A migration prediction method fusing physical priori and double-layer attention mechanism
By integrating physical priors and a two-layer attention mechanism, the migration prediction method solves the problem of data migration across operating conditions and equipment, and realizes efficient adaptation and accurate prediction of the life prediction model of rotating equipment under different conditions, thereby improving the interpretability and robustness of the model.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NORTHEASTERN UNIV CHINA
- Filing Date
- 2026-01-30
- Publication Date
- 2026-06-19
AI Technical Summary
Traditional life prediction models cannot effectively migrate data across operating conditions and equipment, resulting in insufficient adaptability and generalization ability under different conditions.
A transfer prediction method integrating physical priors and a two-layer attention mechanism is adopted. By acquiring the time series of the source and target domains of the rotating device, the data is augmented using a pre-set harmonic enhancement network. Feature extraction and weight allocation are performed by combining a feature extraction network and a two-stage attention mechanism network. Finally, a pre-set lifetime prediction model is constructed through iterative training using a lifetime prediction network.
The model improves the adaptability and generalization ability of the rotating equipment life prediction model under different conditions, enhances the prediction accuracy of the model under small sample conditions, and improves the interpretability and robustness of the model.
Smart Images
Figure CN122242573A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of predictive maintenance technology for rotating machinery, and in particular to a migration prediction method that integrates physical priors and a two-layer attention mechanism. Background Technology
[0002] In modern industry, rotating equipment (such as electric motors, pumps, and fans) is a core component, and its operating status directly affects the reliability and service life of the equipment. Effectively predicting equipment lifespan has become a key technology in equipment maintenance and predictive management.
[0003] Currently, in practical industrial applications, due to differences in equipment models, operating environments, and monitoring systems, obtaining a large number of high-quality labeled samples is costly and difficult. Therefore, cross-condition or cross-equipment transfer training methods are usually adopted. However, there are significant differences in data distribution between different conditions or equipment, and traditional life prediction models cannot effectively perform data transfer across conditions and equipment, resulting in insufficient adaptability and generalization ability of life prediction models under different conditions. Summary of the Invention
[0004] In view of this, this application provides a migration prediction method that integrates physical priors and a two-layer attention mechanism. The main purpose is to solve the data migration problem across operating conditions and equipment, thereby improving the adaptability and generalization ability of the life prediction model of rotating equipment under different conditions.
[0005] According to a first aspect of this application, a transfer prediction method integrating physical priors and a two-layer attention mechanism is provided, the method comprising: The source domain time series and target domain time series of the rotating equipment are obtained, as well as an initial lifetime prediction model. The source domain time series and the target domain time series come from different rotating equipment or from the same rotating equipment under different operating conditions. The initial lifetime prediction model includes a feature extraction network, a two-stage attention mechanism network, and a lifetime prediction network. The source domain time series is input into a preset harmonic enhancement network for data enhancement to obtain an enhanced source domain time series, wherein the preset harmonic enhancement network is embedded with physical constraint equations. The enhanced source domain time series and the target domain time series are input into the feature extraction network for feature extraction to obtain the source domain device state feature vector and the target domain device state feature vector. The source domain device state feature vector and the target domain device state feature vector are respectively input into the two-stage attention mechanism network for adaptive weight allocation to obtain the source domain context vector and the target domain context vector. The weight allocation is based on the contribution of each feature in the time-frequency domain feature corresponding to the enhanced source domain time series at each time step in different stages. The source domain context vector is input into the lifetime prediction network to predict the lifetime, thereby obtaining the predicted lifetime of the rotating equipment. Based on the source domain context vector, the target domain context vector, and the predicted lifetime, the initial lifetime prediction model is iteratively trained to obtain a preset lifetime prediction model.
[0006] According to a second aspect of this application, a migration prediction device integrating physical prior knowledge and a two-layer attention mechanism is provided, the device comprising: The acquisition unit is used to acquire the source domain time series and the target domain time series of the rotating equipment, as well as the initial lifetime prediction model. The source domain time series and the target domain time series come from different rotating equipment or from the same rotating equipment under different operating conditions. The initial lifetime prediction model includes a feature extraction network, a two-stage attention mechanism network, and a lifetime prediction network. An enhancement unit is used to input the source domain time series into a preset harmonic enhancement network for data enhancement to obtain an enhanced source domain time series, wherein the preset harmonic enhancement network is embedded with physical constraint equations; The feature extraction unit is used to input the time-frequency domain features corresponding to the enhanced source domain time series and the time-frequency domain features corresponding to the target domain time series into the feature extraction network to extract features, thereby obtaining the source domain device state feature vector and the target domain device state feature vector. The weight allocation unit is used to input the source domain device state feature vector and the target domain device state feature vector into the two-stage attention mechanism network for adaptive weight allocation to obtain the source domain context vector and the target domain context vector. The weight allocation is based on the contribution of each feature in the time-frequency domain feature corresponding to the enhanced source domain time series at each time step in different stages. The prediction unit is used to input the source domain context vector into the lifetime prediction network to predict the lifetime and obtain the predicted lifetime of the rotating equipment. The training unit is used to iteratively train the initial lifetime prediction model based on the source domain context vector, the target domain context vector, and the predicted lifetime, to obtain a preset lifetime prediction model.
[0007] According to a third aspect of this application, a storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the above-described migration prediction method that integrates physical prior knowledge and a two-layer attention mechanism.
[0008] According to a fourth aspect of this application, an electronic device is provided, including a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, wherein the processor, when executing the program, implements the above-described migration prediction method that integrates physical prior knowledge and a two-layer attention mechanism.
[0009] By employing the aforementioned technical solutions, this application provides a transfer prediction method that integrates physical priors and a two-layer attention mechanism. By using a two-stage attention mechanism, weights are adaptively assigned to each feature at different stages. This design significantly improves the predictive model's adaptability under small sample conditions and enhances its transferability across equipment and operating conditions. It ensures accurate prediction even with small samples and significant differences in distribution between the source domain, thereby improving the adaptability and generalization ability of the life prediction model for rotating equipment under different conditions. Simultaneously, the two-stage attention mechanism enhances the model's interpretability by explaining the contribution of each feature in the source domain at different stages. Furthermore, by introducing physical constraint equations into the harmonic enhancement network, this application ensures that the extracted features not only conform to data patterns but also follow the physical characteristics of the equipment, thus improving the interpretability of the life prediction model while suppressing noise and enhancing the robustness and physical consistency of the prediction model. This multi-layered interpretable mechanism makes the model's decision-making process transparent, enhancing its credibility and usability in engineering applications, and providing an efficient and reliable tool for intelligent maintenance and reliability management of industrial equipment.
[0010] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description
[0011] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 The diagram shows a flowchart of a migration prediction method that integrates physical prior knowledge and a two-layer attention mechanism, as provided in an embodiment of this application. Figure 2 This paper illustrates the overall process of migration prediction provided in an embodiment of this application. Figure 3A flowchart illustrating the preset harmonic enhancement network training method provided in an embodiment of this application is shown. Figure 4 The diagram shows a schematic of a migration prediction device that integrates physical prior knowledge and a two-layer attention mechanism, as provided in an embodiment of this application. Detailed Implementation
[0012] The present application will be described in detail below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in the embodiments of the present application can be combined with each other.
[0013] There are significant differences in data distribution between different operating conditions or equipment. Traditional life prediction models cannot effectively migrate data across operating conditions and equipment, resulting in insufficient adaptability and generalization ability of life prediction models under different conditions.
[0014] To address the aforementioned problems, embodiments of the present invention provide a transfer prediction method that integrates physical prior knowledge and a two-layer attention mechanism, such as... Figure 1 As shown, the method includes: Step 101: Obtain the source domain time series and target domain time series of the rotating equipment, as well as the initial lifetime prediction model.
[0015] The rotating equipment includes rotating machinery such as electric motors, pumps, and fans. The source domain time series and the target domain time series come from different rotating equipment or from the same rotating equipment under different operating conditions. The initial lifetime prediction model includes a feature extraction network, a two-stage attention mechanism network, and a lifetime prediction network.
[0016] In this embodiment of the invention, the source domain time series comes from the training dataset, and the target domain time series comes from the validation dataset. The training dataset of the source domain has sufficient labeled samples, meaning that a large amount of data has been accumulated in similar rotating equipment, operating conditions, or monitoring environments. The validation dataset and test set of the target domain come from new rotating equipment and operating conditions, and the available labeled samples are very limited. The source domain time series includes horizontal vibration signal time series and vertical vibration signal time series, and the target domain time series also includes horizontal vibration signal time series and vertical vibration signal time series. The horizontal vibration signals and vertical vibration signals can be collected separately using sensors when the rotating equipment is operating.
[0017] Since the source domain time series and the target domain time series originate from different rotating devices or from the same rotating device under different operating conditions, there are significant differences in their data distributions. To effectively perform data migration across operating conditions and devices, this invention proposes an initial lifetime prediction model. By iteratively training this initial lifetime prediction model, a preset lifetime prediction model is constructed, enabling effective data migration across operating conditions and devices. This initial lifetime prediction model includes a feature extraction network, a two-stage attention mechanism network, and a lifetime prediction network. Specifically, the feature extraction network can be an LSTM feature extraction network. The two-stage attention mechanism is used to adaptively allocate weights in the time and channel dimensions to focus on the most critical degradation features. The lifetime prediction network can be a one- to three-layer fully connected network.
[0018] This invention is primarily applicable to scenarios where a lifespan prediction model for rotating equipment is trained based on a small sample size. The executing entity of this invention is a device or equipment capable of training a lifespan prediction model for rotating equipment based on a small sample size, specifically, it can be located on a server side.
[0019] Step 102: Input the source domain time series into a preset harmonic enhancement network for data enhancement to obtain the enhanced source domain time series.
[0020] The preset harmonic enhancement network incorporates physical constraint equations, including dynamic equations and harmonic characteristic equations. The source domain time series includes horizontal vibration signal time series and vertical vibration signal time series.
[0021] In this embodiment of the invention, the horizontal vibration signal time series and the vertical vibration signal time series are respectively input into a preset harmonic enhancement network for data enhancement, so as to obtain the enhanced horizontal vibration signal time series and the enhanced vertical vibration signal time series.
[0022] Specifically, in this embodiment of the invention, the dynamic equations and harmonic characteristic equations are embedded into a physical information neural network (PINNS network) to obtain a preset harmonic enhancement network (HPINNs network). This ensures that the subsequently extracted features not only conform to data patterns but also adhere to the physical characteristics of the equipment, thereby improving the interpretability and stability of the lifetime prediction model. The HPINNs network of this embodiment can effectively handle periodic harmonics and non-stationary signals of rotating equipment, suppress noise, and improve the accuracy of lifetime prediction. The specific structure and principle of the preset harmonic enhancement network (HPINNs network) are detailed in the model training process of Embodiment 2.
[0023] Step 103: Input the enhanced source domain time series corresponding to the time-frequency domain features and the target domain time series corresponding to the time-frequency domain features into the feature extraction network to extract features, and obtain the source domain device state feature vector and the target domain device state feature vector.
[0024] Specifically, the feature extraction network can be an LSTM feature extraction network. The target domain time series also includes horizontal vibration signals and vertical vibration signals.
[0025] In this embodiment of the invention, the time-frequency domain features corresponding to the time series of the enhanced horizontal vibration signal from the source domain, the time-frequency domain features corresponding to the time series of the enhanced vertical vibration signal from the source domain, and the time-frequency domain features corresponding to the time series of the target domain are respectively input into the feature extraction network for feature extraction, so as to obtain the horizontal feature vector of the equipment state, the vertical feature vector of the equipment state, and the feature vector of the target domain equipment state.
[0026] Specifically, the time-frequency domain features corresponding to the time series of the horizontal vibration signal of the target domain, and the time-frequency domain features corresponding to the time series of the vertical vibration signal of the target domain, are also input into the LSTM feature extraction network for feature extraction.
[0027] Step 104: Input the source domain device state feature vector and the target domain device state feature vector into the two-stage attention mechanism network for adaptive weight allocation to obtain the source domain context vector and the target domain context vector.
[0028] The weights are allocated based on the contribution of each feature in the time-frequency domain features corresponding to the enhanced source domain time series at each time step in different stages.
[0029] In this embodiment of the invention, the horizontal feature vector and the vertical feature vector of the device state are respectively input into the two-stage attention mechanism network for adaptive weight allocation to obtain the contribution weights of each feature in the time-frequency domain features corresponding to the enhanced horizontal vibration signal time series at each time step under different stages, and the contribution weights of each feature in the time-frequency domain features corresponding to the enhanced vertical vibration signal time series at each time step under different stages. Based on the contribution weights of each feature in the time-frequency domain features corresponding to the enhanced horizontal vibration signal time series at each time step under different stages, and the contribution weights of each feature in the time-frequency domain features corresponding to the enhanced vertical vibration signal time series at each time step under different stages, an attention heatmap, a horizontal source domain context vector, and a vertical source domain context vector are output. The target domain device state feature vector is respectively input into the two-stage attention mechanism network for adaptive weight allocation to obtain the target domain context vector.
[0030] The two-stage attention mechanism of this invention adaptively allocates weights in the time and channel dimensions to focus on the most critical degradation features and integrates the subsequent maximum mean difference (MMD) loss to reduce the inter-domain distribution differences during the migration from the source domain to the target domain. Specifically, in the first stage, the time-step attention mechanism focuses on key time steps in the degradation process and weights the features in each time step. In the second stage, in the channel dimension, weights are assigned to each sensor signal channel (such as horizontal vibration signal and vertical vibration signal) to explain the contribution of the source domain features represented by different channels, thereby further improving the prediction accuracy and interpretability of the model. The attention heatmap records the contribution weights of each feature in the time-frequency domain features corresponding to the time series of the enhanced horizontal vibration signal in the source domain at different stages and at different time steps, as well as the contribution weights of each feature in the time-frequency domain features corresponding to the time series of the enhanced vertical vibration signal in the source domain at different stages and at different time steps.
[0031] It should be noted that the horizontal and vertical feature vectors of the device state in the target domain are also input into the two-stage attention mechanism network for adaptive weight allocation, resulting in the horizontal and vertical target domain context vectors.
[0032] Step 105: Input the source domain context vector into the lifetime prediction network to predict the lifetime and obtain the predicted lifetime of the rotating equipment.
[0033] Specifically, the lifetime prediction network can be a fully connected layer network with one to three layers.
[0034] In this embodiment of the invention, the horizontal source domain context vector and the vertical source domain context vector are input into the lifetime prediction network to predict the lifetime and obtain the predicted lifetime of the rotating device.
[0035] Step 106: Based on the source domain context vector, the target domain context vector, and the predicted lifetime, iteratively train the initial lifetime prediction model to obtain a preset lifetime prediction model.
[0036] In this embodiment of the invention, a total loss function is constructed based on the horizontal source domain context vector, the vertical source domain context vector, the target domain context vector, and the predicted lifetime; based on the total loss function, the initial lifetime prediction model is iteratively trained to obtain a preset lifetime prediction model. The overall training process of the prediction model in this embodiment of the invention is as follows: Figure 2 As shown.
[0037] When constructing the total loss function, a first maximum mean difference loss is calculated based on the horizontal source domain context vector and the target domain context vector; a second maximum mean difference loss is calculated based on the vertical source domain context vector and the target domain context vector; a lifetime loss is calculated based on the predicted lifetime and the actual predicted lifetime of the rotating equipment; and the total loss function is constructed based on the first maximum mean difference loss, the second maximum mean difference loss, and the lifetime loss.
[0038] Specifically, based on the horizontal source domain context vector and the horizontal target domain context vector, a first maximum mean difference loss is calculated. Similarly, based on the vertical source domain context vector and the vertical target domain context vector, a second maximum mean difference loss is calculated. Then, the first maximum mean difference loss, the second maximum mean difference loss, and the lifetime loss are summed to obtain the total loss. Based on this total loss, the initial lifetime prediction model is iteratively trained until a preset number of iterations is reached or the total loss is less than a preset threshold. At this point, iteration stops, and the preset lifetime prediction model is output. The specific formula for the total loss function is as follows:
[0039] in, Represents the total loss; Represents life expectancy loss; This represents the maximum mean difference loss, which is the sum of the first and second maximum mean difference losses. This represents the weighting coefficient.
[0040] In some embodiments, to provide guidance for operators' maintenance work, this invention identifies key features in the time-frequency domain that significantly impact the lifespan of rotating equipment. Based on this, the method further includes: determining a first key feature affecting lifespan prediction based on the attention heatmap; analyzing the contribution of each feature in the time-frequency domain features corresponding to the enhanced horizontal vibration signal time series and the enhanced vertical vibration signal time series using the Shapley additive interpretation method; determining a second key feature affecting lifespan prediction based on the contribution of each feature in the time-frequency domain features corresponding to the enhanced horizontal vibration signal time series and the enhanced vertical vibration signal time series; and outputting a target key feature by taking the intersection of the first and second key features.
[0041] Specifically, the attention heatmap records the contribution weights of each feature in the time-frequency domain of the enhanced horizontal vibration signal time series at different stages and time steps, as well as the contribution weights of each feature in the time-frequency domain of the enhanced vertical vibration signal time series at different stages. The attention heatmap allows identification of the first key feature with a significant impact on lifespan. Simultaneously, the Shapley additive interpretation method (SHAP analysis) is used to calculate the contribution of each input feature to the prediction model output, quantifying the positive and negative impacts of features on the prediction results. This identifies the second key feature with a significant impact on lifespan. Finally, the intersection of the first and second key features pinpoints the target key feature that truly influences lifespan prediction.
[0042] In some embodiments, the migration effect of the prediction model across devices and operating conditions can also be verified. Based on this, the method further includes: performing dimensionality reduction on the time-frequency domain features corresponding to the enhanced source domain time series to obtain dimensionality-reduced source domain time-frequency domain features; performing dimensionality reduction on the time-frequency domain features corresponding to the enhanced target domain time series to obtain dimensionality-reduced target domain time-frequency domain features; analyzing the directional consistency between the dimensionality-reduced source domain time-frequency domain features and the dimensionality-reduced target domain time-frequency domain features; if the directional consistency between the dimensionality-reduced source domain time-frequency domain features and the dimensionality-reduced target domain time-frequency domain features is confirmed, then the data migration effect across operating conditions and devices is determined to be good.
[0043] Specifically, the source domain time-frequency domain features and target domain time-frequency domain features under different time windows are obtained using a lifetime prediction model. Then, methods such as PCA or t-SNE are used to reduce the dimensionality of the source domain time-frequency domain features and target domain time-frequency domain features to obtain two-dimensional source domain time-frequency domain features and two-dimensional target domain time-frequency domain features. Next, it is checked whether the directions of the two-dimensional source domain time-frequency domain features and two-dimensional target domain time-frequency domain features under different time windows are consistent. If they are consistent, it is determined that the data migration effect of the fixed-term operating conditions and equipment is good.
[0044] The embodiments of the present invention provide comprehensive transparency from feature selection to decision-making process through multi-level interpretation techniques such as SHAP feature importance analysis and attention heatmaps, providing a reliable decision-making basis for industrial applications.
[0045] This invention provides a transfer prediction method that integrates physical priors and a two-stage attention mechanism. By employing a two-stage attention mechanism, weights are adaptively assigned to each feature at different stages. This design significantly improves the adaptability of the prediction model under small sample conditions and enhances its transferability across equipment and operating conditions. It ensures accurate prediction even with small samples and significant differences in distribution between the source domain, thereby improving the adaptability and generalization ability of the life prediction model for rotating equipment under different conditions. Simultaneously, the two-stage attention mechanism enhances the interpretability of the model by explaining the contribution of each feature in the source domain at different stages. Furthermore, by introducing physical constraint equations into the harmonic enhancement network, this invention ensures that the extracted features not only conform to data patterns but also follow the physical characteristics of the equipment, thus improving the interpretability of the life prediction model. It also suppresses noise and improves the robustness and physical consistency of the prediction model. This multi-level interpretable mechanism makes the model decision-making process transparent, enhancing its credibility and usability in engineering applications, and providing an efficient and reliable tool for intelligent maintenance and reliability management of industrial equipment.
[0046] Furthermore, as a refinement and extension of the specific implementation of the above embodiments, and to fully illustrate the implementation of this embodiment, this embodiment also provides a training method for a preset harmonic enhancement network, such as... Figure 3 As shown, the method includes: Step 201: Embed the harmonic characteristic equation into the physical information neural network to obtain the initial harmonic enhancement network.
[0047] In the embodiments of the present invention, the harmonic characteristic equation is first... Embedded into a physical information neural network (PINNS network), the embedded PINNS network (initial harmonic enhancement network) is obtained. The specific formula for the harmonic characteristic equation is as follows:
[0048] in, t Represents time, τ Represents the initial (or relative) phase delay, which is a parameter to be trained.
[0049] Furthermore, the embedded PINNS network (initial harmonic enhancement network) can be specifically represented as:
[0050] Step 202: Based on the horizontal and vertical dynamic equations of the rotating device, construct the horizontal residual loss and the vertical residual loss, and based on the horizontal and vertical residual losses, construct the residual loss function based on the dynamic equations.
[0051] For embodiments of the present invention, the horizontal and vertical dynamic equations of the rotating device are as follows:
[0052] The above horizontal and vertical dynamic equations are deformed to introduce residuals, and the specific formulas are as follows.
[0053] in, and These are the residuals introduced in the horizontal and vertical directions, respectively.
[0054] Furthermore, after deforming the dynamic equations, a residual loss function based on the dynamic equations is constructed, with the specific formula as follows:
[0055] in, and These are the horizontal residual loss and the vertical residual loss, respectively. This is the residual loss function.
[0056] Step 203: Based on the residual loss function, iteratively train the initial harmonic enhancement network to obtain the preset harmonic enhancement network.
[0057] By introducing physical constraint equations into the harmonic enhancement network, this invention can ensure that the extracted features not only conform to data patterns but also follow the physical characteristics of the equipment. This improves the interpretability of the lifetime prediction model, while also suppressing noise and enhancing the robustness and physical consistency of the prediction model.
[0058] Furthermore, as Figure 1 and Figure 3 The specific implementation of the method shown in this embodiment provides a migration prediction device that integrates physical prior knowledge and a two-layer attention mechanism, such as... Figure 4 As shown, the device includes: an acquisition unit 31, an enhancement unit 32, a feature extraction unit 33, a weight allocation unit 34, a prediction unit 35, and a training unit 36.
[0059] The acquisition unit 31 can be used to acquire the source domain time series and the target domain time series of the rotating equipment, as well as the initial lifetime prediction model. The source domain time series and the target domain time series come from different rotating equipment or from the same rotating equipment under different operating conditions. The initial lifetime prediction model includes a feature extraction network, a two-stage attention mechanism network, and a lifetime prediction network.
[0060] The enhancement unit 32 can be used to input the source domain time series into a preset harmonic enhancement network for data enhancement to obtain an enhanced source domain time series, wherein the preset harmonic enhancement network is embedded with physical constraint equations.
[0061] The feature extraction unit 33 can be used to input the time-frequency domain features corresponding to the enhanced source domain time series and the time-frequency domain features corresponding to the target domain time series into the feature extraction network for feature extraction, so as to obtain the source domain device state feature vector and the target domain device state feature vector.
[0062] The weight allocation unit 34 can be used to input the source domain device state feature vector and the target domain device state feature vector into the two-stage attention mechanism network for adaptive weight allocation to obtain the source domain context vector and the target domain context vector. The weight allocation is based on the contribution of each feature in the time-frequency domain feature corresponding to the enhanced source domain time series at each time step in different stages.
[0063] The prediction unit 35 can be used to input the source domain context vector into the lifetime prediction network to predict the lifetime and obtain the predicted lifetime of the rotating equipment.
[0064] The training unit 36 can be used to iteratively train the initial lifetime prediction model based on the source domain context vector, the target domain context vector, and the predicted lifetime, to obtain a preset lifetime prediction model.
[0065] In some embodiments, the training unit 36 can also be used to embed the harmonic feature equation into the physical information neural network to obtain an initial harmonic enhancement network; construct a horizontal residual loss and a vertical residual loss based on the horizontal and vertical dynamic equations of the rotating device; construct a residual loss function based on the dynamic equations based on the horizontal and vertical residual losses; and iteratively train the initial harmonic enhancement network based on the residual loss function to obtain the preset harmonic enhancement network.
[0066] In some embodiments, the source domain time series includes a horizontal vibration signal time series and a vertical vibration signal time series. The enhancement unit 32 can be specifically used to input the horizontal vibration signal time series and the vertical vibration signal time series into a preset harmonic enhancement network for data enhancement, so as to obtain the enhanced horizontal vibration signal time series and the enhanced vertical vibration signal time series.
[0067] In some embodiments, the feature extraction unit 33 may be specifically used to input the time-frequency domain features corresponding to the enhanced horizontal vibration signal time series, the time-frequency domain features corresponding to the enhanced vertical vibration signal time series, and the time-frequency domain features corresponding to the target domain time series into the feature extraction network for feature extraction, so as to obtain the device state horizontal feature vector, the device state vertical feature vector, and the target domain device state feature vector.
[0068] In some embodiments, the weight allocation unit 34 may be specifically used to input the device state horizontal feature vector and the device state vertical feature vector into the two-stage attention mechanism network for adaptive weight allocation, thereby obtaining the contribution weights of each feature in the time-frequency domain features corresponding to the enhanced horizontal vibration signal time series at each time step under different stages, and the contribution weights of each feature in the time-frequency domain features corresponding to the enhanced vertical vibration signal time series at each time step under different stages; based on the contribution weights of each feature in the time-frequency domain features corresponding to the enhanced horizontal vibration signal time series at each time step under different stages, and the contribution weights of each feature in the time-frequency domain features corresponding to the enhanced vertical vibration signal time series at each time step under different stages, outputting an attention heatmap, a horizontal source domain context vector, and a vertical source domain context vector; and inputting the target domain device state feature vector into the two-stage attention mechanism network for adaptive weight allocation to obtain the target domain context vector.
[0069] In some embodiments, the prediction unit 35 may be specifically used to input the horizontal source domain context vector and the vertical source domain context vector into the lifetime prediction network to perform lifetime prediction and obtain the predicted lifetime of the rotating device.
[0070] In some embodiments, the training unit 36 includes a construction module and a training module.
[0071] The construction module can be used to construct a total loss function based on the horizontal source domain context vector, the vertical source domain context vector, the target domain context vector, and the predicted lifetime.
[0072] The training module can be used to iteratively train the initial lifetime prediction model based on the total loss function to obtain a preset lifetime prediction model.
[0073] In some embodiments, the construction module may be specifically configured to calculate a first maximum mean difference loss based on the horizontal source domain context vector and the target domain context vector; calculate a second maximum mean difference loss based on the vertical source domain context vector and the target domain context vector; calculate a lifetime loss based on the predicted lifetime and the actual predicted lifetime of the rotating device; and construct the total loss function based on the first maximum mean difference loss, the second maximum mean difference loss, and the lifetime loss.
[0074] In some embodiments, the apparatus further includes a determining unit.
[0075] The determining unit can be used to determine a first key feature affecting lifespan prediction based on the attention heatmap; analyze the contribution of each feature in the time-frequency domain features corresponding to the enhanced horizontal vibration signal time series and the enhanced vertical vibration signal time series using the Shapley additive interpretation method; determine a second key feature affecting lifespan prediction based on the contribution of each feature in the time-frequency domain features corresponding to the enhanced horizontal vibration signal time series and the enhanced vertical vibration signal time series; and output the target key feature by taking the intersection of the first key feature and the second key feature.
[0076] In some embodiments, the determining unit may further be used to perform dimensionality reduction processing on the time-frequency domain features corresponding to the enhanced source domain time series to obtain dimensionality-reduced source domain time-frequency domain features; perform dimensionality reduction processing on the time-frequency domain features corresponding to the target domain time series to obtain dimensionality-reduced target domain time-frequency domain features; analyze the directional consistency between the dimensionality-reduced source domain time-frequency domain features and the dimensionality-reduced target domain time-frequency domain features; if the directional consistency between the dimensionality-reduced source domain time-frequency domain features and the dimensionality-reduced target domain time-frequency domain features is confirmed, then the data migration effect across operating conditions and devices is determined to be good.
[0077] It should be noted that other corresponding descriptions of the functional units involved in the migration prediction device integrating physical prior and two-layer attention mechanism provided in this embodiment of the invention can be found in the following references. Figure 1 and Figure 3 The corresponding description in [the document] will not be repeated here.
[0078] Based on the above, Figure 1 and Figure 3 Accordingly, this embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the above-described method. Figure 1 and Figure 3 The method shown is a transfer prediction method that integrates physical priors and a two-layer attention mechanism.
[0079] Based on this understanding, the technical solution of this application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as CD-ROM, USB flash drive, mobile hard drive, etc.) and includes several instructions to cause an electronic device (such as personal computer, server, or network device, etc.) to execute the methods of various implementation scenarios of this application.
[0080] Based on the above, Figure 1 and Figure 3 The method shown, and Figure 4 To achieve the above objectives, the present application also provides an electronic device, specifically a personal computer, tablet computer, server, or other network device, as shown in the virtual device embodiment. This device includes a storage medium and a processor; the storage medium stores a computer program; the processor executes the computer program to achieve the above-described objectives. Figure 1 and Figure 3 The method shown is a transfer prediction method that integrates physical priors and a two-layer attention mechanism.
[0081] Optionally, the aforementioned physical devices may also include a user interface, a network interface, a camera, radio frequency (RF) circuitry, sensors, audio circuitry, a Wi-Fi module, etc. The user interface may include a display screen, input units such as a keyboard, etc., and optional user interfaces may also include USB interfaces, card reader interfaces, etc. The network interface may optionally include standard wired interfaces, wireless interfaces (such as Wi-Fi interfaces), etc.
[0082] Those skilled in the art will understand that the physical device structure provided in this embodiment does not constitute a limitation on the physical device, and may include more or fewer components, or combine certain components, or have different component arrangements.
[0083] The storage medium may also include an operating system and a network communication module. The operating system is a program that manages the hardware and software resources of the aforementioned physical device, supporting the operation of information processing programs and other software and / or programs. The network communication module is used to enable communication between the various components within the storage medium, as well as communication with other hardware and software in the information processing physical device.
[0084] Through the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware platform, or it can be implemented by hardware.
[0085] This invention employs a two-stage attention mechanism to adaptively assign weights to various features at different stages. This design significantly improves the predictive model's adaptability under small sample conditions and enhances its transferability across equipment and operating conditions. It ensures accurate prediction even with small samples and significant differences in distribution between the source domain, thereby improving the adaptability and generalization ability of the rotating equipment life prediction model under different conditions. Simultaneously, the two-stage attention mechanism enhances the model's interpretability by explaining the contribution of each feature in the source domain at different stages. Furthermore, by introducing physical constraint equations into the harmonic enhancement network, this invention ensures that the extracted features not only conform to data patterns but also follow the physical characteristics of the equipment, thus improving the interpretability of the life prediction model. It also suppresses noise, improving the robustness and physical consistency of the prediction model. This multi-level interpretable mechanism makes the model decision-making process transparent, enhancing its credibility and usability in engineering applications, and providing an efficient and reliable tool for intelligent maintenance and reliability management of industrial equipment.
[0086] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing this application. Those skilled in the art will understand that the modules in the apparatus of the embodiment can be distributed within the apparatus of the embodiment as described, or can be modified to be located in one or more apparatuses different from this embodiment. The modules of the above-described embodiment can be combined into one module, or further divided into multiple sub-modules.
[0087] The serial numbers in this application are for descriptive purposes only and do not represent the superiority or inferiority of any particular implementation scenario. The above disclosures are merely a few specific implementation scenarios of this application; however, this application is not limited thereto, and any variations conceived by those skilled in the art should fall within the protection scope of this application.
Claims
1. A transfer prediction method integrating physical priors and a two-layer attention mechanism, characterized in that, include: The source domain time series and target domain time series of the rotating equipment are obtained, as well as an initial lifetime prediction model. The source domain time series and the target domain time series come from different rotating equipment or from the same rotating equipment under different operating conditions. The initial lifetime prediction model includes a feature extraction network, a two-stage attention mechanism network, and a lifetime prediction network. The source domain time series is input into a preset harmonic enhancement network for data enhancement to obtain an enhanced source domain time series, wherein the preset harmonic enhancement network is embedded with physical constraint equations. The enhanced source domain time series and the target domain time series are input into the feature extraction network for feature extraction to obtain the source domain device state feature vector and the target domain device state feature vector. The source domain device state feature vector and the target domain device state feature vector are respectively input into the two-stage attention mechanism network for adaptive weight allocation to obtain the source domain context vector and the target domain context vector. The weight allocation is based on the contribution of each feature in the time-frequency domain feature corresponding to the enhanced source domain time series at each time step in different stages. The source domain context vector is input into the lifetime prediction network to predict the lifetime, thereby obtaining the predicted lifetime of the rotating equipment. Based on the source domain context vector, the target domain context vector, and the predicted lifetime, the initial lifetime prediction model is iteratively trained to obtain a preset lifetime prediction model.
2. The method according to claim 1, characterized in that, Before inputting the source domain time series into a preset harmonic enhancement network for data enhancement to obtain the enhanced source domain time series, the method further includes: The harmonic characteristic equation is embedded into the physical information neural network to obtain the initial harmonic enhancement network; Based on the horizontal and vertical dynamic equations of the rotating equipment, the horizontal residual loss and the vertical residual loss are constructed. Based on the horizontal residual loss and the vertical residual loss, a residual loss function based on the dynamic equation is constructed; Based on the residual loss function, the initial harmonic enhancement network is iteratively trained to obtain the preset harmonic enhancement network.
3. The method according to claim 1, characterized in that, The source domain time series includes a horizontal vibration signal time series and a vertical vibration signal time series. The step of inputting the source domain time series into a preset harmonic enhancement network for data enhancement to obtain the enhanced source domain time series includes: The horizontal vibration signal time series and the vertical vibration signal time series are respectively input into a preset harmonic enhancement network for data enhancement, so as to obtain the enhanced horizontal vibration signal time series and the enhanced vertical vibration signal time series. The step of inputting the enhanced time-frequency domain features corresponding to the source domain time series and the time-frequency domain features corresponding to the target domain time series into the feature extraction network for feature extraction, to obtain the source domain device state feature vector and the target domain device state feature vector, includes: The time-frequency domain features corresponding to the enhanced horizontal vibration signal time series, the time-frequency domain features corresponding to the enhanced vertical vibration signal time series, and the time-frequency domain features corresponding to the target domain time series are respectively input into the feature extraction network for feature extraction to obtain the equipment state horizontal feature vector, the equipment state vertical feature vector, and the target domain equipment state feature vector. The step of inputting the source domain device state feature vector and the target domain device state feature vector into the two-stage attention mechanism network for adaptive weight allocation to obtain the source domain context vector and the target domain context vector includes: The horizontal feature vector and the vertical feature vector of the device state are respectively input into the two-stage attention mechanism network for adaptive weight allocation, so as to obtain the contribution weight of each feature in the time-frequency domain feature corresponding to the enhanced horizontal vibration signal time series at each time step under different stages, and the contribution weight of each feature in the time-frequency domain feature corresponding to the enhanced vertical vibration signal time series at each time step under different stages. Based on the contribution weights of each feature in the time-frequency domain of the enhanced horizontal vibration signal time series at each time step under different stages, and the contribution weights of each feature in the time-frequency domain of the enhanced vertical vibration signal time series at each time step under different stages, the attention heatmap, the horizontal source domain context vector, and the vertical source domain context vector are output. The target domain device state feature vector is input into the two-stage attention mechanism network for adaptive weight allocation to obtain the target domain context vector.
4. The method according to claim 3, characterized in that, The source domain context vector is input into the lifetime prediction network to predict the lifetime of the rotating equipment, including: The horizontal source domain context vector and the vertical source domain context vector are input into the lifetime prediction network to predict the lifetime and obtain the predicted lifetime of the rotating equipment. Based on the source domain context vector and the target domain context vector, and the predicted lifetime, the initial lifetime prediction model is iteratively trained to obtain a preset lifetime prediction model, including: Based on the horizontal source domain context vector, the vertical source domain context vector, the target domain context vector, and the predicted lifetime, a total loss function is constructed; Based on the total loss function, the initial lifetime prediction model is iteratively trained to obtain the preset lifetime prediction model.
5. The method according to claim 4, characterized in that, The step of constructing a total loss function based on the horizontal source domain context vector, the vertical source domain context vector, the target domain context vector, and the predicted lifetime includes: Calculate the first maximum mean difference loss based on the horizontal source domain context vector and the target domain context vector; Calculate the second maximum mean difference loss based on the vertical source domain context vector and the target domain context vector; Based on the predicted lifespan and the actual predicted lifespan of the rotating equipment, calculate the lifespan loss; The total loss function is constructed based on the first maximum mean difference loss, the second maximum mean difference loss, and the lifetime loss.
6. The method according to claim 3, characterized in that, The method further includes: Based on the attention heatmap, the first key feature affecting lifespan prediction is determined; The contribution of each feature in the time-frequency domain features corresponding to the enhanced horizontal vibration signal time series and the contribution of each feature in the time-frequency domain features corresponding to the enhanced vertical vibration signal time series are analyzed using the Shapley additive interpretation method. Based on the contribution of each feature in the time-frequency domain characteristics corresponding to the enhanced horizontal vibration signal time series and the contribution of each feature in the time-frequency domain characteristics corresponding to the enhanced vertical vibration signal time series, the second key feature affecting life prediction is determined. The intersection of the first key feature and the second key feature is taken to output the target key feature.
7. The method according to claim 1, characterized in that, The method further includes: The time-frequency domain features corresponding to the enhanced source domain time series are subjected to dimensionality reduction processing to obtain the dimensionality-reduced source domain time-frequency domain features; The time-frequency domain features corresponding to the target domain time series are subjected to dimensionality reduction processing to obtain the dimensionality-reduced target domain time-frequency domain features. Analyze the directional consistency between the reduced source domain time-frequency domain features and the reduced target domain time-frequency domain features; If the directions of the reduced source domain time-frequency domain features and the reduced target domain time-frequency domain features are consistent, then the data migration effect across operating conditions and equipment is determined to be good.
8. A migration prediction device integrating physical prior and a two-layer attention mechanism, characterized in that, include: The acquisition unit is used to acquire the source domain time series and the target domain time series of the rotating equipment, as well as the initial lifetime prediction model. The source domain time series and the target domain time series come from different rotating equipment or from the same rotating equipment under different operating conditions. The initial lifetime prediction model includes a feature extraction network, a two-stage attention mechanism network, and a lifetime prediction network. An enhancement unit is used to input the source domain time series into a preset harmonic enhancement network for data enhancement to obtain an enhanced source domain time series, wherein the preset harmonic enhancement network is embedded with physical constraint equations; The feature extraction unit is used to input the time-frequency domain features corresponding to the enhanced source domain time series and the time-frequency domain features corresponding to the target domain time series into the feature extraction network to extract features, thereby obtaining the source domain device state feature vector and the target domain device state feature vector. The weight allocation unit is used to input the source domain device state feature vector and the target domain device state feature vector into the two-stage attention mechanism network for adaptive weight allocation to obtain the source domain context vector and the target domain context vector. The weight allocation is based on the contribution of each feature in the time-frequency domain feature corresponding to the enhanced source domain time series at each time step in different stages. The prediction unit is used to input the source domain context vector into the lifetime prediction network to predict the lifetime and obtain the predicted lifetime of the rotating equipment. The training unit is used to iteratively train the initial lifetime prediction model based on the source domain context vector, the target domain context vector, and the predicted lifetime, to obtain a preset lifetime prediction model.
9. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.
10. An electronic device comprising a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 7.