A method for predicting the remaining life of rolling bearings based on bidirectional temporal convolutional networks and its storage medium.
By performing bidirectional temporal convolution and self-attention processing within a local window, combined with a multilayer perceptron, the problems of insufficient utilization of local information and lack of causal relationships in existing methods are solved, and more accurate prediction of the remaining life of rolling bearings is achieved.
Patent Information
- Application Number
- CN202510037212.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-09
- Publication Date
- 2025-10-31
- Estimated Expiration
- 2045-01-09
AI Technical Summary
Existing methods for predicting the remaining life of rolling bearings are insufficient in mining local information and preserving time-series causal relationships. In particular, temporal convolutional networks (TCNs) rely only on current and past information and fail to make full use of future information.
A method based on bidirectional temporal convolutional networks is adopted, which extracts local information and preserves the causal relationship of the time series by performing forward and backward temporal convolution processing within a local window, combined with self-attention mechanism and multilayer perceptron.
It improves the accuracy and robustness of rolling bearing remaining life prediction by fully mining local information and maintaining overall causal relationships, thereby enhancing prediction performance.
Smart Images
Figure CN119885491B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of fault prediction and health management technology, specifically to a method and storage medium for predicting the remaining life of rolling bearings based on a bidirectional temporal convolutional network. Background Technology
[0002] Rolling bearings are a crucial component of mechanical equipment. Because they operate in extremely challenging environments, they are highly susceptible to factors such as wear and corrosion. Their failure can lead to equipment shutdowns, production interruptions, economic losses, and safety hazards. Therefore, real-time health monitoring and remaining useful life (RUL) prediction during rolling bearing operation are essential. Accurate RUL prediction facilitates preventative maintenance, avoids unexpected failures, and improves equipment reliability and availability.
[0003] Currently, methods for predicting bearing RUL mainly include physics-based methods, statistical model-based methods, and data-driven methods. Physics-based models predict RUL based on the bearing's working principle and wear mechanism using physical equations and models. These methods require a deep understanding of the bearing's physical characteristics and operating environment. Statistical models require historical failure data and life distribution models to estimate the bearing's RUL. These models typically require a large amount of historical data to produce reliable results. Data-driven models use machine learning and deep learning algorithms to predict RUL by analyzing sensor data (such as vibration, temperature, and sound), and have become a hot research area in recent years. Temporal Convolutional Networks (TCNs) are a method based on convolutional neural network architectures. They can compute data across all time steps in parallel and effectively model long-term dependencies with fewer parameters. TCNs have been widely applied in various fields. However, TCNs rely on strict causality, meaning they depend only on current and past information when processing time-series data, without considering future information. Furthermore, local data dependencies are crucial in bearing RUL prediction. Therefore, how to extract important local information while maintaining causality has become a current trend in methodologies. Summary of the Invention
[0004] The technical problem to be solved by this application is to provide a method and storage medium for predicting the remaining life of rolling bearings based on bidirectional temporal convolutional networks. This method has the characteristics of being able to more fully mine local information while maintaining strong correlation of time series and better preserving the causal relationship of the overall time series.
[0005] In a first aspect, one embodiment provides a method for predicting the remaining life of rolling bearings based on a bidirectional temporal convolutional network, comprising:
[0006] The raw signals of the bearing are acquired and preprocessed to obtain time series data. Based on a preset local window, the time series data is divided into multiple window data.
[0007] Perform forward and backward temporal convolution processing on the data in each window within the window;
[0008] Each forward and backward temporal convolution result is subjected to self-attention processing to obtain self-attention weighted features;
[0009] Each self-attention weighted feature is processed by a multilayer perceptron to obtain the corresponding lifetime prediction value.
[0010] In one embodiment, the original signal includes a vibration acceleration signal, a pressure signal, and a torque signal.
[0011] In one embodiment, the process of dividing time series data into multiple window data based on a preset local window includes:
[0012] Based on a preset local window, the step size between windows is smaller than the length of the local window, thus dividing the time series data into multiple window data.
[0013] In one embodiment, the step of performing forward and backward temporal convolution processing on each window's data within the window, specifically performing forward and backward temporal convolution processing on any given window's data within the window, includes:
[0014] Perform forward temporal convolution on the window data within the window to obtain the first feature data;
[0015] Perform backward temporal convolution on the window data within the window to obtain the second feature data;
[0016] The first and second data features are concatenated to obtain the third feature data.
[0017] In one embodiment, the method for performing forward temporal convolution processing on any window of data includes:
[0018] The window data is used as the input of the feedforward temporal convolutional network. Each signal data in the window data is processed by feedforward temporal convolution in turn, and the signal data after feedforward temporal convolution are merged to obtain the first feature data corresponding to the window data.
[0019] For any given window of data, methods for performing backward temporal convolution include:
[0020] The window data is used as the input to the backward temporal convolutional network. After the time series data is reversed, each signal data in the reversed time series data is sequentially processed by forward temporal convolution. The signal data after forward temporal convolution are then merged to obtain the second feature data corresponding to the window data.
[0021] In one embodiment, the forward temporal convolution processing includes multiple layers of residual processing, each residual processing layer including two dilated convolutional layers; for any signal data, the residual processing of any layer includes:
[0022] ,
[0023] ,
[0024] ,
[0025] in, The output of the first dilated convolution layer is represented by l, the index of the residual processing layer is represented by t, the index of the signal data in the window data is represented by 1≤t≤w, and w represents the length of the local window. This represents the output of the second dilated convolution layer. This represents the input for the l-th layer residual processing. This represents the weight matrix for the preset l-th layer residual processing; This represents the output of the l-th layer residual processing; This represents the weight parameters of the first dilated convolution layer. This represents the weight parameters of the second dilated convolution layer. This represents the dilation factor of the first dilated convolution layer. represents the dilation factor of the second layer of dilated convolution; DilatedConv1D represents the one-dimensional dilated convolution processing; WN represents the weight normalization processing; ReLU represents the ReLU activation function; and Dropout represents the regularization processing.
[0026] In one embodiment, the process of concatenating the first feature data and the second feature data to obtain the third feature data includes:
[0027] ,
[0028] in, This represents the third feature data. This represents the first feature data. This represents the second feature data, and n represents the index of the window data.
[0029] In one embodiment, for any forward and backward temporal convolution result, self-attention processing is performed to obtain self-attention weighted features, including:
[0030] ,
[0031] in, This represents the self-attention weighted feature corresponding to any forward and backward temporal convolution result. This indicates self-attention processing. This indicates normalization processing. Indicates the dimension of the key matrix; Represents the query matrix. , The weight matrix representing the query; Represents the key matrix. , The weight matrix representing the key; Represents a value matrix, , The weight matrix represents the values, and T denotes the transpose.
[0032] In one embodiment, for any self-attention weighted feature, multilayer perceptron processing is performed to obtain the corresponding lifetime prediction value, including:
[0033] ,
[0034] ,
[0035] ,
[0036] in, ReLU represents the output of the first fully connected layer in a multilayer perceptron, and ReLU represents the ReLU activation function. This represents the weight matrix of the first fully connected layer. This represents the bias vector of the first fully connected layer; This represents the output of the second fully connected layer in a multilayer perceptron. This represents the weight matrix of the second fully connected layer. This represents the bias vector of the second fully connected layer; Indicates the life expectancy prediction. This represents the weight matrix of the third fully connected layer in a multilayer perceptron. This represents the bias vector of the third fully connected layer.
[0037] In a second aspect, one embodiment 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.
[0038] The beneficial effects of this invention are:
[0039] The raw signals of the bearings are collected and preprocessed to obtain time series data. Based on a preset local window, the time series data is divided into multiple window data. By performing forward and backward temporal convolution processing on each window data within the window, local information can be more fully extracted while maintaining the strong correlation of the time series and better preserving the causal relationship of the overall time series. Attached Figure Description
[0040] Figure 1 This is a schematic flowchart of a rolling bearing remaining life prediction method based on a bidirectional temporal convolutional network according to an embodiment of this application;
[0041] Figure 2 This is a schematic flowchart of a method for performing forward and backward time volume processing on any window of data according to an embodiment of this application. Detailed Implementation
[0042] The present invention will now be described in further detail with reference to specific embodiments and accompanying drawings. 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 this 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 this application are not shown or described in the specification. This is to avoid obscuring the core parts of this 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.
[0043] 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.
[0044] 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).
[0045] To facilitate the explanation of the inventive concept of this application, the following is a brief description of the rolling bearing remaining life prediction technology based on bidirectional temporal convolutional networks.
[0046] Current technologies also include methods for predicting the remaining life of rolling bearings based on bidirectional temporal convolutional networks. However, these methods do not utilize locally defined windows for bidirectional temporal convolution; instead, they often perform forward temporal convolution or simple convolution operations on the data within the window. In contrast, the applicant's research found that performing bidirectional temporal convolution on the data within the window is more effective in fully extracting local information. This approach not only extracts more local information but also maintains strong correlations in the time series, thus better preserving the causal relationships of the overall time series.
[0047] In view of this, this application provides a method and storage medium for predicting the remaining life of rolling bearings based on a bidirectional temporal convolutional network. The method collects the original signal of the bearing and preprocesses it to obtain time series data. Based on a preset local window, the time series data is divided into multiple window data. Forward and backward temporal convolution processing is performed on each window data within the window. This allows for more full mining of local information while maintaining the strong correlation of the time series and better preserving the causal relationship of the overall time series.
[0048] A method for predicting the remaining life of rolling bearings based on bidirectional temporal convolutional networks is available for reference. Figure 1 ,include:
[0049] Step S10: Acquire the raw signal of the bearing and preprocess it to obtain time series data. Based on the preset local window, divide the time series data into multiple window data.
[0050] In one embodiment, it is assumed that the obtained time series data is Where m represents the index of time, 1≤m≤M, and M represents the total length of the time series. The signal data representing time m can include vibration acceleration signal data, pressure signal data, and torque signal data, etc.
[0051] Let the length of each local window be w. Then, the time series X can be divided into multiple window data, with a step size s between windows. If the step size s = w, there is no overlap between windows; if s < w, there will be overlapping areas between windows; if s > w, there will be gaps between windows, and some data may not be covered. Therefore, in one embodiment, based on a preset local window, the step size between windows is made smaller than the length of the local window, i.e., s < w, to divide the time series data into multiple window data. For the nth window, the obtained nth window data can be represented as... Where t represents the index of the signal data in the window data, Let w represent the t-th signal data in the n-th window of data, where 1 ≤ t ≤ w.
[0052] Step S20: Perform forward and backward temporal convolution processing on the data in each window within the window.
[0053] Within each local time window, bidirectional temporal convolution is performed on the window data, including both forward and backward branches. Two independent Temporal Convolution (TCN) branches are used, one for forward convolution and one for backward convolution. Specifically, for any given window of data, the forward and backward temporal convolution processes can include:
[0054] Step S201: Perform forward temporal convolution on the window data within the window to obtain the first feature data.
[0055] In one embodiment, the method for performing forward temporal convolution processing on any window of data includes:
[0056] The window data is used as the input to the feedforward temporal convolutional network. Each signal data in the window data is processed by feedforward temporal convolution in sequence, and the signal data after feedforward temporal convolution are merged to obtain the first feature data corresponding to the window data.
[0057] In one embodiment, the forward temporal convolution processing includes multiple layers of residual processing, each residual processing layer including two dilated convolutional layers.
[0058] Multi-layer residual processing can effectively capture long dependencies in time series. Each dilated convolutional layer can expand the receptive field of the convolutional kernel through the dilation factor, allowing the convolutional operation to cover a wider range of time steps.
[0059] In one embodiment, for any signal data, the residual processing of any layer includes:
[0060] ,
[0061] ,
[0062] ,
[0063] in, This represents the output of the first dilated convolution layer, and l represents the index of the residual processing layer. This represents the output of the second dilated convolution layer. This represents the input for the l-th layer residual processing. This represents the weight matrix for the preset l-th layer residual processing; This represents the output of the l-th residual processing layer and the input of the (l+1)-th residual processing layer. This represents the weight parameters of the first dilated convolution layer. This represents the weight parameters of the second dilated convolution layer. This represents the dilation factor of the first dilated convolution layer. represents the dilation factor of the second layer of dilated convolution; DilatedConv1D represents the one-dimensional dilated convolution processing; WN represents the weight normalization processing; ReLU represents the ReLU activation function; and Dropout represents the regularization processing.
[0064] By merging the signal data after the last layer of residual processing, we obtain the first feature data obtained by forward temporal convolution of the input window data.
[0065] For any window data After forward temporal convolution, the corresponding first feature data can be obtained. .
[0066] Step S202: Perform backward temporal convolution on the window data within the window to obtain the second feature data.
[0067] In one embodiment, the method for performing backward temporal convolution processing on any window of data includes:
[0068] The window data is used as the input to the backward temporal convolutional network. After the time series data is reversed, each signal data in the reversed time series data is sequentially processed by forward temporal convolution. The signal data after forward temporal convolution are then merged to obtain the second feature data corresponding to the window data.
[0069] For any window data First, the time series data is reversed to obtain the reversed window data. Then, the inverted window data After performing forward temporal convolution, the corresponding second feature data can be obtained. Then the second feature data can be represented as:
[0070] ,
[0071] in, Represents backward convolution. .
[0072] Step S203: The first feature data and the second feature data are concatenated to obtain the third feature data.
[0073] In one embodiment, the third feature data can be represented as:
[0074] ,
[0075] in, This represents the third feature data. In this way, information from both the preceding and following periods can be captured simultaneously within each window, while ensuring the causality of the overall time series.
[0076] In one embodiment, the third feature data is used as the result of forward and backward temporal convolution processing.
[0077] Step S30: Perform self-attention processing on the results of each forward and backward temporal convolution to obtain self-attention weighted features.
[0078] In one embodiment, for any forward and backward temporal convolution result, self-attention processing is performed to obtain a self-attention weighted feature, which can be expressed as:
[0079] ,
[0080] in, This represents the self-attention weighted feature corresponding to any forward and backward temporal convolution result. This indicates self-attention processing. This indicates normalization processing. Indicates the dimension of the key matrix; Represents the query matrix. , The weight matrix representing the query; Represents the key matrix. , The weight matrix representing the key; Represents a value matrix, , The weight matrix represents the values, and T denotes the transpose.
[0081] By applying a self-attention mechanism to each local window, complex dependencies and key features in local time series can be captured more effectively.
[0082] Step S40: Process each self-attention weighted feature using a multilayer perceptron to obtain the corresponding lifetime prediction value.
[0083] In one embodiment, for any self-attention weighted feature, multilayer perceptron processing is performed to obtain the corresponding lifetime prediction value, which can be expressed as:
[0084] ,
[0085] ,
[0086] ,
[0087] in, ReLU represents the output of the first fully connected layer in a multilayer perceptron, and ReLU represents the ReLU activation function. This represents the weight matrix of the first fully connected layer. This represents the bias vector of the first fully connected layer; This represents the output of the second fully connected layer in a multilayer perceptron. This represents the weight matrix of the second fully connected layer. This represents the bias vector of the second fully connected layer; Indicates the life expectancy prediction. This represents the weight matrix of the third fully connected layer in a multilayer perceptron. This represents the bias vector of the third fully connected layer.
[0088] Based on the above embodiments, by inputting the self-attention weighted features into the three fully connected layers of the multilayer perceptron, the lifetime prediction value of each window data is obtained.
[0089] In any of the above embodiments, by dividing the local time window and creatively applying a bidirectional temporal convolutional network within the local window, feature extraction is performed using the bidirectional information within the local window. This fully exploits the local information in the bearing sequence while maintaining the overall causal relationship, thereby improving the model's predictive performance. Simultaneously, a self-attention mechanism is introduced to automatically assign attention weights to the data at each time step, enabling it to focus on more important information and further improving the network's performance.
[0090] The rolling bearing remaining life prediction method based on a bidirectional temporal convolutional network in the above embodiments is implemented based on a trained rolling bearing remaining life prediction network. This network is a neural network trained using a constructed basic neural network architecture combined with a constructed loss function. The basic neural network architecture includes a bidirectional temporal convolutional sub-network, a self-attention processing sub-network, and a multilayer perceptron. The bidirectional temporal convolutional sub-network includes forward TCN branches and backward TCN branches, used to perform forward TCN processing and backward TCN processing on the input window data, respectively. The bidirectional temporal convolutional sub-network also includes a concatenation layer for concatenating the results of the forward and backward TCN processing. The self-attention processing sub-network performs self-attention processing on each forward and backward temporal convolution result to obtain self-attention weighted features. The multilayer perceptron processes each self-attention weighted feature to obtain the predicted life value. The loss function calculates the distance between the predicted life value and the actual life value and provides feedback until the network training meets the requirements.
[0091] To verify the effectiveness of the rolling bearing remaining life prediction method based on bidirectional temporal convolutional networks in the above embodiments, one embodiment uses an aerospace bearing dataset for verification experiments. The hardware structure of the civil aircraft bearing component test bench consists of eight parts: a drive motor, gearbox, rolling bearing, connecting shaft, disc flywheel, pulley, coupling, and adjustable load. Sensors used to collect characteristics of the bearing from normal operation to failure include vibration acceleration sensors, torque sensors, temperature sensors, pressure sensors, and air temperature and humidity sensors. In this embodiment, vibration acceleration sensors, torque sensors, and pressure sensors are used to acquire data signals. The sampling interval of the test bench is 10 seconds, each acquisition cycle is 0.1 seconds, and 2500 data points are generated for each bearing sample.
[0092] To verify the accuracy and robustness of the proposed prediction method, two evaluation metrics, MAE and MSE, were introduced. MAE reflects the true error. MSE is more sensitive to larger errors in the sample and can detect both large and less common errors. Overall, these two metrics can well reflect the prediction performance. For completeness, the network structure of this application was compared with classical network structures for predicting bearing RUL. The selected network structures included FCN, LSTM, Bi-LSTM, and TCN. The performance results for bearing 1 and bearing 2 are shown in Table 1.
[0093] Table 1
[0094] ,
[0095] Experimental results show that, for any evaluation metric, the proposed method outperforms other existing models. Analyzing the average evaluation metrics for the two bearings, compared to existing models such as the TCN, FCN, LSTM, and Bi-LSTM models, the proposed method reduces the MAE by 5.1%, 21.5%, 16.7%, and 13.8% for bearing 1, and reduces the MSE by 1.4%, 16.7%, 7.2%, and 4.3%, respectively.
[0096] The above description is merely a specific implementation method of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of protection of the claims.
[0097] One embodiment of this application provides a computer-readable storage medium storing a program, the stored program including methods that can be loaded by a processor and processed in any of the above embodiments.
[0098] 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.
[0099] 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 rolling bearings based on a bidirectional temporal convolutional network, characterized in that, include: The raw signals of the bearing are acquired and preprocessed to obtain time series data. Based on a preset local window, the time series data is divided into multiple window data. Within each window, perform forward and backward temporal convolution processing on the data within that window. Specifically, for any given window of data, the forward and backward temporal convolution processing within the window includes: The window data is used as the input of the feedforward temporal convolutional network. Each signal data in the window data is processed by feedforward temporal convolution in turn, and the signal data after feedforward temporal convolution are merged to obtain the first feature data corresponding to the window data. The window data is used as the input of the backward temporal convolutional network. After the time series data is reversed, each signal data in the reversed time series data is sequentially processed by forward temporal convolution. Then, each signal data after forward temporal convolution is merged to obtain the second feature data corresponding to the window data. The first and second feature data are concatenated to obtain the third feature data, including... in, This represents the third feature data. This represents the first feature data. This represents the second feature data, where n represents the index of the window data; The forward temporal convolution processing includes multiple layers of residual processing, each of which consists of two dilated convolutional layers. For any given signal data, the residual processing at any layer includes: in, The output of the first dilated convolution layer is represented by l, the index of the residual processing layer is represented by t, the index of the signal data in the window data is represented by 1≤t≤w, and w represents the length of the local window. This represents the output of the second dilated convolution layer. This represents the input for the l-th layer residual processing. This represents the weight matrix for the preset l-th layer residual processing; This represents the output of the l-th layer residual processing; This represents the weight parameters of the first dilated convolution layer. This represents the weight parameters of the second dilated convolution layer. This represents the dilation factor of the first dilated convolution layer. represents the dilation factor of the second layer of dilated convolution; DilatedConv1D represents the one-dimensional dilated convolution processing; WN represents the weight normalization processing; ReLU represents the ReLU activation function; Dropout represents the regularization processing. Each forward and backward temporal convolution result is subjected to self-attention processing to obtain self-attention weighted features. Specifically, for any forward and backward temporal convolution result, self-attention processing is performed to obtain self-attention weighted features, including: in, This represents the self-attention weighted feature corresponding to any forward and backward temporal convolution result. This indicates self-attention processing. This indicates normalization processing. Indicates the dimension of the key matrix; Represents the query matrix. , The weight matrix representing the query; Represents the key matrix, , The weight matrix representing the key; Represents a value matrix, , The weight matrix represents the values, and T denotes the transpose; Each self-attention-weighted feature is processed by a multilayer perceptron to obtain the corresponding lifetime prediction value. Specifically, for any self-attention-weighted feature, the multilayer perceptron processing yields the following lifetime prediction value: in, ReLU represents the output of the first fully connected layer in a multilayer perceptron, and ReLU represents the ReLU activation function. This represents the weight matrix of the first fully connected layer. This represents the bias vector of the first fully connected layer; This represents the output of the second fully connected layer in a multilayer perceptron. This represents the weight matrix of the second fully connected layer. This represents the bias vector of the second fully connected layer; Indicates the life expectancy prediction. This represents the weight matrix of the third fully connected layer in a multilayer perceptron. This represents the bias vector of the third fully connected layer.
2. The method for predicting the remaining life of rolling bearings as described in claim 1, characterized in that, The raw signals include vibration acceleration signals, pressure signals, and torque signals.
3. The method for predicting the remaining life of rolling bearings as described in claim 1, characterized in that, The method described above, based on a preset local window, divides the time series data into multiple window data, including: Based on a preset local window, the step size between windows is smaller than the length of the local window, thus dividing the time series data into multiple window data.
4. A computer-readable storage medium, characterized in that, The medium stores a program that can be loaded by a processor and executed as the rolling bearing remaining life prediction method as described in any one of claims 1 to 3.