Research On Condition Recognition Methods Of Rolling Bearings Based On Deep Learning

With the development of industrial Internet and intelligent manufacturing technology,modern mechanical systems can collect a large number of real-time data which reflecting the operation status of mechanical equipment and even rolling bearings.How to effectively diagnose mechanical faults driven by this "big data" has become a front hotspot.Aiming at the fault diagnosis of rolling bearings,this paper puts forward several methods to recognize the running state of rolling bearings based on deep learning model.The main content of the article is displayed as follows:Considering the non-linear mapping and feature extraction capability of the deep neural network and the advantages of LSTM in processing time series data,an end-to-end fault state identification method for bearing based on LSTM is proposed.The original vibration signals with rich information and details are directly input into the LSTM neural network model,and the diagnosis results are directly output.By reducing the steps between the input of the original vibration signal and the final fault diagnosis result,the adaptive ability of the deep neural network can be fully utilized while obtaining more abundant original information.Aiming at the problem of information loss in frequency domain in LSTM model,a fault state identification method for bearing based on time-frequency image and twodimensional convolution neural network is proposed.The original vibration signal is transformed into a time-frequency image,and the time and frequency are regarded as two dimensions of the image.Then the image is input into the 2DCNN model for classification and recognition.By accumulating multi-layer convolutional pooling pairs,2DCNN model can see relatively long history and future information,which ensures that 2DCNN can well express the long-term correlation of vibration signals,while directly retaining the relevant information in frequency domain.Compared with LSTM network structure,2DCNN model has better robustness.Aiming at the problem that 2DCNN model needs short-time Fourier transform and consumes a lot of computing resources,a end-to-end fault state identification method for bearing based on 1DCNN is proposed.The 1DCNN model directly accepts the original vibration signal as input,and then uses one-dimensional convolution kernels with larger size and longer step length in the first convolution layer to replace the process of time-frequency conversion of vibration signal.Then,a combination of multi-layer one-dimensional convolution layer and pooling layer is used to model the vibration signal.This method not only simplifies the process of data processing,but also uses one-dimensional convolution instead of short-time Fourier transform and two-dimensional convolution,which greatly reduces the amount of computation and the consumption of computing resources.In order to validate the effectiveness of the above algorithm,several bearing fault data sets are used for experimental verification.Finally,a software system is developed,which integrates the deep learning models proposed in this paper and provides end-to-end fault diagnosis solutions for bearings.