Research On Bearing Equipment Fault Diagnosis Based On Convolutional Network And Multi-task Learning

Bearing as one of the most common and widely used mechanical equipment in industrial production,its faults have a great impact on production efficiency and cost,so achieving fast and precise bearing fault diagnosis is crucial.Currently,bearing equipment fault diagnosis often relies on the knowledge and guidelines of experts,which can be time-consuming and resource-intensive.Additionally,it is susceptible to errors,particularly in complex fault scenarios.With the development of sensor technology and deep learning,data-driven bearing fault diagnosis method has been widely used.In this paper,deep learning technology is used to study the fault diagnosis of bearing equipment for the following problems: bearing equipment operating environment is bad,vibration signal data collected through the sensor noise is complex;Traditional deep convolutional neural networks have limited feature extraction ability and cannot learn the periodic characteristics of sequential signals.Fault diagnosis and life prediction are two important objectives and tasks in the operation support of bearing equipment.In the operation and maintenance of bearing equipment,it is often desired to obtain bearing health state information and degradation information at the same time.Traditional methods need to train fault diagnosis and life prediction models separately,which is inefficient and can not take advantage of the correlation between tasks.The focus of this study is the fault diagnosis of bearing equipment,utilizing vibration signal data from the bearings.The aim is to address the previously mentioned issues and ensure the stable operation of machinery,thus safeguarding the safety of people and property during production and daily life.The main areas of research covered in this paper are as follows:1.To address the challenges presented by the complex noise and non-stationary signals found in bearing vibration data,this paper proposes a denoising approach.The method is based on an improved complete ensemble empirical mode decomposition with adaptive noise(ICEEMDAN)and wavelet transform adaptive threshold denoising(WTATD).In this method,the original signal is modal decomposed based on the ICEEMDAN,and the solid component with different frequencies is obtained.Then,the high-frequency component is denoised by the improved WTATD.Ultimately,the denoised signal is acquired through a synthesis of various categories.The simulation signal denoising experiment and the fault classification experiment after bearing vibration signal denoising are carried out.The experimental results demonstrate that compared to several traditional denoising methods,the proposed denoising approach yields superior results in the denoising of non-stationary signals.2.A multi-scale fusion lightweight convolutional neural network with periodic analysis was proposed to solve the problems of low training efficiency and lack of timing analysis ability in the classification of bearing equipment fault types of traditional deep convolutional neural networks,which resulted in poor classification accuracy.In this method,a lightweight convolutional neural networks with dual attention mechanisms of channels and positions(ALCNN)is firstly established.Then combined with the Fourier transform periodic analysis method to extract the periodic features,and finally fused the original signal features extracted by one-dimensional ALCNN,so as to improve the efficiency of model training while accurately carrying out bearing equipment fault diagnosis.By comparing the bearing fault diagnosis experiment with a variety of traditional deep convolutional neural networks,the proposed method significantly enhances the accuracy of fault diagnosis,thereby confirming its efficacy in the realm of bearing equipment fault diagnosis.3.In view of the low efficiency of establishing a single-task model for the traditional fault diagnosis and remaining useful life prediction required to ensure the operation of bearing equipment,as well as the low generalization ability of the hard sharing multi-task method,a new multi-task learning method is introduced,which is based on the proposed ALCNN with Fourier periodic analysis(FP-ALCNN).The proposed method is a soft-sharing model with a shared feature extraction structure.The method is based on FP-ALCNN as the shared feature extraction layer,and designed a private feature extraction structure for specific tasks based on the attention mechanism.At the same time,the dynamic task weight is set based on the difficulty of the task,and the multi-task learning method of soft sharing mode is realized.Bearing fault diagnosis and remaining life prediction can be carried out at the same time.The training efficiency and generalization ability of the model are effectively improved.Finally,experiments of fault diagnosis and residual life prediction were carried out on the degradation data set of bearing lifetime operation.Compared with single task learning method and traditional hard shared multi-task learning method,the proposed multi-task learning method in the fault diagnosis and residual life prediction results have been comprehensively improved in accuracy.Experimental results demonstrate that the proposed method possesses superior accuracy and generalization capabilities.