing methods are failed to find fault frequency,MSCNN can accurately distinguish between different fault types of different damage levels of fault samples.Then the sensitivity of the network to the noise level and training samples numbers is studied,and it shows the excellent characteristics of the model on automatically extracting the fault feature,high precision and less required samples.Finally,the influence of the important parameters in the network is discussed in detail,and an effective model building method is proposed.Thirdly,the signal samples with labels are the necessary prerequisites for constructing accurate MSCNN.In order to make the model have the ability to identify early faults,a signal processing method for detecting early fault samples of mechanical equipment is proposed.In case of early mechanical failure,the fault signal is extremely weak and it is difficult to determine whether the signal is a fault signal.A method of noise reduction based on variational mode decomposition(VMD)is proposed,which can be applied to the early fault diagnosis of rolling bearing so that the fault signal samples can be identified at an early stage.Fourthly,the paper uses MSCNN to construct the self-learning fault diagnosis model.Taking the fault diagnosis of rolling bearing as an example,first,for the lack of real fault samples,The multiple scale convolutional neural networks is initialized using the Simulation fault signal,then the simulation samples are gradually replaced with real samples,which makes the model gradually improves,so as to realize the self-learning characteristics of the diagnosis system.