A deep learning-based rolling bearing RUL prediction method

Abstract

The application discloses a rolling bearing RUL prediction method based on deep learning, obtains kurtosis and peak factor from a bearing vibration signal, is used for health monitoring of a rolling bearing, judges a first prediction time of the bearing; uses wavelet transform to perform time-frequency analysis on an original vibration signal of the rolling bearing, obtains a wavelet power spectrum diagram; uses the obtained wavelet power spectrum diagram data set to represent a degradation state of the rolling bearing. A convolutional neural network model based on a Transform structure is trained and used to complete prediction of the remaining useful life of the rolling bearing. The application gives play to the advantage of the Transform in extracting long-distance information, solves the problem of actual application of a model in prediction of the remaining useful life of the rolling bearing, and makes the prediction of the remaining useful life of the rolling bearing more accurate.

Description

Technical Field

[0001] This invention belongs to the field of moving equipment and relates to rolling bearing remaining service life prediction technology, specifically to a rolling bearing RUL prediction method based on deep learning. Background Technology

[0002] Rolling bearings are critical components commonly used in rotating machinery. Failure of a rolling bearing accelerates the damage to other adjacent components and the machine itself, becoming a major cause of rotating equipment failure. Predicting the remaining useful life (RUL) of rolling bearings allows for accurate understanding of the equipment's operating status, thereby determining the degree of failure and remaining operating time, maximizing the bearing's working potential, and reducing maintenance losses. Vibration analysis, which uses vibration signals to obtain bearing failure status information, is currently the most effective and widely used method, as vibration signals contain more information about bearing defects. Currently, life prediction for rotating machinery is mainly divided into physical models and data-driven models. Building models based on physical failure requires estimating model parameters using measurement data, thus requiring extensive experience and expertise. Furthermore, with the increasing complexity of mechanical equipment, the generalization of these models becomes difficult. Data-driven methods build models based on historical condition monitoring data. In the initial stages, performance degradation indicators were constructed by manually extracting bearing failure characteristics.

[0003] The rapid development of IoT technology has led to an explosive growth in data, bringing new opportunities and challenges to the field of machine health monitoring. Furthermore, with the advancement of computer technology and deep learning, deep learning algorithms are increasingly being applied to the prediction and health management of mechanical equipment. Their powerful network structures, capable of capturing the latent characteristics of massive amounts of data, significantly improve the predictive performance of models. The method of adaptively extracting features using deep learning models in conjunction with other methods or theories is currently more widely used, but it also has some limitations. The short-time Fourier transform cannot change the window size during the transformation process, resulting in poor performance for time-frequency analysis of non-stationary signals with large variations. Many existing studies linearly characterize the degradation trend of bearings and normalize it to the range of (0, 1), assuming that degradation occurs from the start of machine operation, which is not the case in reality. Convolution operations excel at capturing local information within a small sliding window, but are weaker in capturing global features. These problems limit the accuracy of bearing remaining service life prediction to some extent; therefore, a suitable method is needed for high-precision prediction of bearing remaining service life. Summary of the Invention

[0004] To address the problems of existing technologies, this invention provides a deep learning-based method for predicting the rolling bearing RUL (Round Upward Length). It employs continuous wavelet transform to obtain the time-frequency information of the original acquired data, and enhances the learning ability of the network through convolution operations and the self-attention mechanism of Transformer, thereby achieving high-precision prediction of the rolling bearing RUL.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0006] On the one hand, this invention provides a deep learning-based method for predicting the RUL (Round Usage Limit) of rolling bearings, comprising the following steps:

[0007] Vibration monitoring experiments were conducted on the rolling bearing until the bearing failed; vibration signal data of the rolling bearing throughout its entire operating cycle and the final failure time were obtained.

[0008] Calculate the kurtosis and peak value of the bearing vibration signal, and determine the first prediction time of the rolling bearing based on the kurtosis and peak value.

[0009] Wavelet transform is performed on the acquired vibration signal data to obtain the wavelet power spectrum.

[0010] The obtained wavelet power spectrum is preprocessed to obtain samples. The remaining rolling bearing life at the time of sampling of each sample is calculated based on the final failure time and the first prediction time as a label to construct training samples.

[0011] Construct a CNN-Transformer joint model, calculate the remaining life of the rolling bearing at the time of each sample sampling as the sample label, and construct training samples;

[0012] The CNN-Transformer joint model is trained using training samples;

[0013] The RUL prediction of rolling bearings is performed using a trained CNN-Transformer joint model.

[0014] Preferably, the method for determining the first prediction time for each rolling bearing based on kurtosis and peak factor is as follows:

[0015] Calculate the mean μ and standard deviation σ of the kurtosis of the bearing during the early operation phase;

[0016] Based on the 3σ principle of normal distribution, the time when the bearing begins to degrade is determined as the first pre-selected time;

[0017] Calculate the mean μ and standard deviation σ of the peak factor of the bearing in the early operation stage. According to the 3σ principle of normal distribution, the time when the bearing begins to degrade is the second pre-selected time.

[0018] The minimum value between the first and second pre-selected times is taken as the first predicted time when the bearing begins to degrade.

[0019] Preferably, in the vibration monitoring experiment, vibration signals in both the horizontal and vertical directions are monitored. When calculating the first prediction time, the first prediction time under both vibration signals is calculated separately, and the earliest first prediction time is selected for calculating the remaining life of the rolling bearing.

[0020] Preferably, the remaining life of the rolling bearing is calculated as follows:

[0021] The remaining lifetime before the first prediction time is set to 100%, and the remaining lifetime at the time of final bearing failure is set to 0%. The formula for calculating the remaining lifetime between the first prediction time and the final bearing failure is as follows:

[0022]

[0023] In the above formula, y i Let t be the remaining life of the bearing when the i-th sample is collected, t be the final failure time of the bearing, and t1 be the first predicted time of the bearing.

[0024] The preferred wavelet transform method is as follows:

[0025] The wavelet coefficient matrix is ​​obtained by transforming the vibration time series signal using wavelet transform;

[0026] The wavelet energy obtained by squaring the wavelet coefficients is superimposed with the wavelet energy corresponding to each scale factor to obtain the wavelet energy spectrum curve that varies with the scale factor. The wavelet power image is obtained by converting the scale into frequency.

[0027] The preferred method for preprocessing the wavelet power spectrum is as follows:

[0028] After removing the coordinate axes and edges from the wavelet power spectrum, a dimensionality reduction process is performed to obtain an M×M wavelet power spectrum.

[0029] Preferably, the CNN-Transformer joint model includes a CNN convolutional neural network, an encoder, and a decoder connected in sequence. The encoder and decoder constitute the Transformer model architecture. The CNN convolutional neural network includes a convolutional layer, an activation layer, a max pooling layer, a batch normalization layer, a mean pooling layer, and a flattening layer in sequence. The output of the flattening layer is used as the input of the encoder.

[0030] The encoder includes a self-attention module and a fully connected feedforward network; the self-attention module performs linear mapping and normalization on the input variables; the fully connected feedforward network includes a linear mapping layer, a normalization layer, a fully connected (FC) layer, an activation layer, and a fully connected (FC) layer.

[0031] The decoder structure is the same as the encoder structure, and it outputs the prediction result in the last layer.

[0032] Preferably, during the preprocessing of the wavelet power spectrum, M is taken as an even power of 2. If the value is too small, less bearing degradation information is included; if the value is too large, a lot of computer resources are consumed and the improvement in prediction effect may not be significant. Therefore, M is generally taken between 16 and 256.

[0033] Preferably, a set of samples can be obtained by conducting vibration monitoring experiments on one rolling bearing. Multiple sets of samples can be obtained by conducting vibration monitoring experiments on multiple bearings, which can improve training accuracy. However, too many experimental bearings will result in excessively long experimental and training times and high costs. Therefore, the optimal number of experimental bearings is 3-15.

[0034] On the other hand, the present invention provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the rolling bearing RUL prediction method.

[0035] On the other hand, the present invention provides a non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that the computer program, when executed by a processor, implements the rolling bearing RUL prediction method as described above.

[0036] On the other hand, the present invention provides a computer program product, including a computer program, characterized in that the computer program implements the rolling bearing RUL prediction method when executed by a processor.

[0037] Compared with the prior art, the beneficial effects of the present invention are:

[0038] This invention provides an online prediction method for the remaining service life (RUL) of rolling bearings based on wavelet power spectrum and Transformer structure. This method enables cloud deployment of the model and online prediction of bearing RUL, providing support for intelligent decision-making. The proposed method employs a first prediction time (FPT) method based on time-domain features, which can accurately determine the bearing failure initiation time while comprehensively considering wear and impact failures, thus allowing the label value to better characterize the linear degradation trend of the bearing. This invention integrates the Transformer structure, which excels at extracting global information, with convolutional neural networks, which are effective at extracting local information, resulting in more accurate RUL prediction. The proposed method uses a novel neural network structure design, leading to higher model training efficiency and more accurate prediction of degradation curves.

[0039] In summary, the method of the present invention is novel, simple, reliable, and has a wide range of applications, making it easy to use in engineering practice. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 This is a schematic diagram of the overall flowchart of the rolling bearing RUL prediction method in an embodiment of the present invention;

[0042] Figure 2 This is a schematic diagram of the time-domain images of bearing 1_1 in the horizontal and vertical directions;

[0043] Figure 3 A schematic diagram illustrating the process of determining FPT;

[0044] Figure 4 A schematic diagram illustrating the evolution of the wavelet power image of a bearing during the degradation process;

[0045] Figure 5 This is a schematic diagram of the specific structure of the CNN-Transformer model;

[0046] Figure 6 The predicted trend of bearing 1_1 at RMES of 0.630 is shown in the graph.

[0047] Attached image labels: 110 - dashed line, 120 - dashed line, 210 - dashed line, 220 - dashed line, 300 - solid line, 400 - solid line. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0049] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0050] Example 1:

[0051] This invention presents a deep learning-based method for predicting the bearing roughness limit (RUL) of rolling bearings. It combines convolutional neural networks, which excel at extracting local information, with Transformer structures, which are adept at extracting global information, to deeply mine the complex mapping relationship between signal features and bearing RUL. Furthermore, it redefines the starting point for bearing degradation, providing a better characterization of the degradation process. The proposed method is validated using the XJTU-SY rolling bearing accelerated life experimental dataset, and its superiority is demonstrated through comparison with other methods.

[0052] The overall flowchart of this embodiment is as follows: Figure 1 As shown, it includes the following steps:

[0053] Step 1: Conduct vibration monitoring experiments on the rolling bearing until the bearing fails; obtain vibration signal data for the entire operating cycle of the rolling bearing and the final failure time.

[0054] Step 2: Calculate the kurtosis and peak factor of the bearing vibration signal, and determine the first prediction time (FPT) of the rolling bearing based on the kurtosis and peak factor.

[0055] Step 3: Perform wavelet transform on the acquired vibration signal data to obtain the wavelet power spectrum;

[0056] Step 4: Preprocess the obtained wavelet power spectrum to obtain samples. Calculate the remaining rolling bearing life at the time of sampling for each sample based on the final failure time and the first prediction time, and use this as a label to construct training samples.

[0057] Step 5: Construct a CNN-Transformer joint model, calculate the remaining life of the rolling bearing at the time of sampling for each sample as the sample label, and construct training samples;

[0058] Step 6: Train the CNN-Transformer joint model using training samples;

[0059] Step 7: Use the trained CNN-Transformer joint model to predict the RUL of the rolling bearing.

[0060] This embodiment uses the XJTU-SY rolling bearing accelerated life test dataset jointly released by the School of Mechanical Engineering of Xi'an Jiaotong University and Zhejiang Changxing Shengyang Technology Co., Ltd. to verify the effectiveness and usability of the present invention.

[0061] The dataset used in this embodiment sets up three different working conditions. The rolling bearing in this dataset has vibration acceleration sensors installed in both the horizontal and vertical directions to simultaneously collect vibration signals in both directions. The sampling frequency is 25.6 kHz, and sampling is performed every 1 minute for 1.28 seconds, resulting in the collection of 32,768 vibration point signals. The dataset is randomly divided into training and testing sets. Details of the dataset are shown in Table 1. The time-domain images of bearing 1_1 in the horizontal and vertical directions are shown in... Figure 2 As shown.

[0062] Table 1. Details of the Rolling Bearing Dataset

[0063]

[0064] Kuroism and peak value are obtained from bearing vibration signals for rolling bearing health monitoring, and the bearing RUL and FPT are determined.

[0065] 1. Calculate the first prediction time:

[0066] Kurtosis is a commonly used characteristic indicator for pitting damage and is suitable for early fault detection in machines, but it is not sensitive to wear damage. In engineering practice, the failure type and location of bearings are variable. Therefore, this invention proposes a new health monitoring method based on time-frequency characteristic parameters to determine FPT (Failure Tolerance Point). The root mean square (RMS) index is more sensitive to wear-related faults, while the peak value is more sensitive to impact-related faults. The peak factor considers both the peak value and the RMS value, thus providing good judgment results for both types of faults. The peak factor and the kurtosis index, which is more sensitive to impact faults, complement each other and are particularly suitable as a method for judging FPT in RUL (Revenue Limitation) prediction.

[0067] The establishment of FPT is based on the kurtosis X of the vibration signal. Kurtosis and peak factor X Crest The formulas for determining whether a bearing has failed are as follows:

[0068]

[0069]

[0070] Where n is the number of points in the discrete sequence, x i Let x be the vibration amplitude corresponding to the i-th discrete point in the time-domain waveform sequence. i The mean value of the bearing's kurtosis is calculated as follows: The mean μ and standard deviation σ of the bearing's kurtosis during the early operating phase are calculated. According to the 3σ principle of normal distribution, the probability that the peak value of the vibration signal under normal conditions is distributed in (μ-3σ, μ+3σ) is 99.74%. When two consecutive kurtosis values ​​exceed this range, it is considered that the bearing has begun to degrade and is an impact-type failure. This time is taken as the first pre-selected time when the bearing begins to degrade.

[0071] Simultaneously, the mean μ and standard deviation σ of the peak factor of the bearing are calculated in the early operation stage. When two consecutive peak factors exceed the (μ-3σ, μ+3σ) interval of the peak factor, it is considered that the bearing has begun to degrade, which is a wear-type failure. This time is taken as the time when the bearing begins to degrade and is the second pre-selected time.

[0072] The minimum value between the first and second pre-selected times is taken as the first predicted time when the bearing begins to degrade.

[0073] Taking bearing 1_1 as an example, the visualization of this process is as follows: Figure 3 The solid line 300 represents the kurtosis of the bearing throughout its entire operating phase, the solid line 400 represents the peak factor of the bearing throughout its entire operating phase, the dashed lines 110 and 120 represent the normal distribution intervals of the kurtosis, and the dashed lines 210 and 220 represent the normal distribution intervals of the peak factor. Considering the different sensitivities of the two indicators to the bearing condition, and considering that under both indicators, the impact and wear failures suppress the interference of random noise and effectively capture the early failure signals of the bearing, the earliest time of degradation obtained under each of the two indicators is used as the FPT to express the first prediction time of the start of degradation in the RUL prediction process of the rolling bearing.

[0074] In the vibration monitoring experiment, vibration signals in both the horizontal and vertical directions are monitored. When calculating the first prediction time, the first prediction time is calculated for both types of vibration signals, and the earliest first prediction time is selected for calculating the remaining life of the rolling bearing.

[0075] 2. Wavelet Transform: The initial vibration signal is converted into a wavelet power spectrum using Morlet wavelet CWT, which is used to predict the bearing degradation performance characteristics in bearing RUL prediction.

[0076] For fault signals in mechanical equipment, the frequency components change over time, requiring a joint representation in the time and frequency domains. Wavelet transform utilizes wavelets as basis functions, incorporating both scale and time variables, and obtains the correlation coefficient between the signal and the wavelet basis functions through inner product transformation. Therefore, wavelet transform is an effective time-frequency localization method, suitable for transient analysis of non-stationary signals.

[0077] The continuous wavelet transform has greater resolution and greater freedom in choosing basis functions, representing continuous signals f(t) ∈ L 2 (R) Wavelet transform with time t, such as:

[0078]

[0079] a represents the frequency-dependent scaling factor, b represents the time shift factor, and ψ * This indicates that the complex conjugate of the wavelet function was used in the transformation. Different wavelet bases ψ(t) yield different signal characterizations. The basic principle for determining the wavelet base is to select a wavelet with a shape similar to the fault vibration signal. The vibration signal of a faulty bearing is similar to the Morlet wavelet. The Morlet wavelet is a complex wavelet, which can extract the amplitude and phase information of the analyzed time process or signal. f0 is the center frequency of the mother wavelet and f0≥0, therefore its expression is simplified as follows:

[0080]

[0081] In formula (4), i is the imaginary part. The time-frequency image extracted from the continuous wavelet transform (i.e., the contour plot of the power spectrum of each frequency component) effectively contains the degradation information of the rolling bearing during operation.

[0082] The vibration time-series signal is transformed using wavelet transform to obtain the wavelet coefficient matrix. The wavelet coefficients are squared to obtain the wavelet energy, which is then superimposed with the wavelet energy corresponding to each scale factor to obtain the wavelet energy spectrum curve that varies with the scale factor. Converting the scale to frequency yields the wavelet power image.

[0083] 3. Preprocessing of wavelet power images: The coordinate axes and edges of the plotted image data are removed, leaving only the portion containing bearing degradation information, and the dimension is reduced to 64×64 (M=64). Based on the full-cycle vibration signals of 15 bearings in the above dataset, 3807 wavelet power images with a dimension of 64×64 are obtained, some of which are as follows: Figure 4 As shown, Figure 4The image shows partial wavelet power maps of bearings 1-1, 1-2, 1-3, and 1-5 at different degradation levels. The wavelet power maps randomly select samples from left to right according to the sampling time order. The degradation process is intuitively reflected in the color changes of the image. The horizontal axis represents the degree of degradation, with 0% on the left indicating that degradation has just occurred and 100% on the right indicating that degradation is complete.

[0084] 4. Construct training samples:

[0085] The training set of the model is (x i ,y i ), x i For the wavelet power map of the i-th sample, y i Let be the label value of the sample, representing the actual remaining life of the rolling bearing at the current moment as a percentage. Based on the first prediction time (FPT) determined for each bearing, before this time point, the remaining life of the bearing is 100%; after this time point, the remaining life of the rolling bearing degrades linearly over time until it reaches 0%. Let t be the time when the bearing finally fails, t1 be the time when the bearing begins to degrade at the first sample point, and t be the time when the bearing is sampled at the ith point. i Therefore, after the degradation begins, y i The calculation method is as follows:

[0086]

[0087] 5. Construct a deep learning network model. When the accuracy of the test samples meets the actual requirements, output the model and use it to predict the RUL of the bearing.

[0088] In the proposed CNN-Transformer joint model, a traditional CNN structure is added before the Transformer structure, and the GELU activation function with random regularization is used to improve the model's generalization ability. Since the model's input is a fixed-size image matrix, the input and output matrix sizes remain fixed after convolution and other operations, thus positional embedding is unnecessary. The model retains the original encoder and decoder structure of the Transformer, directly generating the prediction result in the last layer of the decoder. For each input vector of the encoder, a linear mapping is performed to three vectors q, k, and v, where q and k... T The attention weight matrix is ​​obtained by performing a dot product operation. After weighted normalization using Softmax, the matrix is ​​multiplied by v and then input into a three-layer perceptron for fitting. The specific structure of the model is as follows: Figure 5As shown. The convolutional kernels are all 3×3 in size, and the number of kernels is set to 3, 9, 27, 27, and 27 respectively. The input and output sizes are (64, 64, 3) and 1 respectively. The entire network uses GELU activation and is normalized using BatchNormalization.

[0089] The CNN-Transformer joint model comprises a CNN convolutional neural network, an encoder, and a decoder connected in sequence. The encoder and decoder constitute the Transformer model architecture. The CNN convolutional neural network includes a combination of convolutional layers, activation layers, max pooling layers, batch normalization layers, mean pooling layers, and flattening layers. The output of the flattening layers serves as the input to the encoder. When combined, as... Figure 5 As shown, the first learning layer consists of a convolutional layer, an activation layer, and a max pooling layer, with a kernel size of 3 for the convolutional layers in the first learning layer. The second learning layer consists of a convolutional layer, an activation layer, and a max pooling layer, with a kernel size of 9 for the convolutional layers in the second learning layer. The third learning layer consists of three convolutional layers, one batch normalization layer, one activation layer, and one max pooling layer, with each of the three convolutional layers in the third learning layer having a kernel size of 27.

[0090] The encoder includes a self-attention module and a fully connected feedforward network; the self-attention module performs linear mapping and normalization on the input variables; the fully connected feedforward network includes a fully connected FC layer, an activation layer, and a fully connected FC layer.

[0091] The decoder structure is basically the same as the encoder structure, and the prediction result is output in the last layer.

[0092] All activation layers are activated using GeLU Gaussian error linear units, and the mean pooling layer is AvgPoo12d mean pooling.

[0093] 6. Train the CNN-Transformer joint model using training samples: During training, randomly select the 3rd and 5th bearings under each operating condition as the test set, and the remaining bearings as the training set. Set the number of iterations to 300, the learning rate to 0.01, and the batch size to 128.

[0094] This invention uses Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE) as standards to evaluate the performance of the proposed method. RMSE measures the deviation between the true value and the predicted value. As a rule of thumb, an RMSE value between 0.2 and 0.5 indicates that the model can accurately predict the RUL. MAE is another loss function used in regression models, which is the sum of the absolute values ​​of the differences between the target value and the predicted value. The above two labels are calculated for a set of continuous prediction results with a sample size of n. i It is the actual value. The predicted values, and the formulas for calculating the two indicators are as follows:

[0095]

[0096]

[0097] Test experiments were conducted on the model with the above parameter settings. Three sets of data were collected for each bearing, and the results are shown in Table 2. The average and minimum values ​​of RMSE and MAE for the three sets of data are shown in Table 3. Figure 6 The figure shows the predicted trend of bearing 1_1 at an RMES of 0.630, indicating that the bearing RUL prediction model is performing well. The predicted life trend after smoothing of the six test results is basically consistent with the actual life curve, with a small deviation from the actual value.

[0098] Table 2. Three sets of prediction errors for the CNN-Transformer model

[0099]

[0100]

[0101] Table 3. Average and minimum prediction errors of the CNN-Transformer model

[0102]

[0103] Example 2: Example 2 of this disclosure provides an online prediction method for the remaining service life of rolling bearings based on wavelet power spectrum and Transformer structure.

[0104] The cloud deployment process of the RUL prediction model is as follows: First, the trained model is saved as a file with the .pth extension. Then, the process of processing the bearing vibration signal uploaded from the webpage, loading the model, and performing RUL prediction are put into a Python script. Finally, the output of the Python script is passed to the HTML webpage.

[0105] After completing the cloud deployment of the model, the vibration signal file of the bearing is uploaded to the client. On the server side, a Python script is called through the controller under the ThinkPHP framework, and the output is then directed to the view page. This allows sensors to collect vibration signals from the running rolling bearing. By uploading the signal file for a specific sampling period to the life prediction page, the remaining lifespan and wavelet power image of the bearing during this operating phase can be obtained, thus achieving online prediction of the rolling bearing's RUL (Remaining Lifespan).

[0106] Example 3: The present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the rolling bearing RUL prediction method.

[0107] Example 4: The present invention provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the rolling bearing RUL prediction method.

[0108] Example 5: The present invention provides a computer program product, including a computer program, which, when executed by a processor, implements the rolling bearing RUL prediction method.

[0109] The above description is merely a preferred embodiment of this disclosure and is not intended to limit this disclosure. Various modifications and variations can be made to this disclosure by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A deep learning-based method for predicting the RUL (Round Usage Limit) of rolling bearings, characterized in that, Includes the following steps: Vibration monitoring experiments were conducted on the rolling bearing until the bearing failed; vibration signal data of the rolling bearing throughout its entire operating cycle and the final failure time were obtained. Calculate the kurtosis and peak value of the bearing vibration signal, and determine the first prediction time of the rolling bearing based on the kurtosis and peak value. Wavelet transform is performed on the acquired vibration signal data to obtain the wavelet power spectrum. The obtained wavelet power spectrum is preprocessed to obtain samples. The remaining rolling bearing life at the time of sampling of each sample is calculated based on the final failure time and the first prediction time as a label to construct training samples. Construct a CNN-Transformer joint model, calculate the remaining life of the rolling bearing at the time of each sample sampling as the sample label, and construct training samples; The CNN-Transformer joint model is trained using training samples; RUL prediction of rolling bearings is performed using a trained CNN-Transformer joint model; The method for determining the first prediction time for each rolling bearing based on kurtosis and peak factor is as follows: Calculate the mean kurtosis of the bearing during the early operating phase. and standard deviation ; According to the normal distribution In principle, the time when the bearing begins to degrade is determined as the first pre-selected time; Calculate the mean of the peak factor of the bearing during the early operating phase. and standard deviation According to the normal distribution In principle, the time when the bearing begins to degrade is the second pre-selected time; The minimum value between the first and second pre-selected times is taken as the first predicted time when the bearing begins to degrade. In the vibration monitoring experiment, vibration signals in both the horizontal and vertical directions are monitored. When calculating the first prediction time, the first prediction time under the two vibration signals is calculated separately, and the earliest first prediction time is selected for the remaining life calculation of the rolling bearing. The remaining life of a rolling bearing is calculated as follows: The remaining lifetime before the first prediction time is set to 100%, and the remaining lifetime at the bearing's final failure is set to 0%. The formula for calculating the remaining lifetime between the first prediction time and the bearing's final failure is as follows: In the above formula, For the first i The remaining life of the bearing at the time of sample collection. t This refers to the final failure time of the bearing. This is the first predicted time for the bearing. Is the bearing in the first The time required to collect each sample.

2. The deep learning-based rolling bearing RUL prediction method according to claim 1, characterized in that: The wavelet transform method is as follows: The wavelet coefficient matrix is ​​obtained by transforming the vibration time series signal using wavelet transform. The wavelet energy obtained by squaring the wavelet coefficients is superimposed with the wavelet energy corresponding to each scale factor to obtain the wavelet energy spectrum curve that varies with the scale factor. The wavelet power image is obtained by converting the scale into frequency.

3. The deep learning-based rolling bearing RUL prediction method according to claim 2, characterized in that: The preprocessing method for the wavelet power spectrum is as follows: After removing the coordinate axes and edges from the wavelet power spectrum, a dimensionality reduction process is performed to obtain an M×M wavelet power spectrum.

4. The deep learning-based rolling bearing RUL prediction method according to claim 1, characterized in that: The CNN-Transformer joint model includes a CNN convolutional neural network, an encoder, and a decoder connected in sequence. The encoder and decoder constitute the Transformer model architecture. The CNN convolutional neural network includes a convolutional layer, an activation layer, a max pooling layer, a batch normalization layer, a mean pooling layer, and a flattening layer. The output of the flattening layer serves as the input of the encoder. The encoder includes a self-attention module and a fully connected feedforward network; the self-attention module performs linear mapping and normalization on the input variables; the fully connected feedforward network includes a linear mapping layer, a normalization layer, a fully connected (FC) layer, an activation layer, and a fully connected (FC) layer. The decoder structure is the same as the encoder structure, and it outputs the prediction result in the last layer.

5. An electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the rolling bearing RUL prediction method as described in any one of claims 1 to 4.

6. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the rolling bearing RUL prediction method as described in any one of claims 1 to 4.

7. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the rolling bearing RUL prediction method as described in any one of claims 1 to 4.

Images