Research On Fault Diagnosis Method Based On Attention Mechanism And Deep Learning

Mechanical systems are becoming more and more complex,sophisticated and automated.Once a failure occurs,it will bring serious safety issues.It is vital to detect and identify the failure in time.Rolling bearing is an important part of mechanical equipment,and the research about its fault diagnosis methods is of representative significance to ensure the safety and stability of mechanical systems.Therefore,this paper selects bearing vibration signals as the object,and conducts fault diagnosis methods based on deep learning.The occurrence time of some faults is extremely short,so the effective samples are extremely difficult to obtain.Therefore,using fewer samples to obtain higher recognition accuracy is an important research direction.In order to obtain richer hidden features,this paper uses continuous wavelet transform to extend the bearing vibration signal to the time-frequency domain;And it uses the natural image processing advantages of the two-dimensional convolutional neural network to design fault diagnosis algorithms;The paper enhance the depth Learn the utilization of samples by the signal processing technology.Continuous wavelet transform will expand invalid features and affect the ability of the diagnostic algorithm to resist noise.In order to solve this problem,firstly the paper adjust the parameters according to the influence of wavelet basis parameters on the model diagnosis results;then it uses the principle of average pooling and down-sampling to design a multiscale feature extraction layer,and reduces noise through the multi-mode characteristics of the bearing vibration signal the interference brought by;finally,the channel attention mechanism is used to evaluate the weight of each order feature,which enhances the model's utilization of effective information.In addition to accurate identification,fault diagnosis also needs to be detected in time.Embedded devices are more suitable for diagnostic sites with high real-time requirements due to their low power consumption and easy deployment characteristics.Algorithms based on deep learning usually have the characteristics of high algorithm complexity and large amount of calculation,so their embedded applications need to improve the diagnosis speed.This paper which starts from the two angles of algorithm and hardware,designs a lightweight diagnosis algorithm model based on one-dimensional convolutional neural network,and uses FPGA to accelerate the model reasoning process,which improves the timeliness of diagnosis.Finally,a fault diagnosis system is built based on the design of edge computing framework,and the methods in this paper are integrated.The edge node loads the fast diagnosis model for rapid analysis in the data source,and the cloud platform uses a more complex two-dimensional convolution model to complete the more accurate identification of the bearing state.Validated by the standard data set of Western Reserve University,the methods in this paper have extremely high sample size tolerance and strong anti-noise ability.By building a fault diagnosis system and testing it on a vibration test bench,the system can quickly and accurately diagnose bearing faults.The fault detection rate can reach 100%,The recognition rate remains above 99%.And the model diagnosis time for edge nodes can be compressed to 0.178s.