Rolling bearing life prediction method based on auto-encoder and random forest regression

Abstract

The invention discloses a rolling bearing service life prediction method based on an auto-encoder and random forest regression, and relates to the field of rolling bearing service life evaluation, and the method specifically comprises the steps: building a traction motor bearing state evaluation model through employing the actual operation data of a motor train unit line; establishing a residual life prediction model by using the traction motor bearing full life cycle data; through state evaluation, the starting time point of prediction of the residual service life of the bearing is determined, and the prediction error is reduced; a prediction model is improved, and the change condition of the remaining service life of the bearing in a period of time in the future is achieved. The prediction result can accurately represent the change condition of the remaining service life of the bearing in a future period of time, sufficient time can be provided for bearing maintenance, and the operation risk of the motor train unit is reduced.

Description

Technical Field

[0001] The present invention relates to the field of life assessment of rolling bearings, and in particular to a method for predicting the life of rolling bearings based on autoencoders and random forest regression. Background Art

[0002] Rolling bearings are key mechanical components of traction motors and are also the mechanical components with the highest failure rate. In order to accurately predict the degradation trend of rolling bearings and their remaining service life and reduce the risk of bearing use, it is necessary to evaluate or predict their service life. In the prior art, there are many methods for predicting the life of rolling bearings, such as the rolling bearing remaining service life prediction method using encoders and decoders with publication number CN114925723B, but the defect of this method is that there is no clear prediction starting point, and the prediction timing is wrong, resulting in a large prediction error, so it cannot be used. At the same time, there is also a rolling bearing remaining service life prediction method based on deep learning and particle filtering with publication number CN116204766A and a rolling bearing remaining service life prediction method based on CNN and LSTM with publication number CN109726524B, but the prediction results of these two prediction methods only represent the current state, and there is no remaining life prediction result in the future period of time, which is limited in practical application. Therefore, it is necessary to design a new method or improve the existing method to solve the above problems. Summary of the invention

[0003] In order to solve the problems in the prior art of rolling bearing life prediction that the prediction starting point is unclear or only indicates the current state, the prediction result has large errors or the prediction results are incomplete, the present invention provides an improved rolling bearing life prediction method based on autoencoder and random forest regression.

[0004] The present invention is realized by the following technical scheme: a rolling bearing life prediction method based on autoencoder and random forest regression, including three parts: constructing a bearing state evaluation model, constructing a bearing remaining life prediction model, and bearing state evaluation and life prediction:

[0005] S1: Constructing a bearing condition assessment model:

[0006] S1-1: Form the data set required for model training:

[0007] S1-1-1: Feature extraction:

[0008] The vibration and temperature data of the traction motor running online are used to extract features. For each vibration signal sequence x(N)=(x1,x2,...,x N ), N is the signal length, the vibration signal has m segments, and 15 time domain statistical features T1~T are extracted 15, and extract 7 frequency domain features F1~F7, T1~T 15 And the characteristic expressions of F1~F7 are as follows:

[0009] average value:

[0010] RMS:

[0011] Square root amplitude:

[0012] Absolute average:

[0013] Maximum value: T5 = max(x i )

[0014] Minimum value: T6 = min(x i )

[0015] Peak-to-peak value: T7 = max(x i )-min(x i )

[0016] variance:

[0017] Skewness:

[0018] Kurtosis:

[0019] Waveform indicators:

[0020] Peak indicators:

[0021] Pulse indicator:

[0022] Margin index:

[0023] Skewness indicator:

[0024] Spectral mean:

[0025] Spectral variance:

[0026] Spectral skewness:

[0027] Spectral Kurtosis:

[0028] Frequency Center of Gravity:

[0029] Frequency standard deviation:

[0030] RMS frequency:

[0031] Where s(K) is the FFT spectrum amplitude of the original time domain signal segment x(N), K is the number of spectral lines, N = 2K, k = 1, 2, ..., K-1, K, f k is the frequency value of the kth spectral line;

[0032] The characteristic value of the jth vibration signal is X j ={T1~T 15 ,F1~F7,T e}, T e is the absolute temperature value corresponding to the vibration signal of this section; after all the characteristic indicators of the vibration data are calculated, the characteristic data set X={X1,X2,…X j ,…X m};

[0033] S1-1-2: Normalization of feature dataset:

[0034] Normalize X by feature class, and for one of the features y, set its maximum value to y max , the minimum value is y min , then the normalized feature is expressed as:

[0035]

[0036] S1-1-3: Training set data organization:

[0037] Process the normalized features of the dataset X to obtain the training dataset X train And the test dataset X test ;

[0038] S1-2: Constructing an autoencoding neural network model:

[0039] S1-2-1: Take training set X train Construct the input layer input_layer as input:

[0040] S1-2-2: Set the number and dimension of encoders encoding_dim, and construct encoder encoder with input_layer as input;

[0041] S1-2-3: Set the number and dimension of decoders decoding_dim, and construct a decoder decoder using encoder encoder as input;

[0042] S1-2-4: Construct the output layer output_layer with decoder as input;

[0043] S1-2-5: Form and compile the autoencoder neural network model autoencoder;

[0044] S1-3: Model training and storage:

[0045] S1-3-1: Set the model's save node checkpointer;

[0046] S1-3-2: Set the number of nodes for training num_epoch;

[0047] S1-3-3: Set the batch_size of training times for each node;

[0048] S1-3-4: Use X train 、num_epoch、batch_size、X test , checkpointer trains the constructed autoencoder model autoencoder, saves the trained model in checkpointer, and judges whether the model meets the requirements based on the calculated mean adv and standard deviation std of the loss function. If it does not meet the requirements, reset the model parameters for training until the requirements are met, and save the training model and the mean adv and standard deviation std of the loss function;

[0049] S2: Construct a bearing remaining life prediction model:

[0050] S2-1: Full life vibration temperature data processing:

[0051] From the full life vibration data of the bearing, extract the time domain, frequency domain features and absolute temperature value of each vibration data sample in the same way as S1-1-1, and finally form the feature Y j ={T1~T 15 ,F1~F7,T e}, T e is the absolute temperature value corresponding to the vibration signal, and then the obtained characteristic data set is normalized in the same way as S1-1-2;

[0052] S2-2: Constructing an autoencoding neural network model:

[0053] Construct a random forest regression model as the bearing life prediction model RandomForestRegressor:

[0054] S2-2-1: Set the number of trees in the forest n_estimators;

[0055] S2-2-2: Set the function criterion for measuring segmentation quality;

[0056] S2-2-3: Set the maximum depth of the tree max_depth;

[0057] S2-2-4: Set the minimum number of samples required to split an internal node min_samples_split;

[0058] S2-2-5: Set the minimum number of samples required for leaf nodes min_samples_leaf;

[0059] S2-2-6: Set the minimum weighted fraction min_weight_fraction_leaf of the sum of weights at the leaf node;

[0060] S2-2-7: Set max_features, the number of features to consider when finding the best split;

[0061] S2-2-8: Set the verbose level to control the fitting and prediction;

[0062] S2-3: Model training and saving:

[0063] S2-3-1: Perform PCA analysis on the normalized feature data set, set the principal component score threshold per, select the fusion feature Y with a principal component score greater than per, set the step size of the prediction model, combine the feature data into a new data set, and then split the data set into the training set Y train and the test set Y test , with training set Y train As input, train the prediction model RandomForestRegressor;

[0064] S2-3-2: Use test set Y test Test the model accuracy, set the error threshold Error, and stop training when the actual training error is less than Error and save the model and the model's maximum output value p max , otherwise return to S2-3-1 to retrain the model;

[0065] S3: Bearing condition assessment and life prediction:

[0066] S3-1: Evaluation of the current state of the bearing. The current vibration and temperature data are used as input to evaluate the current state of the rolling bearing, as follows:

[0067] S3-1-1: Use the method of S1-1 to extract features from the vibration and temperature data of the current traction motor and normalize them to obtain the current feature vector Xt ;

[0068] S3-1-2: Use the model saved in S1-3 to generate t As input, get the loss function value loss;

[0069] S3-1-3: Determine the state S of the rolling bearing of the traction motor through the relationship between the loss function value loss and adv and std. 0 indicates normal and 1 indicates abnormal:

[0070]

[0071] S3-2: Bearing remaining life prediction:

[0072] If the result of S1-1-3 evaluation is normal, the remaining life prediction is not performed and the remaining life of the bearing is 100%. When the result of S1-1-3 evaluation is abnormal, the remaining life of the rolling bearing is predicted by taking step vibration and temperature data samples as input;

[0073] S3-2-1: Calculate feature X according to the method of S2-2 k and normalized;

[0074] S3-2-2: According to the method of S2-3, the normalized features are processed by PCA to obtain a new data set Z = {Z|z k-step+1 ,z k-step+2 ,...,z k}, take Z as input and use the model RandomForestRegressor to predict Z k+1 , calculate Z k+1 The corresponding remaining bearing life is

[0075]

[0076] S3-2-3: Repeat S3-2-2, set the number of points n corresponding to the prediction duration, use the model RandomForestRegressor to predict the feature Zk+n at a certain moment, and the model input is Z = {Z|z k-step+n ,z k-step+n+1 ,*,z k+n-1}, the remaining bearing life corresponding to Zk+n is calculated as:

[0077]

[0078] The present invention uses the actual operation data of the EMU line to establish a model for evaluating the state of the traction motor bearing; uses the full life cycle data of the traction motor bearing to establish a remaining life prediction model; through state evaluation, the starting time point for the prediction of the remaining service life of the bearing is determined, the prediction error is reduced, and the prediction of the change of the remaining service life of the bearing in a period of time in the future is realized.

[0079] Compared with the prior art, the present invention has the following beneficial effects: the present invention provides a rolling bearing life prediction method based on autoencoder and random forest regression, and the prediction result can more accurately represent the change of the remaining service life of the bearing in the future, which can provide sufficient time for bearing maintenance and reduce the operation risk of EMUs. BRIEF DESCRIPTION OF THE DRAWINGS

[0080] Figure 1 This is the remaining life result of the multi-step prediction of a specific embodiment of the present invention. DETAILED DESCRIPTION

[0081] The present invention is further described below in conjunction with specific embodiments.

[0082] The CR400BF EMU train traction motor bearing NU214 cylindrical roller bearing is used as the instance object, and Python is selected as the development platform.

[0083] A rolling bearing life prediction method based on autoencoder and random forest regression includes three parts: building a bearing condition assessment model, building a bearing remaining life prediction model, and bearing condition assessment and life prediction:

[0084] S1: Constructing a bearing condition assessment model:

[0085] S1-1: Form the data set required for model training: Use 414 samples of vibration and temperature data of the traction motor running online to calculate the time domain and frequency domain characteristics of the vibration data. When splitting the data set, take the training data set X train There are 280 samples in total, and the test data set has X test 134 samples;

[0086] S1-2: Constructing an autoencoding neural network model:

[0087] S1-2-1: Use "tensorflow.keras.layers.Input" to construct the input layer input_layer with 280 layers.

[0088] S1-2-2: There are 4 encoders in total, and the dimensions of each encoder are 4, 8, 16, and 32 respectively. The settings of each encoder are as follows: the first encoder encoder1: the input is input_layer, the activation function selects "tensorflow.nn.swish", the bias selects "tensorflow.keras.regularizers.l1", and the parameter is set to (10e-5); the inputs of the second encoder encoder2, the third encoder encoder3, and the fourth encoder encoder4 are encoder1, encoder2, and encoder3 respectively, and the remaining parameters are selected the same as encoder1.

[0089] S1-2-3: There are 3 decoders in total, and the dimensions of each decoder are 32, 16, and 8 respectively. The settings of each decoder are as follows; the inputs of the first decoder decoder1, the second decoder decoder2, and the third decoder decoder3 are encoder4, decoder1, and decoder2 respectively, and the activation function is "tensorflow.nn.swish".

[0090] S1-2-4: The input of the output layer outputs is decoder3, and the activation function selects the "relu" function.

[0091] S1-2-5: Select input_layer as input and outputs as output, use the "tensorflow.keras.Model" module to build the autoencoder model and compile it. Use the "compile" function, select "adam" for the optimizer of the compilation function, select "mean_squared_error" for the loss function, and select "mae" for the indicator.

[0092] S1-3: Model training and storage:

[0093] S1-3-1: Save the checkpointer node parameters as follows: the file name is set to AE_model.h5, the redundancy is set to 0, and only the best model is saved.

[0094] S1-3-2: Use "autoencoder.fit" to train the model with the following parameters: select Xtrain for both input and output data, num_epoch for training nodes = 200, batch_size for training each node = 24, select Xtest for both input and output of validation data, set redundancy to 1, and select checkpointer for callback point. The mean adv of the loss function obtained from training is 0.234 and the standard deviation std is 0.003, adv>6std, and the model meets the requirements.

[0095] S2: Construct a bearing remaining life prediction model:

[0096] S2-1: Full life vibration and temperature data processing: A total of 1528 bearing full life vibration and temperature data samples are used to calculate the characteristics, normalize and PCA analysis to obtain Y. The step length of the prediction model is set to 500 to obtain the training data set Y. train ={Y1,Y2,…,Y 1000}, Y1=Y(i),i=1,2,…,500; Y2=Y(j),j=2,3,…,501; Y 1000 =Y(k), k = 1000, 1001, …, 1049; the label set is Y(l), l = 501, 502, …, 1528.

[0097] S2-2: Construct model RandomForestRegressor:

[0098] S2-2-1: The number of trees in the forest n_estimators is 20.

[0099] S2-2-2: The function criterion for measuring segmentation quality is “absolute error”.

[0100] S2-2-3: The maximum depth of the tree max_depth is set to "false".

[0101] S2-2-4: The minimum number of samples required to split an internal node, min_samples_split, is 3.

[0102] S2-2-5: The minimum number of samples required for a leaf node, min_samples_leaf, is 1.

[0103] S2-2-6: The minimum weighted fraction min_weight_fraction_leaf of the sum of weights at the leaf node is 0.1.

[0104] S2-2-7: The number of features to consider when finding the best split is max_features, which is set to "auto".

[0105] S2-2-8: Control the level of detail during fitting and prediction. Set verbose to 1.

[0106] S2-3: Using the above training set and label set data as input, use "RandomForestRegressor" to train the model.

[0107] S2-4: Model prediction and storage: test set data Y test = {Y 1001 ,Y 1002 ,…,Y 1028},Y 1001 =Y(i),i=1001,1002,…,1500; Y 1002 =Y(j),j=1002,1003,…,1501; Y 1028 =Y(k), k=1028,1030,…,1527; the label set is Y(m), m=1501,1502,…,1528. The absolute error of the result obtained compared with the label set is 0.002, which is less than the training error threshold of 0.01, meeting the requirements. The model RandomForestRegressor and the maximum value p of the model output are saved. max =1.82.

[0108] S3: Bearing condition assessment and life prediction:

[0109] Input the vibration and temperature data of the current bearing, extract the features and normalize them to get the current feature vector X t , with X t As input, the loss function value loss is 0.224. According to formula (2), the result of evaluation is abnormal. The 500 sets of vibration and temperature data samples up to the current time are successively subjected to feature extraction, normalization and PCA processing. The processed data Y test Input the prediction model, substitute the output value into formula (4) to calculate that the current remaining service life is 96%, and the remaining service life is 0 at the 410th point, such as Figure 1 shown.

[0110] The scope of protection claimed by the present invention is not limited to the above specific embodiments. For those skilled in the art, the present invention may have various deformations and changes. Any modifications, improvements and equivalent substitutions made within the concept and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A rolling bearing life prediction method based on autoencoder and random forest regression, characterized in that: It includes three parts: building a bearing condition assessment model, building a bearing remaining life prediction model, and bearing condition assessment and life prediction: S1: Building a bearing condition assessment model: S1-1: Form the data set required for model training: S1-1-1: Feature extraction: The vibration and temperature data of the traction motor running online are used to extract features. For each vibration signal sequence x(N)=(x1,x2,...,x N ), N is the signal length, the vibration signal has m segments, and 15 time domain statistical features T1~T are extracted 15 , and extract 7 frequency domain features F1~F7, T1~T 15 And the characteristic expressions of F1~F7 are as follows: average value: RMS: Square root amplitude: Absolute average: Maximum value: T5 = max(x i ) Minimum value: T6 = min(x i ) Peak-to-peak value: T7 = max(x i )-min(x i ) variance: Skewness: Kurtosis: Waveform indicators: Peak indicators: Pulse indicator: Margin index: Skewness indicator: Spectral mean: Spectral variance: Spectral skewness: Spectral Kurtosis: Frequency Center of Gravity: Frequency standard deviation: RMS frequency: Where s(K) is the FFT spectrum amplitude of the original time domain signal segment x(N), K is the number of spectral lines, N = 2K, k = 1, 2, ..., K-1, K, f k is the frequency value of the kth spectral line; The characteristic value of the jth vibration signal is X j ={T1~T 15 ,F1~F7,T e }, T e is the absolute temperature value corresponding to the vibration signal of this section; after all the characteristic indicators of the vibration data are calculated, the characteristic data set X={X1,X2,…X j ,…X m }; S1-1-2: Normalization of feature data set: Normalize X by feature class, and for one of the features y, set its maximum value to y max , the minimum value is y min , then the normalized feature is expressed as: S1-1-3: Training set data organization: Process the normalized features of the dataset X to obtain the training dataset X train And the test dataset X test ; S1-2: Constructing an autoencoding neural network model: S1-2-1: Take training set X train Construct the input layer input_layer as input: S1-2-2: Set the number and dimension of encoders encoding_dim, and construct encoder encoder with input_layer as input; S1-2-3: Set the number and dimension of decoders decoding_dim, and construct a decoder decoder using the encoder encoder as input; S1-2-4: Construct the output layer output_layer with decoder as input; S1-2-5: Form and compile the autoencoder neural network model autoencoder; S1-3: Model training and storage: S1-3-1: Set the model's save node checkpointer; S1-3-2: Set the number of nodes for training num_epoch; S1-3-3: Set the batch_size of training times for each node; S1-3-4: Use X train 、num_epoch、batch_size、X test , checkpointer trains the constructed autoencoder model autoencoder, saves the trained model in checkpointer, and judges whether the model meets the requirements based on the calculated mean adv and standard deviation std of the loss function. If it does not meet the requirements, reset the model parameters for training until the requirements are met, and save the training model and the mean adv and standard deviation std of the loss function; S2: Construct a bearing remaining life prediction model: S2-1: Full life vibration temperature data processing: From the full life vibration data of the bearing, extract the time domain, frequency domain features and absolute temperature value of each vibration data sample in the same way as S1-1-1, and finally form the feature Y j ={T1~T 15 ,F1~F7,T e }, T e is the absolute temperature value corresponding to the vibration signal, and then the obtained characteristic data set is normalized in the same way as S1-1-2; S2-2: Constructing an autoencoding neural network model: Construct a random forest regression model as the bearing life prediction model RandomForestRegressor: S2-2-1: Set the number of trees in the forest n_estimators; S2-2-2: Set the function criterion for measuring segmentation quality; S2-2-3: Set the maximum depth of the tree max_depth; S2-2-4: Set the minimum number of samples required to split an internal node min_samples_split; S2-2-5: Set the minimum number of samples required for leaf nodes min_samples_leaf; S2-2-6: Set the minimum weighted fraction min_weight_fraction_leaf of the sum of weights at the leaf node; S2-2-7: Set max_features, the number of features to consider when finding the best split; S2-2-8: Set the verbose level to control the fitting and prediction; S2-3: Model training and saving: S2-3-1: Perform PCA analysis on the normalized feature data set, set the principal component score threshold per, select the fusion feature Y with a principal component score greater than per, set the step size of the prediction model, combine the feature data into a new data set, and then split the data set into the training set Y train and the test set Y test , with training set Y train As input, train the prediction model RandomForestRegressor; S2-3-2: Use test set Y test Test the model accuracy, set the error threshold Error, and stop training when the actual training error is less than Error and save the model and the model's maximum output value p max , otherwise return to S2-3-1 to retrain the model; S3: Bearing condition assessment and life prediction: S3-1: Evaluation of the current state of the bearing. The current vibration and temperature data are used as input to evaluate the current state of the rolling bearing, as follows: S3-1-1: Use the method of S1-1 to extract features from the vibration and temperature data of the current traction motor and normalize them to obtain the current feature vector X t ; S3-1-2: Use the model saved in S1-3 to generate t As input, get the loss function value loss; S3-1-3: Determine the state S of the rolling bearing of the traction motor through the relationship between the loss function value loss and adv and std. 0 indicates normal and 1 indicates abnormal: S3-2: Bearing remaining life prediction: If the result of S1-1-3 evaluation is normal, the remaining life prediction is not performed and the remaining life of the bearing is 100%. When the result of S1-1-3 evaluation is abnormal, the remaining life of the rolling bearing is predicted by taking step vibration and temperature data samples as input; S3-2-1: Calculate feature X according to the method of S2-2 k and normalized; S3-2-2: According to the method of S2-3, the normalized features are processed by PCA to obtain a new data set Z = {Z|z k-step+1 ,z k-step+2 ,...,z k }, take Z as input and use the model RandomForestRegressor to predict Z k+1 , calculate Z k+1 The corresponding remaining bearing life is S3-2-3: Repeat S3-2-2, set the number of points n corresponding to the prediction duration, use the model RandomForestRegressor to predict the feature Zk+n at a certain moment, and the model input is Z = {Z|z k-step+n ,z k-step+n+1 ,...,z k+n-1 }, the remaining bearing life corresponding to Zk+n is calculated as:

Images