A life prediction method for rolling bearings

Abstract

The embodiment of the present invention discloses a life prediction method for rolling bearings, which relates to the technical field of rolling bearings and can improve the accuracy of life prediction of rolling bearings. The present invention comprises: extracting the signal characteristics of the vibration signal from the collected vibration signal of the rolling bearing; obtaining an initial characteristic matrix by limited dimensionality reduction processing according to the signal characteristics, and then performing nonlinear fitting processing using the initial characteristic matrix to obtain a characteristic matrix; forming a sample segment by the obtained characteristic matrix, and uploading it to a training database and a test database; training the established time convolution network model using the training database; inputting the test data in the test database into the trained convolution network model, and obtaining a fault diagnosis result. The present invention is applicable to the design of rolling bearings.

Description

Technical Field

[0001] The present invention relates to the technical field of rolling bearings, and in particular to a life prediction method for rolling bearings. Background Art

[0002] Rolling bearings are known as the "joints of industry" and are widely used in aerospace, wind power generation and other fields. Many failures of rotating machinery are caused by rolling bearing failures. Once a failure occurs, it can cause property losses or even casualties. Therefore, it is necessary to perform status detection on rolling bearings, obtain the status of mechanical equipment, estimate the life of rolling bearings, and thus provide early warning of possible failures.

[0003] Existing intelligent prediction methods usually extract features from the information in the historical data of a variable, and use these features to predict the change trend of the variable in the future. However, when the historical data is relatively large, the overall temporal correlation of the features extracted by traditional data prediction methods will be weakened, and they will not be able to reflect the rules implied by the entire historical data, which will lead to poor prediction accuracy. Some methods are also prone to overfitting problems. Summary of the invention

[0004] An embodiment of the present invention provides a life prediction method for a rolling bearing, which can improve the accuracy of the life prediction of the rolling bearing.

[0005] To achieve the above object, the embodiments of the present invention adopt the following technical solutions:

[0006] Extracting a signal feature of the vibration signal from the collected vibration signal of the rolling bearing;

[0007] An initial feature matrix is ​​obtained by performing a dimensionality reduction process based on the signal feature limitations, and then a nonlinear fitting process is performed using the initial feature matrix to obtain a feature matrix;

[0008] The obtained feature matrix is ​​used to form a sample segment and uploaded to the training database and the test database;

[0009] Using the training database to train the established temporal convolutional network model;

[0010] The test data in the test database is input into the trained convolutional network model to obtain a fault diagnosis result.

[0011] The life prediction method for rolling bearings provided in the embodiment of the present invention predicts the life of rolling bearings based on a time convolutional network. Compared with traditional neural networks, time convolutional networks can process time series data in parallel and flexibly change the receptive field. Under the same conditions, more information can be extracted than traditional neural networks, and the ability to process longer and more complex equipment historical data is improved, which can improve the accuracy of life prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0013] Figure 1 A schematic diagram of an implementation flow of a specific example in a possible implementation method provided by an embodiment of the present invention;

[0014] Figure 2 A schematic diagram of obtaining a final feature matrix provided by an embodiment of the present invention;

[0015] Figure 3 A schematic diagram of the temporal convolutional network training and testing process provided by an embodiment of the present invention;

[0016] Figure 4 A schematic diagram of a temporal convolutional network training process provided by an embodiment of the present invention;

[0017] Figure 5 A schematic diagram of a residual link portion of a temporal convolutional network provided by an embodiment of the present invention;

[0018] Figure 6 A schematic diagram of a method flow provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0019] In order to enable those skilled in the art to better understand the technical solution of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. The embodiments of the present invention will be described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements with the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be interpreted as limiting the present invention. It can be understood by those skilled in the art that, unless specifically stated, the singular forms "one", "one", "said" and "the" used herein may also include plural forms. It should be further understood that the term "including" used in the specification of the present invention refers to the presence of the features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or their groups. It should be understood that when we say that an element is "connected" or "coupled" to another element, it can be directly connected or coupled to other elements, or there may also be intermediate elements. In addition, the "connection" or "coupling" used here may include wireless connection or coupling. The term "and / or" used herein includes any unit and all combinations of one or more associated listed items. It will be understood by those skilled in the art that, unless otherwise defined, all terms (including technical terms and scientific terms) used herein have the same meaning as generally understood by those of ordinary skill in the art to which the present invention belongs. It should also be understood that terms such as those defined in general dictionaries should be understood to have meanings consistent with the meanings in the context of the prior art, and will not be interpreted with idealized or overly formal meanings unless defined as herein.

[0020] The embodiment of the present invention provides a life prediction method for a rolling bearing, which can be specifically understood as a life prediction method for a rolling bearing based on deep learning, comprising:

[0021] S1. Extracting signal features of the vibration signal from the collected vibration signal of the rolling bearing.

[0022] The step of collecting the vibration signal of the rolling bearing includes: collecting the vibration signal through a portable dynamic signal collector and an acceleration sensor installed on the rolling bearing. The vibration direction includes horizontal and vertical directions, and the sampling frequency is 25.6kHz.

[0023] S2. Performing dimension reduction processing based on the signal feature limitations to obtain an initial feature matrix, and then performing nonlinear fitting processing using the initial feature matrix to obtain a feature matrix.

[0024] S3. The obtained feature matrix is ​​used to construct a sample segment and uploaded to the training database and the test database.

[0025] S4. Using the training database to train the established temporal convolutional network model.

[0026] S5. Input the test data in the test database into the trained convolutional network model to obtain a fault diagnosis result.

[0027] Specifically, the execution process of the scheme of this embodiment can also be divided according to the process of data interaction, which can be roughly divided into four stages: data acquisition stage: collecting the vibration signal of the rolling bearing throughout its life cycle, and dividing the obtained full life cycle vibration signal into training data and test data; data preprocessing stage: extracting multiple time domain, frequency domain and time-frequency domain signal features respectively, and adding linearization processing and dimensionality reduction features to form multiple sample segments, and the above sample segments constitute a training database and a test database; model training stage: constructing a time convolutional network model and training the network on the training set; testing stage: inputting the test data into the trained time convolutional network model to obtain fault diagnosis results.

[0028] In this embodiment, the time convolution network can process time series data in parallel and flexibly change the receptive field compared to the traditional neural network. Under the same conditions, more information can be extracted than the traditional neural network, and the ability to process longer and more complex equipment historical data is improved, and the life of the rolling bearing can be predicted. In addition, the method of introducing the features obtained by nonlinear fitting in the feature extraction stage improves the accuracy of life prediction compared with the recurrent neural network (RNN), long short-term memory network (LSTM) and gated recurrent neural network (GRU).

[0029] In this embodiment, the initial feature matrix is ​​obtained by the limited dimensionality reduction processing based on the signal features, including: using the signal features to form an m×n feature matrix. Then, the feature matrix to be reduced and the target dimension k are input, and the matrix A composed of the k-dimensional sample feature vectors after the dimensionality reduction is output fpca And the principal component vector W.

[0030] The nonlinear fitting process using the initial feature matrix includes: establishing a multi-order polynomial function model f(x) and a model loss function Loss that fit the trend of real data changes. The initial feature matrix is ​​optimized by f(x) and Loss, and the optimized feature matrix A' is obtained. fpca The signal features include: time domain signal features, frequency domain signal features and time-frequency domain signal features, m is the number of samples, and n is the feature dimension. The gradient descent method is used to optimize the model loss function Loss. After each update, the corresponding loss function value decreases until the feature matrix A fpcaThe sum of the numerical value in and the numerical value spacing obtained by fitting f(x) is the smallest. Specifically, the processes that need to be executed in the rolling bearing data preprocessing stage include: vibration signal feature extraction: extract multiple time domain signals with strong life correlation from the collected rolling bearing horizontal or vertical full life cycle vibration signal time series, multiple frequency domain signals obtained by wavelet decomposition and denoising of the original vibration signal, and multiple time-frequency domain signals obtained by wavelet packet decomposition of the original vibration signal as preliminary features for life prediction. Feature dimensionality reduction and linearization processing: perform fast PCA dimensionality reduction on the features obtained above to obtain the reduced feature time series, and perform feature linear fitting on each dimension of this series to obtain the final feature.

[0031] In this embodiment, after uploading the collected vibration signal data to the training database and the test database, the data needs to be preprocessed: first, the time domain, frequency domain and time-frequency domain features need to be extracted: the time domain features, frequency domain features and time-frequency domain features of the original vibration signal are extracted. The features extracted above are formed into an m×n feature matrix with the number of rows being the number of samples m and the number of columns being the feature dimension n. The obtained feature matrix is ​​then subjected to dimensionality reduction processing: the feature matrix A to be reduced and the target dimension k are input, and the matrix A consisting of the k-dimensional sample feature vectors after dimensionality reduction is output. fpca And the principal component vector W. Then perform nonlinear feature fitting: the feature matrix A obtained by learning fpca , get a multi-time polynomial function f(x) that can fit the trend of real data changes, so that the feature matrix A fpca The sum of the numerical values ​​in and the numerical distance obtained by fitting f(x) is the smallest, and the newly obtained characteristic matrix is ​​recorded as A' fpca Specifically, the learned f(x) and loss function Loss need to be set as follows:

[0032] Then optimize the loss function. To achieve the optimal value, derive each coefficient so that its partial derivative is 0. Then use the gradient descent method to optimize the loss function. For example: ① Initialize k coefficient values ​​and start iteration; ② Calculate each coefficient n in each iteration k' Corresponding gradient value; ③ Use gradient value and learning rate to update each coefficient n k' ; ④ Ensure that the corresponding loss function value decreases after the parameter update. In practical applications, the task goal is to calculate n k' The corresponding gradient value and update n k' . A' fpca And the obtained feature matrix A fpca Combine to form the final feature matrix F fpca ; Finally, the feature matrix F fpcaDivide into training set and test set and normalize them, and use the training set feature data as the input of the temporal convolutional network.

[0033] In this embodiment, the obtained feature matrix is ​​used to form a sample segment and uploaded to the training database and the test database, including: after normalizing the final features of the training data and the final features of the test data within the range of [0,1], the final features of the training data and the final features of the test data are used to form a training set and a test set. Then the training set and the test set are uploaded to the training database and the test database respectively. The training database is used to train the established time convolutional network model, including: constructing a time convolutional network model and initializing network parameters, wherein the network parameters include: step size τ, input dimension k and output dimension k', initial learning rate η, and number of training times e. The average relative error is used as the model loss function, and the minimum MRE is used as the optimization goal during the training process. Among them, the ideal life degradation curve is used as the data label, and the life range is recorded as [0,f] and normalized to [0,1] as the remaining life. During the training process, the network weights are updated by the Adam algorithm, and the training stops when the loss function value of the model loss function stabilizes and no longer decreases.

[0034] It can also be understood that the model training stage in this embodiment includes: constructing a time convolutional network model; training the time convolutional network model to obtain a trained time convolutional network model, which can realize the prediction of the rolling bearing life under the target working condition. Among them, the time convolutional network model construction and training:

[0035] First, initialize the network parameters: step size τ, input dimension k and output dimension k', initial learning rate η, number of training times e, etc. Secondly, it is necessary to determine the model loss function: select the mean relative error (MRE) as the model loss function, and the training process takes the minimum MRE as the optimization goal. After that, set the data label: the data label uses the ideal life degradation curve, that is, the time from the initial operation to the final scrapping of the bearing is taken as the life, the life range is recorded as [0,f], and then it is normalized to [0,1] as the remaining life. Finally, update the network weights: use the Adam algorithm to update the network weights. When the loss function of the model stabilizes and almost no longer decreases, the training stops, and the trained network can be used for life prediction.

[0036] In this embodiment, the test data in the test database is input into the trained convolutional network model to obtain the fault diagnosis result, including: inputting the test set into the trained convolutional network model, and outputting a one-dimensional array for characterizing the remaining life percentage, and denormalizing the one-dimensional array to obtain: Among them, L trueis the predicted remaining life, L is the sample length of the entire life cycle, T1 and T2 are the first and last data sample numbers of the test set, l is the length of the test set data, and y is the remaining life percentage output by the model.

[0037] For example, the above embodiment can be divided into four stages in practical application, wherein the first stage is data collection. The second stage is data preprocessing. The third stage is model training. The fourth stage is testing. It includes the following processes:

[0038] like Figure 1 As shown, the first stage includes: 1 collecting the original vibration signal of the rolling bearing; the second stage includes: 2 extracting the time domain, frequency domain and time-frequency domain features of the signal; 3 performing dimensionality reduction processing on the above extracted features and forming an initial feature matrix; as shown Figure 2 As shown, 4 performs nonlinear fitting on the features in the feature matrix obtained in 3, and forms the final feature matrix for life prediction with the original feature matrix; the third stage includes: Figure 3 As shown in Figure 5, a temporal convolutional network model is constructed and the network is trained. The network structure is as follows: Figure 4 and Figure 5 As shown; the fourth stage includes: 6 inputting the test sample into the trained network model to obtain the percentage of remaining life; 7 inputting the remaining life percentage into the life calculation formula to obtain the predicted remaining life.

[0039] Specifically, the first step of collecting the original vibration signal of the rolling bearing includes collecting the vibration signal obtained by the acceleration sensor and the portable dynamic signal collector during the operation of the rolling bearing. The vibration signal direction is divided into horizontal and vertical directions, and the sampling frequency is 25.6kHz.

[0040] Specifically, the 2 stage of extracting time domain, frequency domain and time-frequency domain features of the vibration signal includes: extracting time domain features, frequency domain features and time-frequency domain features of the original vibration signal. The features extracted above are formed into an m×n feature matrix with m rows and n columns, which is the number of samples and n feature dimensions.

[0041] Specifically, the 3rd step is to perform dimensionality reduction processing on the features extracted above and form an initial feature matrix stage, including: performing fast PCA dimensionality reduction processing on the feature matrix obtained in step 1: inputting the feature matrix A to be reduced and the target dimension k, and outputting a matrix A composed of k-dimensional sample feature vectors after dimensionality reduction fpca And the principal component vector W.

[0042] Specifically, the features in the feature matrix obtained in step 3 are subjected to nonlinear fitting to obtain the feature matrix A' fpcaThe feature matrix A obtained by learning 3 is formed with the original feature matrix to form the final feature matrix for life prediction. fpca , get a multi-time polynomial function f(x) that can fit the trend of real data changes, so that the feature matrix A fpca The sum of the intervals between the values ​​in and the values ​​obtained by fitting f(x) is the smallest. The specific steps are as follows:

[0043] (1) Set the learned f(x) and loss function Loss as follows:

[0044] f(x)=n0+n1x+n2x 2 +...+n k' x k′ ...+n k x k

[0045]

[0046] (2) Optimize the loss function. To achieve the optimal value, derive each coefficient so that its partial derivative is 0.

[0047] (3) Use the gradient descent method to optimize the loss function. The specific steps are as follows:

[0048] ① Initialize k coefficient values ​​and start iteration;

[0049] ② Calculate each coefficient n separately in each iteration k' The corresponding gradient value;

[0050] ③ Use gradient value and learning rate to update each coefficient n k' ;

[0051] ④ Ensure that the corresponding loss function value decreases after the parameter is updated.

[0052] In summary, the task goal is to calculate n k' The corresponding gradient value and update n k' .

[0053] (4) Substitute the obtained f(x) with the feature matrix A obtained in step 2 fpca Combine to form the final feature matrix F fpca ;

[0054] (5) The feature matrix F fpca Divide into training set and test set and normalize them, and use the training set feature data as the input of the temporal convolutional network.

[0055] Specifically, the step 5 of constructing a temporal convolutional network model and training the network includes:

[0056] (1) Initialize network parameters: step size τ, input dimension k and output dimension k', initial learning rate η, number of training times e, etc.

[0057] (2) Determine the model loss function:

[0058] The mean relative error (MRE) is selected as the model loss function, and the training process takes minimizing MRE as the optimization goal.

[0059] (3) Set data labels:

[0060] The data label uses the ideal life degradation curve, which takes the time from the initial operation to the final scrapping of the bearing as the life, records the life range as [0, f], and then normalizes it to [0, 1] as the remaining life.

[0061] (4) Network weight update:

[0062] The Adam algorithm is used to update the network weights. When the loss function of the model stabilizes and almost no longer decreases, the training stops and the trained network can be used for life prediction.

[0063] Specifically, the step 6 inputs the test sample into the trained network model to obtain the percentage stage of the remaining life, including:

[0064] By inputting the test set data into the trained network model, a one-dimensional array with a range of [0, 1] that can represent the remaining life percentage can be output.

[0065] Specifically, the remaining life percentage is input into the life calculation formula to obtain the predicted remaining life stage, including:

[0066] Substitute the one-dimensional array obtained in 6 into the remaining life calculation formula:

[0067] L true =L-(T2-(y*l))

[0068] l=T2-T1

[0069] Where L true is the predicted remaining life, L is the length of the full life cycle sample, T1 and T2 are the first and last data sample numbers of the test set, l is the length of the test set data, and y is the remaining life percentage output by the model. In this way, the remaining life can be obtained.

[0070] The time-based convolutional network proposed in this embodiment as a life prediction algorithm can be used to achieve life prediction of rolling bearings. This method increases network memory by introducing convolution in the network of the cyclic architecture, improves the ability to process longer and more complex equipment history data, and can predict the life of rolling bearings. In addition, the method of introducing features obtained by nonlinear fitting in the feature extraction stage improves the accuracy of life prediction.

[0071] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment. The above is only a specific implementation method of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or replacements that can be easily thought of by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A method for predicting the life of a rolling bearing, characterized in that: include: Extracting a signal feature of the vibration signal from the collected vibration signal of the rolling bearing; An initial feature matrix is ​​obtained by performing a dimensionality reduction process based on the signal feature limitations, and then a nonlinear fitting process is performed using the initial feature matrix to obtain a feature matrix; The obtained feature matrix is ​​used to form a sample segment and uploaded to the training database and the test database; Using the training database to train the established temporal convolutional network model; Inputting the test data in the test database into the trained convolutional network model and obtaining a fault diagnosis result; The method of obtaining an initial feature matrix by performing limited dimensionality reduction processing based on the signal features includes: using the signal features to form an m×n feature matrix, wherein the signal features include: time domain signal features, frequency domain signal features and time-frequency domain signal features, m is the number of samples, and n is the feature dimension; inputting the feature matrix to be reduced and the target dimension k, and outputting a matrix A composed of k-dimensional sample feature vectors after dimensionality reduction fpca and principal component vector W; The nonlinear fitting process using the initial feature matrix includes: establishing a multi-time polynomial function model f(x) and a model loss function Loss that fit the change trend of real data; optimizing the initial feature matrix through f(x) and Loss, and obtaining an optimized feature matrix A' fpca , where the gradient descent method is used to optimize the model loss function Loss. After each update, the corresponding loss function value decreases until the feature matrix A fpca The sum of the intervals between the values ​​in and the values ​​obtained by fitting f(x) is the smallest.

2. The method according to claim 1, characterized in that Collecting the vibration signal of the rolling bearing includes: Vibration signals are collected by a portable dynamic signal collector and an acceleration sensor installed on the rolling bearing, wherein the vibration direction includes horizontal and vertical directions, and the sampling frequency is 25.6 kHz.

3. The method according to claim 1, characterized in that The obtained feature matrix is ​​used to form a sample segment and uploaded to the training database and the test database, including: After normalizing the final features of the training data and the final features of the test data within the range of [0,1], the final features of the training data and the final features of the test data are used to form the training set and the test set respectively; Then, the training set and the test set are uploaded to the training database and the test database respectively.

4. The method according to claim 1, characterized in that The step of training the established temporal convolutional network model using the training database includes: Construct a temporal convolutional network model and initialize network parameters, wherein the network parameters include: step size τ, input dimension k and output dimension k', initial learning rate η, and number of training times e; The average relative error is used as the model loss function, and the minimum MRE is used as the optimization goal during the training process.

5. The method according to claim 4, characterized in that Also includes: The ideal life degradation curve is used as the data label, where the life range is recorded as [0,f] and normalized to [0,1] as the remaining life; During the training process, the network weights are updated through the Adam algorithm. When the loss function value of the model loss function stabilizes and no longer decreases, the training stops.

6. The method according to claim 1, characterized in that The step of inputting the test data in the test database into the trained convolutional network model and obtaining the fault diagnosis result comprises: The test set is input into the trained convolutional network model, and a one-dimensional array representing the remaining life percentage is output, and the one-dimensional array is denormalized to obtain: Among them, L true is the predicted remaining life, L is the sample length of the entire life cycle, T1 and T2 are the first and last data sample numbers of the test set, l is the length of the test set data, and y is the remaining life percentage output by the model.

Images