2DCNN, 1DCNN, BP, SVM fault diagnosis and result visualization of matlab

0、前言

        This paper aims at ten kinds of bearing fault diagnosis problems,Four classical methods are used2DCNN、1DCNN、BP、SVM进行建模,and compare the final results.

1、理论介绍

        BP和SVMThe theory is no longer described.1DCNNRefers to the use of one-dimensional convolution,2DCNNtwo-dimensional convolution,For related theories, please refer to the paper《A two-dimensional convolutional neural network optimization method for bearing fault diagnosis》.This paper proposes a new data preprocessing method,将原始time domain signal data转换成2D grayscale imageto extract the transformed image features,Eliminate the effects of handcrafted features;同时,Added noise reduction processing to the experimentally collected fault dataset before validating the classification,Parameter adaptive learning rate optimization is carried out for the gradient descent algorithm of convolutional neural network.所提2DCNNThe method achieved good results,A new way of thinking is provided for fault diagnosis.

2、方法对比

      

2.1 BP建模结果

2.2 SVM建模结果

 2.3 1DCNN建模结果

 2.4 2DCNN建模结果

 2.5 Comparison of test classification results

3、Feature visualization and comparative analysis

3.1  BPHidden layer feature visualization

 3.2 SVM特征可视化（PCAThe feature visualization of the sample in the kernel space after dimensionality reduction）

3.3 1DCNNHidden layer feature visualization 

3.4 2DCNNHidden layer feature visualization

 4、结果分析

        上述BP与SVM效果不太理想,This is because the input of these two methods is raw signal data,The shallow model has limited processing effect on the original signal data.CNNDeep models can automatically learn abstract representations of raw data,This avoids handcrafted features designed by engineers,And compared with traditional machine learning methods, good results have been achieved.The most common data type is the time domain signal,Various deep learning methods for processing one-dimensional signals have been applied in real-time motor fault diagnosis.But they are all one-dimensional time series signals,Feature extraction is prone to feature loss,However, the current mainstream two-dimensional convolutional neural network structure is not directly applicable to one-dimensional vibration signals,As a result, it is necessary to deepen the depth of the commonly used one-dimensional convolutional neural network to obtain a larger receptive field,从而抑制过拟合,This increases the difficulty of design to a certain extent.This paper refers to the data preprocessing method proposed in the above paper,Convert the raw time-domain signal data into a two-dimensional grayscale image,There are no predefined parameters,This eliminates expert experience as much as possible,And can use convolutional neural network widely used in image recognition（2DCNN）.

        Inspired by the paper,将原始1Convert dimensional signal data to 2D data（即图像）就可以采用Image recognition is widely used in convolutional neural networks（2DCNN）,The conversion method is not limited to the two-dimensional grayscale image proposed in the paper,Raw signal data can also be passed through小波变换、EEMD变换、VMD变换Obtained by isochronous frequency domain analysis method时频图（二维图像）,然后采用2DCNN对时频图进行训练,实现分类（或预测）.