A method for predicting the remaining life of rolling bearings

Abstract

This invention discloses a method for predicting the remaining service life (RUL) of rolling bearings. First, data detected under different operating conditions of various mechanical equipment is divided into source domain data and target domain data, and then training and validation sets are constructed. The data undergoes batch standardization. A deep convolutional neural network is used to extract the time-domain and frequency-domain features of the original vibration signals, and these features are concatenated to form a surrogate feature set. Using deep time-series features, a fully connected layer is output, and the deep time-series features are compared with the RUL labels. Finally, a large amount of labeled source domain data is used to repeatedly learn the data features under different operating conditions, calibrating the correspondence between features and RUL values, constructing a mapping relationship between features and RUL, and then adding a Dropout layer to build the model and verify it on the source domain test set. This method can significantly reduce the manpower and material costs associated with data relabeling, better uncover the inherent degradation trend of bearings, and effectively improve the accuracy of remaining service life prediction.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of mechanical component life prediction, and particularly relates to a rolling bearing residual life prediction method. BACKGROUND

[0002] With the rapid development of modern industry, the high integration of mechanical equipment puts forward higher requirements for the safety of the system. Bearings, as one of the key basic parts of mechanical equipment, according to relevant statistical data, more than 30% of the failure of rotating machinery equipment is caused by bearing failure. Machine prediction and health management (PHM) plays a crucial role in maintaining the stable operation of mechanical equipment. The degradation of mechanical equipment bearings is related to many factors, such as operating mode, working load, material defects and operating temperature. In addition, the long degradation period and the randomness of bearing failure make the prediction of the residual service life of the bearing a challenging task. In view of the current problems of different bearing operating conditions, less effective data, unlabeled data and low prediction accuracy, a rolling bearing residual life prediction method is proposed, which can more effectively predict the residual bearing life under different equipment and different working conditions. SUMMARY

[0003] Therefore, the application provides a rolling bearing residual life prediction method, which can solve the problem of low prediction accuracy caused by different bearing operating conditions, less effective data and unlabeled data in the prior art.

[0004] The purpose of the application can be achieved by the following technical solutions:

[0005] A rolling bearing residual life prediction method, comprising the steps of:

[0006] Step 1, obtaining data detected under different working conditions of different mechanical equipment, dividing the data into source domain data and target domain data, and then constructing a training set and a validation set according to the source domain data and the target domain data;

[0007] Step 2, batch standardizing the source domain data and the target domain data;

[0008] Step 3, using a deep convolutional neural network DCNN to extract features of time domain features and frequency domain features of the original vibration signal, and splicing the time domain and frequency domain features to form a replaceable feature set;

[0009] Step 4, using deep time sequence features, outputting a full connection layer, and comparing the deep time sequence features with RUL labels;

[0010] Step 5, repeat learning data characteristics of different working conditions using a large amount of labeled data in the source domain, calibrate the characteristics and the RUL corresponding value, construct the mapping relationship between the characteristics and the RUL, then add a Dropout layer, thereby constructing the source domain model, and testing in the source domain test set;

[0011] Step 6, using the source domain model and the target domain unlabeled data set, analyzing the parameter characteristics of the source domain data and the target domain data, learning similar characteristics, and minimizing the distribution difference between the target domain and the source domain;

[0012] Step 7, based on the source domain model, performing migration training, updating the parameters of the model, and training a migration prediction model;

[0013] Step 8, using the migration prediction model to predict and verify the RUL of the rolling bearing under different device and different working condition signals.

[0014] Further, the source domain data and the target domain data in step 1 include: obtaining monitoring data under different working conditions of a mechanical device to construct a source domain, and dividing training set and verification set; monitoring data under different working conditions of the same mechanical device or different mechanical devices to construct a target domain, and dividing training set and verification set.

[0015] Further, the training set and the verification set include: the training set of the source domain data is labeled with RUL, and the test set is not labeled with RUL; the training set and the test set of the target domain data are not labeled with RUL.

[0016] Further, the batch normalization processing in step 2 includes: mapping the source domain data and the target domain data to 0-1, thereby ensuring the stability of the data.

[0017] Further, the vibration original signal extracted by the deep convolutional neural network DCNN in step 3 includes: 11 time domain indexes of vibration signals such as mean, skewness, and margin factor, 12 frequency domain indexes of vibration signals such as amplitude variance, amplitude average, and spectral bandwidth, a total of 23 characteristic indexes to construct a selectable feature set; the first 15 features with high scores are selected by a screening equation as the input of the RUL prediction model, as the bearing remaining useful life prediction; the screened time domain indexes and frequency domain indexes are standardized.

[0018] Further, in step 3, the deep convolutional neural network DCNN can gradually learn high-level semantic features from low-level edge and texture features through multi-layer convolution and pooling operations, and the convolution operation process expression is as follows:

[0019]

[0020] In the formula: is the weight of the i-th convolution kernel in the l-th layer; is the bias of the i-th convolution kernel in the l-th layer; X l (j) is the j-th local region in the l-th layer; is the input of the j-th neuron in the i-th convolution kernel operation result of the l+1-th layer;

[0021] After the convolution operation, a ReLU function is used as the activation function of the model to speed up the convergence of the model and prevent overfitting. The expression of the ReLU function is as follows:

[0022] f(x) = max{0, x}

[0023] In the deep convolutional neural network DCNN, a pooling layer is added after the convolution layer to reduce the computational burden. There are two common pooling operations, namely average pooling and maximum pooling. The expressions of the two pooling operations are as follows:

[0024] v l(i,j) = max H(j-1)-1＜n＜jH {x l(i,n)}

[0025]

[0026] wherein: l(i, n) is the activation value of the n-th neuron in the i-th channel in the l-th layer; H is the size of the sliding window; v l(i,j) is the result of the j-th neuron in the current l-th layer after pooling;

[0027] The deep convolutional neural network DCNN is used to extract features from the original signal, and the maximum pooling layer is used to improve the speed of feature extraction. Then the time domain and frequency domain features are spliced to form a replaceable feature set.

[0028] Further, in step 7, a transfer learning bearing life prediction model is established, transfer learning is performed on the DCNN model, and a large number of source domain data sets are used to pre-train the transfer prediction model to ensure that the model learns as many different working condition data features as possible. At the same time, transfer learning is used to extract similar features of vibration signals under different working conditions and different equipment to reduce the difference between the source domain and the target domain. On the basis of the DCNN model, the similar features learned by transfer learning are combined to retrain a transfer model for residual life prediction in the target domain.

[0029] Further, in step 8, after the bearing residual life prediction is completed, the source domain test set and the target domain test set are verified. For this purpose, the method uses two indicators, mean absolute error MAE and root mean square error RMSE, to evaluate the prediction effect, and the formulas are as follows:

[0030]

[0031]

[0032] In the formula: m is the number of prediction points; i is the serial number of the prediction point y i Actual value; Predicted value;

[0033] The MAE value ranges from 0 to +infinity, and the smaller the value is, the higher the accuracy is.

[0034] Compared with the prior art, the rolling bearing residual life prediction method provided by the application has the beneficial effects that:

[0035] Firstly, according to the data detected under different working conditions of different mechanical equipment, the data is divided into source domain data and target domain data, and then a training set and a verification set are constructed according to the source domain data and the target domain data; secondly, the input data is batch standardized; thirdly, a deep convolutional neural network (DCNN) is used to extract the time domain features and the frequency domain features of the original vibration signal, and the time domain features and the frequency domain features are spliced to form a replaceable feature set; fourthly, deep time sequence features are used to output a full connection layer, and the deep time sequence features are compared with RUL labels; finally, a large amount of source domain labeled data is used to repeatedly learn the data features of different working conditions, calibrate the corresponding values of the features and RUL, and construct the mapping relationship between the features and RUL, and then a Dropout layer is added, so as to construct a model and test it in a source domain test set; the rolling bearing residual life prediction model constructed by the DCNN network and the transfer learning can effectively solve the problems of different bearing working conditions, less effective data, no labeled data, and low prediction accuracy.

[0036] The source domain model and the target domain unlabeled data set are used to analyze the parameters and features of the source domain data and the target data, learn similar features, and minimize the distribution difference between the target domain and the source domain; then the source domain model is used for transfer training, the parameters of the model are updated, and a transfer model is trained; finally, the transfer prediction model is used to predict and verify the RUL of the rolling bearing of different equipment under different working conditions. Compared with the prior art, the method can greatly reduce the consumption of manpower and material resources caused by data re-labeling, better tap the inherent degradation trend of the bearing, and effectively improve the residual life prediction accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0037] The above and other objects, features and advantages of the exemplary embodiments of the present application will be more apparent from the following detailed description taken in conjunction with the accompanying drawings, in which several embodiments of the present application are shown by way of example, and wherein like or corresponding elements refer to like or corresponding parts throughout the several views, and in which:

[0038] Figure 1 The prediction flowchart of the present application.

[0039] Figure 2 Figure 1 is a schematic diagram of a DNCC network structure according to the present application. DETAILED DESCRIPTION

[0040] Exemplary embodiments of the present disclosure will be described in greater detail below with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it is to be understood that the present disclosure can be variously implemented without being limited to the embodiments set forth herein. Rather, the embodiments are provided so that the present disclosure can be thoroughly understood and complete conveyance of the scope of the present disclosure to those skilled in the art is achieved. Technical means used in the examples are conventional means well known to those skilled in the art unless otherwise specified.

[0041] It is to be noted that technical terms or scientific terms used in the present application should be understood as their general meanings understood by those skilled in the art unless otherwise specified. The terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or apparatus including a series of elements includes not only those elements but also other elements not explicitly listed or inherent to such a process, method, article or apparatus. Without further limitation, the elements defined by the phrase "comprise" do not exclude the presence of additional identical elements in the process, method, article or apparatus including the elements.

[0042] As shown in FIG. 1, it is a flowchart of a rolling bearing residual life prediction method provided by the present application, and the method includes the following steps: Figure 1

[0043] Step 1, according to the data detected under different working conditions of different mechanical equipment, the data is divided into source domain data and target domain data, and then a training set and a validation set are constructed according to the source domain data and the target domain data;

[0044] Step 2, batch standardization processing is performed on the input data;

[0045] Step 3, a deep convolutional neural network (DCNN) is used to extract time domain features and frequency domain features of the original vibration signal, and the time domain and frequency domain features are spliced to form a replaceable feature set;

[0046] Step 4, deep time sequence features are used to output a full connection layer, and the deep time sequence features are compared with RUL labels;

[0047] Step 5, a large amount of source domain labeled data is used to repeatedly learn data features of different working conditions, calibrate feature and RUL corresponding values, construct a mapping relationship between features and RUL, and then add a Dropout layer, so as to construct a model and test in a source domain test set; ​

[0048] Step 6, using the source domain model and the target domain unlabeled data set, analyzing the parameters and characteristics of the source domain data and the target data, learning similar characteristics, and minimizing the distribution difference between the target domain and the source domain;

[0049] Step 7, based on the source domain model, the migration training is carried out, the parameters of the model are updated, and the migration model is trained;

[0050] Step 8, the RUL of the rolling bearing is predicted by the migration prediction model for different device and different working condition signals.

[0051] The specific implementation of the above steps is described in detail as follows:

[0052] In step 1, the data is collected by using an acceleration sensor to collect vibration signals in horizontal and vertical directions respectively, the sampling frequency of the data is 25.6 kHz, the data is recorded once every 10 s, the collection time is 0.1 s, and the vibration data collected each time is summarized; In the experiment, cross-validation method is used to divide the data, according to different test data, it is divided into test data and training data, and then it is superimposed to the input model for training.

[0053] In step 2, the input data is standardized to map the data to 0-1 to ensure the stability of the data. This method uses Z-score standardization criterion to process the selected features, and the formula of Z-score standardization criterion is as follows:

[0054] Z=(X-μ) / σ

[0055] (In the formula: Z is the Z-score value, X is the original data, μ is the average value of the original data, and σ is the standard deviation of the original data)

[0056] In step 3, as shown in the following formula, the DNCC network structure diagram is shown in the following formula, and the DCNN network model can gradually learn high-level semantic features from low-level edge, texture and other features through multi-layer convolution and pooling operation, and the convolution operation process expression is as follows: Figure 2 Figure 2

[0057]

[0058] (In the formula: W is the weight of the i-th convolution kernel in the l-th layer; b is the bias of the i-th convolution kernel in the l-th layer; X is the input of the j-th local region in the l-th layer; Y is the output of the j-th neuron in the i-th convolution kernel operation result in the l+1-th layer) l (j) is the j-th local region in the l-th layer;

[0059] ​​​​​After the convolution operation, the ReLU function is usually used as the activation function of the model to speed up the convergence of the model and prevent overfitting. The ReLU function expression is as follows:

[0060] f(x) = max{0, x}

[0061] In the DCNN network structure, a pooling layer is usually added after the convolution layer to reduce the computational burden. There are two common pooling operations, namely average pooling and maximum pooling. The expressions of the two pooling operations are as follows:

[0062] v l(i,j) = max H(j-1)-1＜n＜jH {x l(i,n)}

[0063]

[0064] (In the formula: l(i, n) is the activation value of the nth neuron in the ith channel in the lth layer; H is the size of the sliding window; v l (i,j) is the result of pooling of j neurons in the current l layer)

[0065] The DCNN network is used to extract features from the original signal, and the maximum pooling layer is used to improve the feature extraction speed, and then the time domain and frequency domain features are spliced to form a replaceable feature set.

[0066] In step 7, the transfer learning bearing life prediction model is established, the transfer learning is performed on the DCNN model, and the transfer prediction model is pre-trained through a large number of source domain data sets to ensure that the model learns as many different working condition data features as possible. At the same time, the transfer learning extracts similar features of vibration signals under different working conditions and different equipment to reduce the difference between the source domain and the target domain. The transfer model is retrained based on the DCNN model combined with the similar features learned by transfer learning, which is used for residual life prediction in the target domain.

[0067] In step 8, after the bearing residual life prediction is completed, the source domain test set and the target domain test set are verified. For this, the method uses two indicators, mean absolute error (MAE) and root mean square error (RMSE), to evaluate the prediction effect, and the formulas are as follows:

[0068]

[0069] (In the formula: m is the number of prediction points; i is the sequence number of the prediction point y i actual value; predicted value)

[0070] The MAE value range is [0, +∞), the smaller the better, and the smaller the RMSE value, the higher the accuracy.

[0071] In the present application, first, according to the data detected under different working conditions of different mechanical equipment, the data is divided into source domain data and target domain data, and then a training set and a verification set are constructed according to the source domain data and the target domain data; second, the input data is batch standardized; third, the deep convolutional neural network (DCNN) is used to extract the time domain features and the frequency domain features of the original vibration signal, and the time domain features and the frequency domain features are spliced to form a replaceable feature set; fourth, the deep time sequence features are used to output a full connection layer, and the deep time sequence features are compared with RUL labels; finally, a large amount of source domain labeled data is used to repeatedly learn the data features of different working conditions, calibrate the corresponding values of the features and RUL, and construct the mapping relationship between the features and RUL, and then a Dropout layer is added, so as to construct a model and test it in the source domain. The prediction model of the rolling bearing residual life constructed by the DCNN network and the transfer learning can effectively solve the problems of different working conditions of bearings, less effective data, no labeled data, and low prediction accuracy.

[0072] The source domain model and the target domain unlabeled data set are used to analyze the parameters and features of the source domain data and the target data, learn similar features, and minimize the distribution difference between the target domain and the source domain; then the source domain model is used for transfer training, the parameters of the model are updated, and a transfer model is trained; finally, the transfer prediction model is used to predict and verify the RUL of the rolling bearing under different equipment and different working condition signals. Compared with the existing method, the method can greatly reduce the consumption of manpower and material resources caused by data re-labeling, better tap the inherent degradation trend of the bearing, and effectively improve the prediction accuracy of the residual service life.

[0073] Test process and verification: In order to verify the effectiveness of the rolling bearing residual life prediction method, the following test process and test scheme are designed. The test process is divided into three parts: data acquisition, model training and verification, and result analysis.

[0074] 1. The data acquisition aspect includes experimental equipment conditions and data recording: (1) Experimental equipment conditions: build a corresponding bearing degradation test platform, select experimental bearing models (such as 6007, 6008, 6009, etc.), arrange high sampling frequency vibration acceleration in the radial direction (horizontal and vertical directions) of the bearing, and set multiple load and speed combinations to simulate the bearing degradation process under different working conditions; (2) Data recording: record data once at a set time, collect the required number of vibration signals each time, obtain normal state data at the initial stage of operation and full life cycle data before failure, and divide the data into source domain data (labeled degradation signal) and target domain data (unlabeled signal).

[0075] 2、Model training and verification: (1) Feature extraction includes time domain features, frequency domain features and feature selection; (2) Model training: using deep convolutional neural network (DCNN) to learn the characteristics of the vibration signal, set the network structure: contains convolution layer, each layer uses ReLU activation function, convolution kernel size, step is, the pooling layer uses maximum pooling, the parameter setting of the Dropout layer is used to prevent overfitting; (3) Transfer learning: first train the initial model based on the source domain data, use the unlabeled target domain data for transfer learning, extract the similar features of the source domain and the target domain, reduce the distribution difference, update the parameters of the transfer model, get the transfer prediction model; (4) Verification test: respectively on the source domain test set and the target domain test set for the remaining useful life (RUL) prediction, using the mean absolute error (MAE) and root mean square error (RMSE) for evaluation.

[0076] 3、Test result analysis: (1) Compared with traditional machine learning methods, the advantages of this method are reflected in that when in different working conditions, the MAE and RMSE values of the transfer learning model are better than those of the traditional method, especially in the test of unlabeled target domain data; The high-order semantic features extracted by the deep convolutional neural network can effectively reflect the degradation trend of the bearing and have stronger robustness; At the same time, the transfer learning method has stronger adaptability between different working conditions and devices, the feature difference between the source domain and the target domain can be effectively reduced, and finally the prediction accuracy can be improved.

[0077] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any skilled person in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A rolling bearing residual life prediction method characterized by, Comprise the steps of: Step 1, obtain the data detected under different working conditions of different mechanical equipment, divide the data into source domain data and target domain data, and then construct a training set and a validation set according to the source domain data and the target domain data; Step 2, batch standardization processing is performed on the source domain data and the target domain data; Step 3, using deep convolutional neural network DCNN to extract the time domain features and frequency domain features of the original vibration signal, and splicing the time domain and frequency domain features to form a replaceable feature set; Step 4, using deep time sequence features, outputting a fully connected layer, comparing the deep time sequence features with the RUL label; Step 5, using a large amount of labeled data in the source domain to repeatedly learn the data features of different working conditions, calibrating the feature and the RUL corresponding value, constructing the mapping relationship between the feature and the RUL, and then adding a Dropout layer to construct a source domain model, and testing in the source domain test set; Step 6, using the source domain model and the target domain unlabeled data set, analyzing the parameter features of the source domain data and the target domain data, learning similar features, and minimizing the distribution difference between the target domain and the source domain; Step 7, based on the source domain model, the migration training is carried out, and the parameters of the model are updated, and a migration prediction model is trained; In step 7, a migration bearing life prediction model is established, migration learning is carried out on the DCNN model, and a large amount of source domain data set is used to pretrain the migration prediction model, so that the model can learn as many different working condition data features as possible, and the similar features of the vibration signals under different working conditions and different equipment are extracted by using migration learning to reduce the difference between the source domain and the target domain; The migration model is retrained based on the DCNN model combined with the similar features learned by migration, and is used for residual life prediction in the target domain; Step 8, the migration prediction model is used to predict and verify the RUL of the rolling bearing of different equipment and different working condition signals.

2. The rolling bearing residual life prediction method according to claim 1, characterized in that, The source domain data and the target domain data of step 1 comprise: obtaining the monitoring data under different working conditions of a mechanical equipment to construct a source domain, and dividing the training set and the validation set; the monitoring data under different working conditions of the same mechanical equipment or different mechanical equipment to construct a target domain, and dividing the training set and the validation set.

3. The rolling bearing residual life prediction method according to claim 2, characterized in that, The training set and the validation set comprise: the training set of the source domain data is with RUL label, and the test set is without RUL label; the training set and the test set of the target domain data are without RUL label.

4. The rolling bearing residual life prediction method according to claim 1, characterized in that, The batch standardization processing of step 2 comprises: mapping the source domain data and the target domain data to 0~1, so as to ensure the stability of the data.

5. The rolling bearing residual life prediction method according to claim 1, characterized in that, The vibration original signal extracted by the deep convolutional neural network DCNN of step 3 comprises: 11 time domain indexes of vibration signals including mean, skewness and kurtosis factor, 12 frequency domain indexes of vibration signals including amplitude variance, amplitude average and frequency bandwidth, a total of 23 characteristic indexes to construct a selectable feature set; the first 15 features with high scores are selected by a screening equation as the input of the RUL prediction model, which is used for bearing residual life prediction; the screened time domain indexes and frequency domain indexes are standardized.

6. The rolling bearing residual life prediction method according to claim 1, characterized in that, In step 3, the deep convolutional neural network DCNN is able to gradually learn high-level semantic features from low-level edge and texture features through multi-layer convolution and pooling operations. The convolution operation process expression is shown as follows: ； In the formula: For the first The first in the layer The weights of each convolutional kernel; For the first The first in the layer The bias of each convolutional kernel; for The j-th local region in the layer; For the first The first layer The input of the j-th neuron in the result of each convolution kernel operation; After the convolution operation, the ReLU function is used as the activation function of the model to speed up the convergence of the model and prevent overfitting. The ReLU function expression is shown as follows: ； In the deep convolutional neural network DCNN, a pooling layer is added after the convolution layer to reduce the computational burden. The common pooling operations are average pooling and maximum pooling. The expressions of the two pooling operations are as follows: ； ； In the formula: is the first layer in activation value of the nth neuron in the channel H is the size of the sliding window; is the current layer in the result after pooling of the neurons The deep convolutional neural network DCNN is used to extract features from the original signal, and the maximum pooling layer is used to improve the feature extraction speed. Then the time domain and frequency domain features are spliced to form a replaceable feature set.

7. The rolling bearing residual life prediction method according to claim 1, characterized in that, In step 8, after the bearing residual life prediction is completed, the source domain test set and the target domain test set are verified. For this, the method uses the mean absolute error MAE and the root mean square error RMSE to evaluate the prediction effect. The formulas are as follows: ； ； In the formula, m is the number of prediction points; The serial number of the prediction point is Actual value; Prediction value; The MAE value range is [0, +∞), the smaller the better, and the smaller the RMSE value, the higher the accuracy.

Images