A method, system, and medium for predicting the remaining life of rolling bearings.

By improving the temporal convolutional network and combining multi-sensor signal processing and feature extraction methods, the problem of low accuracy in rolling bearing life prediction was solved, and higher accuracy life prediction was achieved.

CN117744487BActive Publication Date: 2026-11-10CIVIL AVIATION UNIV OF CHINA
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202311765498.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-12-21
Publication Date
2026-11-10
Estimated Expiration
2043-12-21

AI Technical Summary

Technical Problem

Existing technologies for predicting the life of rolling bearings have low accuracy and struggle to effectively capture short-term, medium-term, and long-term time correlations.

Method used

An improved temporal convolutional network is used to extract bearing performance degradation features through multi-level feature extraction methods, including normalization of multi-sensor signals, sliding window slicing, hierarchical dilated temporal convolution, channel attention processing, and matrix multiplication dimensionality reduction.

Benefits of technology

It improves the accuracy of rolling bearing life prediction, enhances the ability to capture short-term, medium-term and long-term time correlations, and strengthens the stability and efficiency of prediction.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117744487B_ABST
    Figure CN117744487B_ABST
Patent Text Reader

Abstract

This invention relates to a method, system, and medium for predicting the remaining life of rolling bearings. After normalizing and preprocessing the acquired multi-sensor signals of the bearing, a sliding window slicing process is performed according to the sampling time series to obtain a slice dataset. For any slice dataset, hierarchical dilated temporal convolution is performed, and the outputs of each of the three temporal convolution layers are extracted and stacked to obtain a first three-dimensional feature data block. Channel attention processing is then applied to the first three-dimensional feature data block to obtain a second three-dimensional feature data block, followed by matrix multiplication dimensionality reduction to obtain a corresponding one-dimensional feature vector indicating bearing performance degradation. The one-dimensional feature vector is then processed through a fully connected layer to obtain prediction point values. Connecting all prediction point values ​​sequentially yields the remaining life prediction curve. This application's solution can more fully capture the advantages of short-term, medium-term, and long-term temporal correlations, thereby improving the accuracy of rolling bearing life prediction.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of fault prediction and health management technology, specifically to a method, system, and medium for predicting the remaining life of rolling bearings. Background Technology

[0002] Rolling bearings are a crucial component of modern machinery, promising to optimize equipment performance and extend service life while reducing vibration and shock, ensuring machine stability. As a core part of rotating machinery, bearings operate in extremely challenging environments. Their complexity and unpredictability make them susceptible to factors such as wear and corrosion, potentially leading to equipment failure, impacting operational efficiency, and even posing safety hazards.

[0003] Remaining life prediction methods are generally divided into two categories: model-based methods and data-driven methods. Model-based methods mainly involve the development of mathematical or physical degradation models based on failure mechanisms to describe the degradation process of mechanical components. However, the specific framework of data-driven prediction methods is equally crucial. In recent years, deep neural networks have been introduced to replace feature extraction in data-driven methods and have been applied to the prediction field. However, the accuracy of predictions for rolling bearings still remains a problem. Summary of the Invention

[0004] The technical problem to be solved by this application is to provide a method, system and medium for predicting the remaining life of rolling bearings, which has the feature of higher accuracy in predicting the remaining life of rolling bearings.

[0005] In a first aspect, one embodiment provides a method for predicting the remaining life of a rolling bearing, comprising:

[0006] Acquire multi-sensor signals from the bearing;

[0007] The multi-sensor signals are subjected to normalization preprocessing;

[0008] The normalized preprocessed multi-sensor signals are sliced ​​using a sliding window according to the sampling time series to obtain a slice dataset, which includes multiple slice data.

[0009] For any slice of data, perform hierarchical dilated temporal convolution processing, including: feeding the slice of data into a temporal convolutional network to achieve three layers of temporal convolution processing; wherein, each temporal convolution layer uses a convolutional kernel of a different size, and in two adjacent temporal convolution layers, the output of the previous layer is used as the input of the next layer;

[0010] For any slice of data, extract the output of each of the three temporal convolutional layers and stack them to obtain the first three-dimensional feature data block;

[0011] For any slice of data, channel attention processing is performed on the first three-dimensional feature data block to obtain the second three-dimensional feature data block;

[0012] For any slice of data, matrix multiplication is used to reduce the dimensionality, so as to obtain the corresponding one-dimensional feature vector that indicates the performance degradation of the bearing.

[0013] By performing a fully connected operation on the one-dimensional feature vector of any slice of data, the predicted point values ​​are obtained. By sequentially connecting all the predicted point values, the predicted curve of the remaining life of the rolling bearing is obtained.

[0014] In one embodiment, the normalization preprocessing of the multi-sensor signals includes:

[0015]

[0016] in, This represents the data acquired from a sensor x on the bearing at time t. X represents the normalized data. min X represents the minimum value in the dataset obtained for a specific sensor x. max This represents the maximum value obtained from the dataset of a specific sensor x.

[0017] In one embodiment, each temporal convolution layer uses a convolution kernel of a different size, including: in two adjacent temporal convolution layers, the convolution kernel of the current temporal convolution layer is larger than the convolution kernel of the previous temporal convolution layer.

[0018] In one embodiment, feeding the sliced ​​data into a temporal convolutional network to achieve three-layer temporal convolution processing includes:

[0019]

[0020]

[0021]

[0022] Among them, V d=1 This represents the output of the first temporal convolution layer, where d represents the dilation factor, and M represents the value of M. T×μ V represents a two-dimensional tensor of sliced ​​data, where T represents the time dimension, μ represents the feature dimension, and α represents the width of the convolutional layer generated during the convolution process of the first temporal convolutional network; d=2 V represents the output of the second temporal convolution, and β represents the width of the convolutional layer generated during the first temporal convolution process; d=4 This represents the output of the third temporal convolution, where γ represents the width of the convolutional layer generated during the third temporal convolution process; k Tk represents the size of the convolution kernel. T3 >k T2 >k T1 .

[0023] In one embodiment, the step of extracting the output of each layer of the three-layer temporal convolutional network and stacking them to obtain a first three-dimensional feature data block for any slice of data includes:

[0024] Using the maximum width of the three layers of the hierarchically dilated temporal convolution as a reference, the outputs of each of the three temporal convolutional layers are zero-paded in the width dimension and then stacked to obtain the first three-dimensional feature data block; the length of the zero-padding sequence includes:

[0025]

[0026] Where, δ p κ represents the length of the zero-padded sequence. p κ represents the width of any one of the three layers. p =α,β,γ, where p represents the index of the slice data.

[0027] In one embodiment, the step of performing matrix multiplication dimensionality reduction processing on any slice of data to obtain a corresponding one-dimensional feature vector indicating bearing performance degradation includes:

[0028] The three layers of two-dimensional data in the second three-dimensional feature data block are multiplied element by element with the corresponding column vectors to obtain three column vectors after dimensionality reduction. The three column vectors are aligned, summed and added to obtain the corresponding one-dimensional feature vector indicating the performance degradation of the bearing.

[0029] In one embodiment, the step of multiplying the three layers of two-dimensional data in the second three-dimensional feature data block element-wise with the corresponding column vectors to obtain three column vectors after dimensionality reduction, and then aligning and summing the three column vectors to obtain the corresponding one-dimensional feature vector indicating bearing performance degradation, includes:

[0030]

[0031] Among them, Z ξ Let l represent the corresponding one-dimensional feature vector indicating bearing performance degradation, and l represent the index of the two-dimensional data layer, where 1 ≤ l ≤ L. This represents the two-dimensional data of the corresponding layer. This represents the column vector of the corresponding layer.

[0032] In one embodiment, the process of fully connecting the one-dimensional feature vector of any slice of data includes:

[0033] The one-dimensional feature vector of any slice of data is processed by two fully connected layers.

[0034] Secondly, this application provides a rolling bearing remaining life prediction system, including a memory and a processor. The memory stores a program that can be loaded and run by the processor, and the program includes the rolling bearing remaining life prediction method described in any of the above embodiments.

[0035] Thirdly, this application provides a computer-readable storage medium storing a program that can be loaded by a processor and executed by the rolling bearing remaining life prediction method described in any of the above embodiments.

[0036] The beneficial effects of this invention are:

[0037] Since the proposed solution is based on an improved temporal convolutional network, it extracts multi-level feature changes from time series, which can more fully capture the advantages of short-term, medium-term and long-term temporal correlations, thereby improving the accuracy of prediction in the life prediction of rolling bearings. Attached Figure Description

[0038] Figure 1 This is a flowchart illustrating a method for predicting the remaining life of rolling bearings according to an embodiment of this application.

[0039] Figure 2 This application Figure 1 A schematic diagram of the first layer of dilated temporal convolution processing in a hierarchical dilated temporal convolution processing embodiment of step S40;

[0040] Figure 3 This application Figure 1 A schematic diagram of the second layer of dilated temporal convolution processing in a hierarchical dilated temporal convolution processing step S40 of one embodiment;

[0041] Figure 4 This application Figure 1 A schematic diagram of the third layer of dilated temporal convolution processing in a hierarchical dilated temporal convolution processing embodiment of step S40;

[0042] Figure 5 This application Figure 1 A schematic diagram of the global and local feature data stacking process in one embodiment of step S50;

[0043] Figure 6 This application Figure 1 A schematic diagram of the process of fusing multidimensional features in one embodiment of step S70;

[0044] Figure 7 This is a schematic diagram of the RUL prediction results for bearing 1-1 in the test set of this application;

[0045] Figure 8 This is a schematic diagram of the RUL prediction results for bearings 1-5 in the test set of this application;

[0046] Figure 9 This is a schematic diagram of the RUL prediction results for bearings 2-4 in the test set of this application;

[0047] Figure 10 This is a schematic diagram of the predicted results of bearing 2-5RUL in the test set of this application. Detailed Implementation

[0048] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. Similar elements in different embodiments are referred to by associated similar element reference numerals. In the following embodiments, many details are described to facilitate a better understanding of the present application. However, those skilled in the art will readily recognize that some features may be omitted in different situations, or may be replaced by other elements, materials, or methods. In some cases, certain operations related to the present application are not shown or described in the specification. This is to avoid obscuring the core parts of the present application with excessive description. For those skilled in the art, detailed description of these related operations is not necessary; they can fully understand the related operations based on the description in the specification and general technical knowledge in the art.

[0049] Furthermore, the features, operations, or characteristics described in the specification can be combined in any suitable manner to form various embodiments. At the same time, the steps or actions in the method description can be rearranged or adjusted in a manner obvious to those skilled in the art. Therefore, the various orders in the specification and drawings are only for the clear description of a particular embodiment and do not imply a necessary order, unless otherwise stated that a particular order must be followed.

[0050] The serial numbers assigned to components in this document, such as "first" and "second," are used only to distinguish the described objects and have no sequential or technical meaning. The terms "connection" and "linkage" used in this application, unless otherwise specified, include both direct and indirect connections (linkages).

[0051] To facilitate the explanation of the inventive concept of this application, the rolling bearing life prediction technology will be briefly described below.

[0052] Real-time health monitoring and remaining useful life (RUL) prediction during the operation of rolling bearings are crucial. In recent years, this technology has become a focal point of Predictive Failure and Health Management (PHM), and has significant practical implications for ensuring the stable and safe operation of mechanical equipment.

[0053] The feature extraction capability of conventional neural networks is limited by their structure, and most are sequential neural networks. They tend to focus more on high-level features in the final output, rarely considering the influence of low-level features. This can lead to the loss of useful degradation feature information, making it insufficient to capture all features related to the degradation state. Furthermore, few studies have considered fusing features from multiple sensors to extract performance degradation features, resulting in the current problem of low prediction accuracy in rolling bearing life prediction.

[0054] In view of the above problems, this application provides a method for predicting the remaining life of rolling bearings. This method is based on an improved temporal convolutional network, which extracts multi-level feature changes from time series data. This allows for a more comprehensive capture of short-term, medium-term, and long-term temporal correlations, thereby improving the accuracy of rolling bearing life prediction. Please refer to... Figure 1 The method includes:

[0055] Step S10: Acquire the multi-sensor signals of the bearing.

[0056] In one embodiment, the system may include horizontal and vertical signals for acquiring vibration data of the bearing, with the sensor recording the horizontal and vertical vibration signals at a sampling rate of 232,768 data points per minute.

[0057] Step S20: Perform normalization preprocessing on the multi-sensor signals.

[0058] In one embodiment, the normalization preprocessing of multi-sensor signals includes:

[0059]

[0060] in, This represents the data acquired from a sensor x on the bearing at time t. X represents the normalized data. min X represents the minimum value in the dataset obtained for a specific sensor x. max This represents the maximum value obtained from the dataset of a specific sensor x.

[0061] Step S30: The normalized preprocessed multi-sensor signals are processed by sliding window slicing according to the sampling time series to obtain a slice dataset, which includes multiple slice data.

[0062] In one embodiment, a sliding window of size T is used to segment the sampling time series. Starting from the first operation point, data with time step T are selected according to the implementation order to generate the first sample M1 = {m1, m2, ..., m}. T}∈i T ×μThen, the sliding window is moved to the second operation point to create the second sample M2 = {m2, m3, ..., m T+1}∈i T×μ , where i T×μ This represents a two-dimensional bearing time series data of size T×μ, where μ represents the feature dimension (the number of sensor types acquired). This process is repeated until the sliding window covers all the data. All slices generated by the sliding window are stacked together as a three-dimensional data block. Each slice can be represented by a two-dimensional tensor, whose dimensions are the width and height of the three-dimensional data block.

[0063] Step S40: For any slice of data, perform hierarchical dilated temporal convolution processing, including: feeding the slice data into a temporal convolutional network to achieve three-layer temporal convolution processing; wherein, each temporal convolution layer uses a convolutional kernel of a different size, and in two adjacent temporal convolution layers, the output of the previous layer is used as the input of the next layer.

[0064] The preprocessed slice data 2D tensor is fed into a temporal convolutional network, and the slice data is processed by three layers of temporal convolution based on the constructed three-layer temporal convolutional network.

[0065] In one embodiment, in two adjacent temporal convolutional layers, the kernel of the current temporal convolution is larger than the kernel of the previous temporal convolution. Therefore, the kernel sizes, from largest to smallest, are: kernel size of the third layer > kernel size of the second layer > kernel size of the first layer. The input two-dimensional tensor has a temporal dimension of T and a feature dimension of μ.

[0066] Please refer to Figure 2 , Figure 3 and Figure 4 The convolutional kernel moves along the time dimension to obtain temporal and feature information from the input data, resulting in a channel dimension ( Figure 2 The middle is α, Figure 3 The middle is β, Figure 4 The value of γ depends on the number of filters (i.e., convolution kernels).

[0067] In one embodiment, sliced ​​data is fed into a temporal convolutional network to achieve three-layer temporal convolutional processing, including:

[0068]

[0069]

[0070]

[0071] In this system, each layer of the temporal convolutional network generates an intermediate output, where V represents the output of the convolutional layer. d=1This represents the output of the first-layer temporal convolutional network, where d represents the dilation factor, and M represents the output of the first-layer temporal convolutional network. T×μ V represents a two-dimensional tensor representing sliced ​​data, where α represents the width of the convolutional layer generated during the first temporal convolutional network convolution process; d=2 V represents the output of the second temporal convolutional network, and β represents the width of the convolutional layer generated during the convolution process of the first temporal convolutional network; d=4 This represents the output of the third temporal convolutional network, where γ represents the width of the convolutional layer generated during the convolution process of the third temporal convolutional network; k T k represents the size of the convolution kernel. T3 >k T2 >k T1 .

[0072] The three temporal convolutional layers employ different dilation factors: 1 for the first layer, 2 for the second, and 4 for the third, to expand the receptive field. The exponential growth of the dilation factor leads to an exponential increase in the receptive field size. Even with only a few convolutional layers, a large receptive field can be achieved while maintaining high computational efficiency. These dilation factors represent different levels of feature extraction, covering local and global features in the bearing degradation stage, corresponding to the capture of short-term, medium-term, and long-term information during the bearing degradation process.

[0073] The hierarchical dilated temporal convolution processing proposed in this application can better handle features at different levels and fully leverage the advantage of effectively capturing short-term and long-term temporal correlations.

[0074] Step S50: For any slice of data, extract the output of each layer of the three-layer temporal convolutional network and stack them to obtain the first three-dimensional feature data block.

[0075] In one embodiment, please refer to Figure 5 For any slice of data, the output of each layer of the three-layer temporal convolutional network is extracted and stacked to obtain the first three-dimensional feature data block G. s ,include:

[0076] Using the maximum width of the three layers of the hierarchically dilated temporal convolution as a reference, the outputs of each of the three temporal convolutional networks are zero-paded along the width dimension and then stacked to obtain the first 3D feature data block; the length of the zero-padding sequence includes:

[0077]

[0078] Where, δ p κ represents the length of the zero-padded sequence. p κ represents the width of any one of the three layers. p =α,β,γ, where p represents the index of the slice data.

[0079] Based on the above embodiments, the sensor data are stacked by combining the global features corresponding to the last convolutional layer in the temporal convolution process and the local features corresponding to the first and second convolutional layers to form a first three-dimensional feature data block G. s The process of stacking feature data involves stacking the local and global feature matrices extracted by the hierarchical dilated temporal convolutional network to obtain the first three-dimensional feature data block G. s This process allows local features to be incorporated into subsequent computations, thus enabling the mixing of global and local features.

[0080] The feature data stacking proposed in this application extracts more comprehensive and accurate information, enhances the diversity of data, and integrates features from different time scales. All features are included in the prediction, thereby improving the accuracy of bearing remaining life prediction.

[0081] Step S60: For any slice of data, perform channel attention processing on the first three-dimensional feature data block to obtain the second three-dimensional feature data block.

[0082] By using channel attention processing, channels containing important features in both global and local features are given larger weights, while channels with unnecessary features are given smaller weights, resulting in a new data feature block adjusted by channel attention, namely the second three-dimensional feature data block.

[0083] Step S70: For any slice of data, perform matrix multiplication to reduce dimensionality and obtain the corresponding one-dimensional feature vector indicating bearing performance degradation.

[0084] In one embodiment, for any slice of data, matrix multiplication is performed to reduce dimensionality, resulting in a corresponding one-dimensional feature vector indicating bearing performance degradation, including:

[0085] The three layers of two-dimensional data in the second three-dimensional feature data block are multiplied element by element with the corresponding column vectors to obtain three column vectors after dimensionality reduction. The three column vectors are aligned, summed and added to obtain the corresponding one-dimensional feature vector indicating the performance degradation of the bearing.

[0086] In this embodiment, the second three-dimensional feature data block, after channel attention adjustment, is further subjected to matrix multiplication dimensionality reduction transformation to fuse multidimensional features, ultimately obtaining a one-dimensional feature vector indicating bearing performance degradation. Furthermore, the innovative matrix multiplication dimensionality reduction transformation method employed in this application enhances the stability of the overall rolling bearing remaining life prediction while reducing data dimensionality and merging multidimensional features, and also improves computational efficiency.

[0087] In one embodiment, please refer to Figure 6This demonstrates the process of this dimensionality reduction transformation. In the structure of deep learning networks, a "fusion layer" is introduced after channel attention processing. The input to this fusion layer is a second 3D feature data block. Second 3D feature data block Element-wise multiplication of the three corresponding vectors in the fusion layer along the channel dimension yields a result with dimension [missing value]. The three column vectors. Finally, for the dimension of The three column vectors are aligned, summed, and added together to achieve the fusion of global and local features. Finally, the second three-dimensional feature data block... When passing through this fusion layer, the three-dimensional matrix is ​​transformed into a one-dimensional eigenvector Z that indicates bearing performance degradation. ξ ,include:

[0088]

[0089] Among them, Z ξ Let l represent the corresponding one-dimensional feature vector indicating bearing performance degradation, and l represent the index of the two-dimensional data layer, where 1 ≤ l ≤ L. This represents the two-dimensional data of the corresponding layer. This represents the column vector of the corresponding layer. and These represent the second and third-dimensional feature data blocks, respectively. Height and width.

[0090] By using the above methods, the dimensionality of the data is reduced and multidimensional features are merged, allowing for the extraction of more representative degradation information, thereby improving prediction efficiency.

[0091] Step S80: The one-dimensional feature vector of any slice of data is processed by a fully connected layer to obtain the predicted point value. All the predicted point values ​​are then connected in sequence to obtain the predicted curve of the remaining life of the rolling bearing.

[0092] In one embodiment, processing the one-dimensional feature vector of any slice of data through a fully connected layer includes: processing the one-dimensional feature vector of any slice of data through a fully connected layer once, and then processing it through a fully connected layer again, for a total of two fully connected layer processing steps.

[0093] Therefore, since the temporal dilation convolution process involves three dilations, and the extracted features represent the degradation process of long-term, medium-term, and short-term bearing features, the final prediction result includes both the predicted RUL value and the contribution rate of each feature data to RUL.

[0094] To verify the effectiveness of the aforementioned method for predicting the remaining life of rolling bearings, we conducted validation experiments using the XJTU-SY bearing dataset. The experimental setup consisted of an AC motor, a speed controller, a shaft, bearings, and a hydraulic loading system. The experiment began with direct drive from the AC motor, using the speed controller to precisely adjust the output speed of the shaft motor. Then, two unidirectional accelerometers were installed on both sides (vertical and horizontal) of the test bearing to collect vibration acceleration signals. Finally, the hydraulic loading system was applied to the test bearing. The azimuth signal was sampled every 1 minute, with each sampling lasting 1.28 seconds at a sampling frequency of 25600 Hz; therefore, the number of sampling points was 32768.

[0095] To verify the accuracy and robustness of the proposed prediction method, two sets of evaluation metrics are introduced: one set quantifies the fitting effect of the proposed model, and the other set evaluates the performance of actual RUL predictions. Therefore, the mean absolute error (MAE) and root mean square error (RMSE) are first used to evaluate the accuracy of the fitting effect, defined as follows:

[0096]

[0097]

[0098] Where: R i This represents the predicted value of the neural network. σ represents the actual label value, σ is the number of test units, and i represents the index of the test unit. The smaller the value of the above evaluation index, the better the prediction fit.

[0099] Experiment: To verify the superiority of this method, a comparative experiment was conducted with existing research methods based on the verification set bearings 1_1, 1_5, 2_4 and 2_5 under two working conditions. Figure 7 , Figure 8 , Figure 9 and Figure 10 The figure shows the prediction results of the method in this application. In this experiment, existing technologies such as Long Short-Term Memory Neural Network (LSTM), Gate Recurrent Unit Neural Network (GRU) and Temporal Convolutional Network (TCN) were also used to compare with the method of this invention. Table 1 lists the detailed values ​​of the two evaluation indicators.

[0100] Table 1

[0101]

[0102]

[0103] Experimental results show that, for any evaluation metric, the proposed method outperforms other existing methods. Analysis of the average evaluation metrics for four bearings reveals that, compared to the LSTM, GRU, and TCN models in existing methods, the proposed method reduces RMSE by 44.7%, 36.4%, and 36.3%, respectively, and MAE by 50.9%, 48.9%, and 51.9%, respectively.

[0104] In one embodiment, the application provides a rolling bearing remaining life prediction system, including a memory and a processor. The memory stores a program that can be loaded and run by the processor, and the program includes the rolling bearing remaining life prediction method of any of the above embodiments.

[0105] One embodiment of this application provides a computer-readable storage medium storing a program, the stored program including the rolling bearing remaining life prediction method of any of the above embodiments, which can be loaded by a processor and processed.

[0106] Those skilled in the art will understand that all or part of the functions of the various methods in the above embodiments can be implemented by hardware or by computer programs. When all or part of the functions in the above embodiments are implemented by computer programs, the program can be stored in a computer-readable storage medium, which may include: read-only memory, random access memory, disk, optical disk, hard disk, etc., and the program is executed by a computer to achieve the above functions. For example, the program can be stored in the memory of a device, and when the program in the memory is executed by the processor, all or part of the above functions can be achieved. In addition, when all or part of the functions in the above embodiments are implemented by computer programs, the program can also be stored in a server, another computer, disk, optical disk, flash drive, or external hard drive, etc., and can be downloaded or copied to the memory of a local device, or the system of the local device can be updated. When the program in the memory is executed by the processor, all or part of the functions in the above embodiments can be achieved.

[0107] The above examples illustrate the present invention only to aid in understanding it and are not intended to limit the scope of the invention. Those skilled in the art can make various simple deductions, modifications, or substitutions based on the principles of this invention.

Claims

1. A method for predicting the remaining life of a rolling bearing, characterized in that, include: Acquire multi-sensor signals from the bearing; The multi-sensor signals are subjected to normalization preprocessing; The normalized preprocessed multi-sensor signals are sliced ​​using a sliding window according to the sampling time series to obtain a slice dataset, which includes multiple slice data. For any slice of data, perform hierarchical dilated temporal convolution processing, including: feeding the slice of data into a temporal convolutional network to achieve three layers of temporal convolution processing; wherein, each temporal convolution layer uses a convolutional kernel of a different size, and in two adjacent temporal convolution layers, the output of the previous layer is used as the input of the next layer; For any slice of data, extract the output of each of the three temporal convolutional layers and stack them to obtain the first three-dimensional feature data block; For any slice of data, channel attention processing is performed on the first three-dimensional feature data block to obtain the second three-dimensional feature data block; For any slice of data, matrix multiplication is used to reduce the dimensionality, so as to obtain the corresponding one-dimensional feature vector that indicates the performance degradation of the bearing. The one-dimensional feature vector of any slice of data is processed through a fully connected layer to obtain the predicted point values. All the predicted point values ​​are then connected in sequence to obtain the prediction curve of the remaining life of the rolling bearing. Each of the aforementioned temporal convolutions uses a convolution kernel of a different size, including: in two adjacent temporal convolutions, the convolution kernel of the current temporal convolution is larger than the convolution kernel of the previous temporal convolution; For any given slice of data, the output of each layer of the three-layer temporal convolutional network is extracted and stacked to obtain a first three-dimensional feature data block, including: Using the maximum width of the three layers of the hierarchically dilated temporal convolution as a reference, the outputs of each of the three temporal convolutional layers are zero-paded in the width dimension and then stacked to obtain the first three-dimensional feature data block; the length of the zero-padding sequence includes: in, Indicates the length of the zero-padding sequence. This represents the width of any one of the three layers. p represents the index of the slice data; This represents the width of the convolutional layer generated during the convolution process of the first temporal convolutional network. This indicates the width of the convolutional layer generated by the second temporal convolution process. This represents the width of the convolutional layer generated by the third temporal convolution process; The process of performing matrix multiplication dimensionality reduction on any given slice of data to obtain a corresponding one-dimensional feature vector indicating bearing performance degradation includes: The three layers of two-dimensional data in the second three-dimensional feature data block are multiplied element-wise by the corresponding column vectors to obtain three column vectors after dimensionality reduction. The three column vectors are then aligned, summed, and added together to obtain the corresponding one-dimensional feature vector indicating bearing performance degradation, including: in, Let l represent the corresponding one-dimensional feature vector indicating bearing performance degradation, and l represent the index of the two-dimensional data layer, where 1 ≤ l ≤ L. This represents the two-dimensional data of the corresponding layer. This represents the column vector of the corresponding layer.

2. The method for predicting the remaining life of rolling bearings as described in claim 1, characterized in that, The normalization preprocessing of the multi-sensor signals includes: in, This represents the data acquired from a sensor x on the bearing at time t. This represents the normalized data. This represents the minimum value obtained from the dataset of a specific sensor x. This represents the maximum value obtained from the dataset of a specific sensor x.

3. The method for predicting the remaining life of rolling bearings as described in claim 1, characterized in that, The process of feeding the sliced ​​data into a temporal convolutional network to achieve three-layer temporal convolution processing includes: in, This represents the output of the first temporal convolution, where d represents the dilation factor. This represents a two-dimensional tensor of sliced ​​data, where T represents the time dimension. Indicates feature dimension; This represents the output of the second temporal convolution layer; This represents the output of the third temporal convolution layer; k T k represents the size of the convolution kernel. T3 >k T2 >k T1 .

4. The method for predicting the remaining life of rolling bearings as described in claim 1, characterized in that, The process of performing a fully connected operation on the one-dimensional feature vector of any slice of data includes: The one-dimensional feature vector of any slice of data is processed by two fully connected layers.

5. A rolling bearing remaining life prediction system, characterized in that, It includes a memory and a processor, wherein the memory stores a program that can be loaded and run by the processor, the program including the rolling bearing remaining life prediction method as described in any one of claims 1 to 4.

6. A computer-readable storage medium, characterized in that, The medium stores a program that can be loaded by a processor and executed as described in any one of claims 1 to 4 for predicting the remaining life of rolling bearings.

Citation Information

Patent Citations

  • Rolling bearing residual life prediction method based on attention enhancement time frequency Transformer

    CN116150901A

  • Method for predicting remaining useful life of railway train bearing based on can-lstm

    US20230153608A1