Bearing life prediction practice based on deep learning, development of CNN, fusion of LSTM/GRU/ATTENTION

Most of the previous bearing-related projects were of the type of fault identification and diagnosis, and few of them involved regression prediction. On weekends, Zhaijia found a data set of bearing life acceleration experiments and wanted to use it for life prediction.

First look at the dataset as follows:

You can search directly on Baidu, so I won’t go into details here.

Learning_set is the training set

Test_set is the test set

Here I directly use Learning_set as the total data set for simple processing, and randomly divide the specified proportion as the test set.

Of course, you can also choose to read and load the two parts of the data separately as the training set and the test set.

Each directory contains a bunch of csv files, examples are as follows:

The content of the sample data is as follows:

9,11,19,1.1879e+05,0.059,-0.372
9,11,19,1.1883e+05,0.603,-0.085
9,11,19,1.1887e+05,0.613,0.112
9,11,19,1.1891e+05,0.465,0.185
9,11,19,1.1894e+05,-0.216,-0.256
9,11,19,1.1898e+05,-0.806,0.177
9,11,19,1.1902e+05,-0.653,0.113
9,11,19,1.1906e+05,-0.007,0.398
9,11,19,1.191e+05,0.888,0.145
9,11,19,1.1914e+05,1.037,-0.542
9,11,19,1.1918e+05,0.299,-0.201
9,11,19,1.1922e+05,-0.552,-0.022
9,11,19,1.1926e+05,-1.237,0.264
9,11,19,1.193e+05,-1.059,0.155
9,11,19,1.1934e+05,-0.269,0.163
9,11,19,1.1937e+05,0.662,0.269
9,11,19,1.1941e+05,0.949,0.078
9,11,19,1.1945e+05,0.403,-0.065
9,11,19,1.1949e+05,-0.279,-0.411
9,11,19,1.1953e+05,-0.856,0.033
9,11,19,1.1957e+05,-0.736,0.201
9,11,19,1.1961e+05,0.098,0.326
9,11,19,1.1965e+05,0.718,-0.183
9,11,19,1.1969e+05,0.61,-0.038
9,11,19,1.1973e+05,0.201,0.092
9,11,19,1.1976e+05,-0.3,0.01
9,11,19,1.198e+05,-0.378,0.447
9,11,19,1.1984e+05,0.149,-0.189
9,11,19,1.1988e+05,0.499,-0.421
9,11,19,1.1992e+05,0.325,0.024
9,11,19,1.1996e+05,-0.265,0.49
9,11,19,1.2e+05,-0.708,0.487
9,11,19,1.2004e+05,-0.443,0.157
9,11,19,1.2008e+05,-0.042,-0.437
9,11,19,1.2012e+05,0.238,-0.025
9,11,19,1.2016e+05,0.46,0.193
9,11,19,1.202e+05,0.192,0.036
9,11,19,1.2023e+05,-0.093,0.118
9,11,19,1.2027e+05,-0.344,0.148
9,11,19,1.2031e+05,-0.174,0.117
9,11,19,1.2035e+05,-0.029,-0.026
9,11,19,1.2039e+05,0.026,0.469

For data processing, you can directly use the official code:

%% 批量读取IEEE PHM 2012轴承全寿命数据
clc
clear all
close all
%% 文件夹路径
file_path =  'Learning_set/';
%% 全寿命振动信号
csv_acc_path_list = dir(strcat(file_path,'acc*.csv'));
csv_acc_num = length(csv_acc_path_list);%获取文件总数量
if csv_acc_num > 0 %有满足条件的文件
        for j = 1:csv_acc_num %逐一读取文件
            csv_acc_name = csv_acc_path_list(j).name;% 文件名
            csv_acc =  csvread(strcat(file_path,csv_acc_name));
            csv_acc_data(:,:,j)=csv_acc;
            fprintf('%d %d %s\n',csv_acc_num,j,strcat(file_path,csv_acc_name));% 显示正在处理的文件名
        end
end
% 合并矩阵 时间*通道
channel=6;   %信号的通道数
csv_acc_data_change=permute(csv_acc_data,[2 1 3]);
csv_acc_data=reshape(csv_acc_data_change,channel,prod(size(csv_acc_data))/channel)';

%% 全寿命温度信号
csv_temp_path_list = dir(strcat(file_path,'temp*.csv'));%获取该文件夹中所有csv格式的文件
csv_temp_num = length(csv_temp_path_list);%获取文件总数量
delimiter = ',';
formatSpec = '%s%s%s%s%s%s%[^\n\r]';
if csv_temp_num > 0 %有满足条件的文件
        for j = 1:csv_temp_num %逐一读取文件
            csv_temp_name = csv_temp_path_list(j).name;% 文件名
            csv_temp_fileID = fopen(strcat(file_path,csv_temp_name),'r');
            csv_temp = textscan(csv_temp_fileID, formatSpec, 'Delimiter', delimiter);
            for i=1:size(csv_temp{1,1},1)
                csv_temp_data(i,:,j)=str2num(csv_temp{1,1}{i,1})';
            end
            fprintf('%d %d %s\n',csv_temp_num,j,strcat(file_path,csv_temp_name));% 显示正在处理的文件名
            fclose(csv_temp_fileID);
        end
end

% 合并矩阵 时间*通道
channel=5;   %信号的通道数
csv_temp_data_change=permute(csv_temp_data,[2 1 3]);
csv_temp_data=reshape(csv_temp_data_change,channel,prod(size(csv_temp_data))/channel)';

%% 全寿命振动信号和温度信号的时域图
clearvars -except csv_acc_data csv_temp_data
figure;subplot 211;plot(csv_acc_data(:,5));title('水平振动信号');
       subplot 212;plot(csv_acc_data(:,6));title('竖直振动信号');
figure;plot(csv_temp_data(:,5));title('温度信号')

I don't know the data background very well here, so I won't explain it much. If you have knowledge, you can share it.

In order to analyze the data intuitively, it is visualized as follows:

After the data processing is completed, the required model can be built. Here, many different models have been developed for the purpose of experimentation: CNN, CNN-LSTM, CNN-GRU, CNN-LSTM-ATTENTION, CNN-GRU-ATTENTION, etc. , here is a simple look at the actual structure of CNN as an example: