CN114330141A - A prediction method of fan main shaft bearing life based on GRU hyperparameter optimization - Google Patents

Abstract

The invention discloses a fan main shaft bearing life prediction method based on GRU hyperparameter optimization, and belongs to the technical field of wind power main bearings. The specific steps of the prediction method are as follows: (1) searching for optimal parameters; (2) bearing life prediction; the invention The long-term iteration method is used to train the prediction model, and the test set is input into the trained model, and the remaining life prediction curve of the bearing is drawn and analyzed, which can improve the accuracy of the diagnosis model and the efficiency of manual parameter search, and does not require manual Set parameters and do not need manual modeling, just input the bearing vibration signal directly into the model to predict the current remaining life of the bearing, making the operation process simple and easy to operate.

Description

A prediction method of fan spindle bearing life based on GRU hyperparameter optimization

technical field

The invention relates to the technical field of wind power main bearings, in particular to a method for predicting the life of a wind turbine main shaft bearing based on GRU hyperparameter optimization.

Background technique

The prediction of the remaining life of the fan main shaft bearing refers to the use of a suitable prediction model to determine the remaining time for the safe and economical operation of the fan according to the current health status of the fan. At present, the commonly used life prediction methods mainly include prediction methods based on failure physical models and data-based There are two driving prediction methods. Generally speaking, the structure of the main shaft bearing of the fan is very complicated. Considering the reasons such as the changeable operation state of the fan and the unclear failure mechanism, it is usually difficult to build a single failure physical model of the equipment. Therefore, through data-driven It is a relatively feasible way to analyze the vibration data of the fan main shaft bearing, mine the information related to the equipment performance and then predict the remaining life. Although the traditional neural network can predict the bearing life, the premise is that there is a difference between the data. Independent distribution, it is not suitable for sequence problems with dependencies between data, and the algorithms used in bearing life prediction require a large amount of test data as support. Redundant calculation; therefore, it becomes particularly important to invent a method for predicting the life of fan main shaft bearings based on GRU hyperparameter optimization;

After retrieval, Chinese Patent No. CN201810898506.7 discloses a wind power main bearing fault prediction and life evaluation system and method. Although the invention ensures that the main bearing is always in a good lubricated state and prolongs the bearing life, the diagnostic model Therefore, we propose a fan spindle bearing life prediction method based on GRU hyperparameter optimization.

SUMMARY OF THE INVENTION

The purpose of the present invention is to solve the defects existing in the prior art, and propose a method for predicting the life of a fan main shaft bearing based on GRU hyperparameter optimization.

In order to achieve the above object, the present invention adopts the following technical solutions:

A method for predicting the life of a fan main shaft bearing based on GRU hyperparameter optimization. The specific steps of the prediction method are as follows:

(1) Find the optimal parameters: Divide the original samples into a training set and a test set according to the pre-set rules, train the model through the training samples, and then use the test samples to verify the trained model to evaluate the model For small data samples, the leave-one-out method is used for cross-validation;

(2) Bearing life prediction: Collect the bearing vibration signal, determine the parameter selection of the GRU neural network, input the bearing vibration signal, and conduct a prediction experiment on the remaining service life of the bearing.

As a further solution of the present invention, the specific steps of cross-validation described in step (1) are as follows:

Step 1: Select one observation data from N groups of observation data sets as validation data;

Step 2: Use the remaining observation data to fit a test model, and use the first excluded observation to verify the accuracy of the test model, and calculate the predictive ability of the prediction model through the root mean square error, and repeat n Second-rate;

Step 3: Parameter optimization is performed on the generated precision parameters.

As a further scheme of the present invention, the specific calculation formula of the root mean square error described in step 2 is as follows:

Among them, E(y _i ) represents the ith actual observed value, _yi is the ith predicted value obtained by the model inversion, and n is the total number of observation samples.

As a further scheme of the present invention, the specific steps of the optimization process described in step 3 are as follows:

The first step: initialize the parameter range, and set the learning rate η = [0.0001, 0.1], the step size is 0.0001;

The second step: establish a data sample, and list all possible data results at the same time, a total of 1000 sets of data;

Step 3: Divide the samples. For each set of data, select any subset as the test set, and the remaining 999 subsets as the training set. After training the model, make predictions on the test set, and count the root mean square error of the test results;

Step 4: Find the optimal parameter combination, at the same time replace the test set with another subset, then take the remaining 999 subsets as the training set, and count the root mean square error again, until all 1000 sets of data are predicted once. The combination parameter corresponding to the minimum RMSE is selected as the optimal parameter in the data interval.

As a further scheme of the present invention, the specific steps of the prediction experiment described in step (2) are as follows:

S1: Collect bearing vibration signals in real time through sensors, preprocess the vibration signals at the same time, and extract characteristic parameters through time domain and frequency domain methods;

S2: Screen out the characteristic parameters that can represent the bearing degradation information, and screen out the characteristic parameters with poor representation ability;

S3: Set the bearing sample life label, and the label is set to the normalized value of the remaining service life of the bearing corresponding to the current sample;

S4: Divide the bearing vibration data samples into a training set and a test set, and standardize the training set;

S5: Send the training samples to the GRU network model, set the specific parameters of the model, train the prediction model using the long-term iteration method, input the test set into the trained model, draw the bearing remaining life prediction curve, and analyze it.

As a further scheme of the present invention, the specific calculation steps of the normalized value described in S3 are as follows:

Among them, rul _max represents the remaining life prediction threshold of the bearing, rul _real represents the actual remaining service life of the bearing; rul represents the remaining service life of the current sample after trimming; rul _t represents the normalized life value of the sample data.

As a further scheme of the present invention, the specific calculation formula of the standardized processing described in S4 is as follows:

Among them, x represents the proposed feature parameters; mean(x) represents the average processing of the proposed feature parameters; std(x) represents the standard deviation of the feature parameters.

Compared with the prior art, the beneficial effects of the present invention are:

1. In this GRU-based hyperparameter optimization method for predicting the life of the fan spindle bearing, the computer collects the bearing vibration signal in real time through the sensor, preprocesses the vibration signal, and extracts the characteristic parameters through the time domain and frequency domain methods. The characteristic parameters that can represent the bearing degradation information, and filter out the characteristic parameters with poor representation ability, set the bearing sample life label as the normalized value of the remaining service life of the bearing corresponding to the current sample, and divide the bearing vibration data samples into Training set and test set, standardize the training set at the same time, send the training samples to the GRU network model, set the specific parameters of the model, use the long-term iteration method to train the prediction model, and input the test set into the trained model. , draw the bearing residual life prediction curve and analyze it, which can improve the accuracy of the diagnostic model and the efficiency of manual parameter search. At the same time, it does not require manual parameter setting and manual modeling, and only needs to directly input the bearing vibration signal into the model. The current remaining life of the bearing can be predicted, making the operation process simple and easy to operate.

Description of drawings

The accompanying drawings are used to provide a further understanding of the present invention, and constitute a part of the specification, and are used to explain the present invention together with the embodiments of the present invention, and do not constitute a limitation to the present invention.

FIG. 1 is a flow chart of a method for predicting the life of a fan main shaft bearing based on GRU hyperparameter optimization proposed by the present invention.

Detailed ways

Referring to Figure 1, a method for predicting the life of a fan main shaft bearing based on GRU hyperparameter optimization, the prediction method mainly includes two stages:

The first stage: find the optimal parameters through the leave-one-out cross-validation algorithm;

The second stage: determine the parameter selection of the GRU neural network, input the bearing vibration signal, and predict the remaining service life of the bearing.

Find the optimal parameters: Divide the original samples into training sets and test sets according to the pre-set rules, train the model through the training samples, and then use the test samples to verify the trained model to evaluate the accuracy of the model , for small data samples, the leave-one-out method is used for cross-validation.

Specifically, select one observation from the N groups of observation data as the validation data, use the remaining observation data to fit a test model, and use the first excluded observation to verify the accuracy of the test model, and pass the root mean square The error is used to calculate the predictive ability of the predictive model, which is repeated n times, and the parameter optimization process is performed on the generated precision parameters.

It should be further explained that the specific calculation formula of the root mean square error is as follows:

Among them, E(y _i ) represents the ith actual observed value, _yi is the ith predicted value obtained by the model inversion, and n is the total number of observation samples.

It should be further explained that the specific steps of the optimization process are as follows: First, initialize the parameter range, set the learning rate η=[0.0001, 0.1], the step size is 0.0001, establish the data sample, and list all possible data results at the same time, a total of 1000 For each group of data, any subset is selected as the test set, and the remaining 999 subsets are used as the training set. After training the model, the test set is predicted, and the root mean square error of the test results is counted.

Bearing life prediction: Collect the bearing vibration signal, determine the parameter selection of the GRU neural network, input the bearing vibration signal, and conduct a prediction experiment on the remaining service life of the bearing.

Specifically, first, the computer collects the bearing vibration signal in real time through the sensor, and preprocesses the vibration signal at the same time, and extracts the characteristic parameters through the time domain and frequency domain methods, and filters out the characteristic parameters that can represent the bearing degradation information. Except for the characteristic parameters with poor characterization ability, set the bearing sample life label. The label is set to the normalized value of the remaining bearing service life corresponding to the current sample. After the setting is completed, the bearing vibration data samples are divided into training set and test set. Standardize the training set, send the training samples to the GRU network model, set the specific parameters of the model by yourself, train the prediction model using the long-term iteration method, input the test set into the trained model, and draw the remaining bearing life. Predict curves and analyze them.

It should be further explained that the specific calculation steps of the normalized value are as follows:

Among them, rul _max represents the predicted threshold value of the remaining life of the bearing, rul _real represents the actual remaining service life of the bearing; rul represents the remaining service life of the current sample after cutting; rul _t represents the normalized life value of the sample data;

The specific calculation formula for normalization is as follows:

Among them, x represents the proposed feature parameters; mean(x) represents the average processing of the proposed feature parameters; std(x) represents the standard deviation of the feature parameters, and the parameter optimization algorithm is used to calculate the error of each group of parameters. After several hours of optimization model training, the best model parameters are obtained, which improves the accuracy of the diagnostic model and the efficiency of manual parameter search. In addition, because there is no need to manually set parameters and no manual modeling is required, only the bearing vibration signal needs to be directly Input into the model can predict the current remaining life of the bearing, making the operation process simple and easy to operate.

Claims (7)

1. A method for predicting the service life of a fan main shaft bearing based on GRU (generalized regression Unit) super-parameter optimization is characterized by comprising the following specific steps:

(1) finding the optimal parameters: dividing original samples into a training set and a testing set according to a preset rule, training a model through the training samples, verifying the trained model through the testing samples, evaluating the accuracy of the model, and performing cross verification on small data samples by adopting a leave-one-out method;

(2) and (3) bearing life prediction: and acquiring a bearing vibration signal, determining the parameter selection of the GRU neural network, inputting the bearing vibration signal, and performing a prediction experiment on the residual service life of the bearing.

2. The method for predicting the life of the main shaft bearing of the wind turbine based on GRU hyper-parameter optimization according to claim 1, wherein the cross validation in the step (1) comprises the following specific steps:

the method comprises the following steps: selecting one observation data from the N groups of observation data sets as verification data;

step two: fitting a test model by using the rest observation data, verifying the precision of the test model by using the observation value which is excluded firstly, calculating the prediction capability of the prediction model by the root mean square error, and repeating the operation for n times;

step three: and performing parameter optimization processing on the generated precision parameters.

3. The method for predicting the life of the main shaft bearing of the fan based on GRU over-parameter optimization according to claim 2, wherein the specific calculation formula of the root mean square error in the second step is as follows:

wherein, E (y)_i) Represents the ith actual observed value, y_iFor the ith predictor of the model inversion, n is the total number of observed samples.

4. The method for predicting the life of the main shaft bearing of the fan based on GRU hyper-parameter optimization according to claim 2, wherein the optimization in the third step specifically comprises the following steps:

the first step is as follows: initializing a parameter range, and enabling a learning rate eta to be [0.0001, 0.1] and a step length to be 0.0001;

the second step is that: establishing a data sample, and listing all possible data results at the same time, wherein the total data is 1000 groups;

the third step: dividing samples, selecting any subset as a test set and the rest 999 subsets as training sets for each group of data, predicting the test set after training the model, and counting the root mean square error of the test result;

the fourth step: and (3) solving an optimal parameter combination, simultaneously replacing the test set with another subset, taking the remaining 999 subsets as training sets, counting the root mean square error again until 1000 groups of data are predicted once, and selecting the corresponding combination parameter when the RMSE is minimum as the optimal parameter in the data interval.

5. The method for predicting the life of the main shaft bearing of the fan based on GRU hyper-parameter optimization according to claim 1, wherein the prediction experiment in the step (2) comprises the following specific steps:

s1: acquiring a bearing vibration signal in real time through a sensor, simultaneously carrying out preprocessing work on the vibration signal, and extracting characteristic parameters through a time domain and frequency domain method;

s2: screening out characteristic parameters capable of representing bearing degradation information and screening out characteristic parameters with poor characterization capability;

s3: setting a bearing sample life label, wherein the label is set as a normalization value of the residual service life of the bearing corresponding to the current sample;

s4: dividing a bearing vibration data sample into a training set and a testing set, and carrying out standardized processing on the training set;

s5: and (3) conveying the training samples to a GRU network model, setting specific parameters of the model, training the prediction model by adopting a long-term iteration method, inputting the test set into the trained model, drawing a prediction curve of the residual life of the bearing, and analyzing the prediction curve.

6. The method for predicting the life of the main shaft bearing of the wind turbine based on GRU hyper-parameter optimization according to claim 5, wherein the step of specifically calculating the normalized value in S3 is as follows:

wherein, rul_maxIndicating a predicted threshold for residual life of the bearing, rul_realRepresenting the actual remaining service life of the bearing; rul, the remaining service life of the current sample after being cut; rul_tRepresenting a sample data lifetime normalization value.

7. The method for predicting the life of the main shaft bearing of the wind turbine generator based on GRU hyper-parameter optimization of claim 5, wherein the concrete calculation formula of the standardization processing in S4 is as follows:

wherein x represents a proposed characteristic parameter; mean (x) represents the average processing of the characteristic parameters; std (x) represents the standard deviation of the characteristic parameter.

Images