A method, system, medium, and equipment for predicting the temperature field of wet friction elements based on a cross-domain transfer learning model.

Abstract

This invention relates to the field of temperature field prediction for vehicle transmission components, and discloses a method, system, medium, and device for predicting the temperature field of wet friction components based on a cross-domain transfer learning model. The method includes: acquiring temperature field data and interface morphology data of the friction components, constructing a Transformer-LSTM-AdaBoost hybrid neural network model using heterogeneous physical quantity corresponding sample inputs, and pre-training the model as a feature extractor; constructing a DGDAN cross-domain transfer model, using interface morphology data as the source domain and temperature field data as the target domain, and inputting both into the DGDAN cross-domain transfer model for training, initially extracting common features from the two sets of data through the feature extractor; conducting adversarial training between the feature extractor and the domain discriminator through a gradient inversion layer, explicitly aligning the feature distributions of the two domains using the maximum mean difference metric, and obtaining the final common features; and inferring the full surface temperature field distribution online through the input interface morphology data, achieving dynamic state synchronization between the physical entity and the virtual model.

Description

Technical Field

[0001] This invention relates to the field of temperature field prediction technology for vehicle transmission components, and in particular to a method, system, medium, and device for predicting the temperature field of wet friction components based on a cross-domain transfer learning model. Background Technology

[0002] Wet friction elements are core components of heavy-duty vehicle transmission systems, and their safety and reliability directly affect the effectiveness of power transmission. Abnormal temperature distribution at the friction interface can lead to problems such as component warping, fluctuations in the coefficient of friction, and unstable torque transmission, and in severe cases, component failure. Therefore, accurately predicting the temperature field distribution of friction components is crucial for improving their service life and reliability.

[0003] In the study of temperature fields of wet friction elements, data acquisition is limited not only by the size and installation location of thermocouples, making it difficult to obtain a complete surface temperature distribution through direct measurement, but also by the extremely limited temperature field data obtained even when successful experimental acquisition, far from meeting the data scale required to support the modeling and analysis of temperature field failure mechanisms. However, in stark contrast to zero-sample temperature field data, interface morphology data is relatively easy to acquire in large quantities. Overcoming the bottleneck of modeling under zero-sample temperature field conditions has become an urgent technical problem to be solved.

[0004] In recent years, with the development of machine learning and transfer learning technologies, some scholars have begun to try to predict temperature fields using data-driven methods. For example, they have used models such as physical information neural networks and generative adversarial networks to reconstruct and predict temperature fields. However, most of these methods require that the data distributions of the source and target domains be consistent or similar. When temperature field data of friction elements is scarce, but interface morphology data is relatively abundant, traditional methods are difficult to apply directly. Summary of the Invention

[0005] To address the aforementioned problems, the present invention aims to provide a method, system, medium, and device for predicting the temperature field of wet friction elements based on a cross-domain transfer learning model. By using interface morphology data as the source domain and temperature field data as the target domain, the common features of the two are extracted, thereby effectively solving the zero-sample problem of temperature field in wet friction elements.

[0006] To achieve the above objectives, in a first aspect, the technical solution adopted by the present invention is as follows: a method for predicting the temperature field of wet friction elements based on a cross-domain transfer learning model, comprising: acquiring temperature field data and interface morphology data of friction elements using a wet clutch friction and wear test bench, forming samples corresponding to heterogeneous physical quantities; wherein, the amount of interface morphology data is much larger than the amount of temperature field data; inputting the samples corresponding to heterogeneous physical quantities into a constructed Transformer-LSTM-AdaBoost hybrid neural network model, pre-training the Transformer-LSTM-AdaBoost hybrid neural network model, and using the trained Transformer-LSTM-AdaBoost hybrid neural network model as a feature extractor for the DGDAN cross-domain transfer learning model; based on the feature extractor, the method further refines the prediction method based on the domain transfer learning model. A DGDAN cross-domain transfer model is constructed using an adversarial mechanism and a domain adaptation mechanism guided by the maximum mean difference. Interface morphology data is used as the source domain, and temperature field data is used as the target domain. Both are input into the DGDAN cross-domain transfer model for training. A feature extractor initially extracts common features from the two sets of data. A gradient reversal layer is used to conduct adversarial training between the feature extractor and the domain discriminator in the domain adversarial mechanism, driving the DGDAN cross-domain transfer model to learn domain-invariant features. At the same time, the feature distributions of the two domains are explicitly aligned using the maximum mean difference metric to obtain the final common features for predicting the temperature distribution of the friction element. The trained DGDAN cross-domain transfer model is integrated into a digital twin system, and the full surface temperature field distribution is inferred online using real-time input interface morphology data to achieve dynamic state synchronization between the physical entity and the virtual model.

[0007] Furthermore, in the Transformer-LSTM-AdaBoost hybrid neural network model constructed by inputting samples corresponding to heterogeneous physical quantities, the Transformer-LSTM-AdaBoost hybrid neural network model is pre-trained, including:

[0008] The measured rotational speed, pressure, sliding time, oil temperature, degradation stage, and initial interface morphology are used as input features, and the interface morphology corresponding to the same working condition under different degradation stages is used as the prediction target to construct a Transformer-LSTM-AdaBoost hybrid neural network model. Transformer employs a multivariate self-attention mechanism to globally encode the interface morphology data of friction elements, constructing a high-dimensional contextual relevance matrix to capture non-local dependencies and spatial mapping relationships. LSTM introduces a dynamic temporal information flow through cell state and gating mechanism to model the evolution of interface morphology with friction time and degradation stage, and learns the time dependence and state transition law in wear dynamics. As a meta-learner, AdaBoost can hierarchically integrate multiple weak learners and adaptively adjust sample weights to enhance the learning ability of complex samples under complex conditions.

[0009] Furthermore, common feature extraction includes: High-dimensional abstract features were extracted from interface morphology data and temperature field data samples using a feature extractor. The extracted high-dimensional abstract features are simultaneously input into two downstream modules: one is a label predictor, whose task is to regress the temperature field value based on the features; the other is a domain discriminator, which is trained to distinguish which domain the features come from. The DGDAN cross-domain transfer model introduces a gradient inversion layer during training. By inverting the gradient lost by the domain discriminator during backpropagation, it forces the feature extractor to generate domain-invariant features, thereby reducing the distribution difference between interface topography data and temperature field data. At the same time, the DGDAN cross-domain transfer model explicitly minimizes the distance between the feature distributions of the two domains through the maximum mean difference module, thus strengthening feature alignment. By combining adversarial training, common physical features related to the heating mechanism can be extracted from morphological data and transferred for the prediction of the temperature field of friction elements.

[0010] Furthermore, domain adversarial mechanisms include feature extractors. G f Tag predictor G p Domain discriminant G d ; Feature extractor G f Learn invariant feature representations from input data of interface morphology data and temperature field data; Tag predictor G p Predicting interface morphology data; Domain discriminator G d Attempt to distinguish between samples of interface morphology data and temperature field data; Domain adversarial training utilizes GAN-based mini-maximum games to reduce the distributional differences between interface topography data and temperature field data.

[0011] Furthermore, the objective loss function used in domain adversarial training is:

[0012] in, n s This indicates the number of labeled samples in the interface topography data; n t This indicates the number of unlabeled samples in the temperature field data;L p express G p The predicted loss; L d express G d Cross-entropy loss; G f ( x i , θ f ) represents the feature extractor G f High-level features generated; Let represent the true temperature field value corresponding to the i-th interface morphology sample in the source domain Ds; Ds represents the interface morphology data; Dt represents the temperature field data. Indicates a field label; This represents the set of all trainable parameters of the feature extractor Gf; The set of all trainable parameters for the domain discriminator Gd; This represents the set of all trainable parameters for the label predictor Gp.

[0013] Furthermore, the gradient inversion layer ensures that the gradient flowing through the feature extractor is negative, enabling the DGDAN cross-domain transfer model to simultaneously minimize the prediction loss and maximize the domain adversarial loss, and employs stochastic gradient descent to iteratively update parameters and minimize the overall objective loss.

[0014] Furthermore, the domain adaptation mechanism guided by the mean of the maximum mean difference includes: During the domain adaptation phase, the interface morphology data D s The sample is defined as { x s 1, x s 2, …, x s ns}, n s This indicates the number of interface topography data samples and temperature field data. D t The sample is defined as { x t 1, x t 2, …, x t nt}, n t This indicates the number of temperature field data samples. The maximum mean difference is used to measure the difference in distribution between samples from two domains.

[0015] In the formula, x s ns Indicates the first n s One sample of interface morphology data; x t nt Indicates the first n t A temperature field data sample; p represents the probability of the characteristic distribution of the interface morphology data; q represents the probability of the characteristic distribution of the temperature field data. Represents the feature mapping function; Represents the regenerated nucleus Hilbert space; This indicates that the interface topography data sample is in The following expectations; Indicates temperature field data samples in The expectations below.

[0016] Secondly, the technical solution adopted by this invention is as follows: a wet friction element temperature field prediction system based on a cross-domain transfer learning model, comprising: a data acquisition module, which uses a wet clutch friction and wear test bench to acquire temperature field data and interface morphology data of the friction element, forming samples corresponding to heterogeneous physical quantities; wherein, the amount of interface morphology data is much larger than the amount of temperature field data; a pre-training module, which inputs the samples corresponding to heterogeneous physical quantities into a constructed Transformer-LSTM-AdaBoost hybrid neural network model, pre-trains the Transformer-LSTM-AdaBoost hybrid neural network model, and uses the trained Transformer-LSTM-AdaBoost hybrid neural network model as a feature extractor for the DGDAN cross-domain transfer learning model; and a DGDAN cross-domain transfer learning module, which, based on the feature extraction... A DGDAN cross-domain transfer model is constructed using a feature extractor, a domain adversarial mechanism, and a domain adaptation mechanism guided by the maximum mean difference. Interface morphology data is used as the source domain, and temperature field data as the target domain, both input into the DGDAN cross-domain transfer model for training. A feature extractor initially extracts common features from the two datasets. A gradient inversion layer is used to train the feature extractor against the domain discriminator in the domain adversarial mechanism, driving the DGDAN cross-domain transfer model to learn domain-invariant features. Simultaneously, the feature distributions of the two domains are explicitly aligned using the maximum mean difference metric to obtain the final common features for predicting the temperature distribution of friction elements. The prediction output module integrates the trained DGDAN cross-domain transfer model into a digital twin system, using real-time input interface morphology data to infer the full-surface temperature field distribution online, achieving dynamic state synchronization between the physical entity and the virtual model.

[0017] Thirdly, the technical solution adopted by the present invention is: a computer-readable storage medium for storing one or more programs, wherein the one or more programs include instructions, which, when executed by a computing device, cause the computing device to perform any of the methods described above.

[0018] Fourthly, the technical solution adopted by the present invention is: a computing device comprising: one or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the one or more programs include instructions for performing any of the methods described above.

[0019] The present invention has the following advantages due to the adoption of the above technical solutions: 1. This invention uses readily available interface morphology data to predict temperature fields that are difficult to measure directly, thus overcoming the zero-sample limitation of temperature fields.

[0020] 2. This invention combines the global attention of Transformer with the temporal modeling capability of LSTM to improve the physical consistency of feature extraction.

[0021] 3. This invention achieves feature transfer across physical quantities by aligning with domain adversarial and MMD, thereby enhancing the model's generalization ability.

[0022] 4. This invention can serve as the core engine for digital twins, enabling virtual-physical state synchronization and dynamic prediction. Attached Figure Description

[0023] Figure 1 This is a flowchart of the wet friction element temperature field prediction method based on a cross-domain transfer learning model in an embodiment of the present invention. Figure 2 This is a schematic diagram of the network structure of the domain adversarial mechanism in an embodiment of the present invention. Detailed Implementation

[0024] To overcome the bottleneck of temperature field modeling under zero-sample conditions, this invention provides a method, system, medium, and device for predicting the temperature field of wet friction elements based on a cross-domain transfer learning model. By introducing a cross-domain transfer learning framework, a large amount of topographic data is used as the source domain, and a relatively small amount of temperature field data is used as the target domain to construct a distance-guided domain adversarial network (DGDAN). This enables high-precision prediction of the temperature field and transfer of failure features under imbalanced data conditions, achieving physically interpretable and high-precision prediction of the temperature field of wet friction elements.

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention are within the scope of protection of the present invention.

[0026] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0027] In one embodiment of the present invention, a method for predicting the temperature field of wet friction elements based on a cross-domain transfer learning model is provided. Based on the cross-domain transfer learning model, a distance-guided domain adversarial mechanism is used to establish a cross-physical quantity feature transfer mapping for wet friction elements, from "easy to measure, many samples" interface morphology to "difficult to measure, zero samples" temperature field. This achieves physically interpretable and high-precision temperature field prediction of friction elements, exhibiting good virtual physical consistency, thus providing a reliable foundation for state prediction and optimization decision-making. In this embodiment, as... Figure 1 As shown, the method includes the following steps: 1) Temperature field data and interface morphology data of friction elements are obtained using a wet clutch friction and wear test bench to form samples corresponding to heterogeneous physical quantities; among them, the amount of interface morphology data is much larger than the amount of temperature field data.

[0028] 2) In the Transformer-LSTM-AdaBoost hybrid neural network model constructed by inputting samples corresponding to heterogeneous physical quantities, the Transformer-LSTM-AdaBoost hybrid neural network model is pre-trained, and the trained Transformer-LSTM-AdaBoost hybrid neural network model is used as the feature extractor of the DGDAN cross-domain transfer model.

[0029] 3) Based on a feature extractor, a domain adversarial mechanism, and a domain adaptation mechanism guided by the maximum mean difference (MMD), a DGDAN cross-domain transfer model is constructed. Interface morphology data is used as the source domain, and temperature field data is used as the target domain. Both are input into the DGDAN cross-domain transfer model for training. The feature extractor initially extracts common features from the two sets of data. A gradient reversal layer is used to conduct adversarial training between the feature extractor and the domain discriminator in the domain adversarial mechanism, driving the DGDAN cross-domain transfer model to learn domain-invariant features. At the same time, the feature distributions of the two domains are explicitly aligned using the maximum mean difference metric to obtain the final common features for predicting the temperature distribution of friction elements. This effectively reduces the inter-domain differences, thereby enhancing the model's cross-domain generalization accuracy and stability.

[0030] 4) Integrate the trained DGDAN cross-domain transfer model into the digital twin system, and infer the full surface temperature field distribution online through real-time input interface morphology data to achieve dynamic state synchronization between the physical entity and the virtual model.

[0031] Because it is difficult to construct a holistic prediction model directly from the limited temperature measurement points of wet friction elements, and because there is a significant difference between the distribution of interface morphology and temperature field data, this invention adopts a cross-domain transfer learning framework. It uses easily obtainable interface morphology data as the source domain and temperature field data, which is difficult to measure directly, as the target domain, to effectively achieve transfer learning. This invention establishes a DGDAN cross-domain transfer learning model based on Transformer-LSTM-AdaBoost, a high-performance prediction core designed for digital twins, possessing the potential for bidirectional data flow. It is a real-time predictive digital twin model. This invention, through this DGDAN cross-domain transfer learning model, solves the problems of difficult acquisition of temperature field data and zero-sample acquisition for wet friction elements.

[0032] In step 1) above, accelerated life tests were conducted on a constructed wet clutch friction and wear test bench. Local surface temperature data was acquired by embedding thermocouples inside the friction plates. Based on a limited number of temperature points, the overall temperature field distribution of the disc was predicted, thus constructing the temperature field experimental samples required for subsequent analysis. Simultaneously, after each friction test, complete interface morphology data of the friction pair surface was acquired using a white light interferometer for offline scanning. Interface morphology determines local contact pressure and heat generation intensity, and is a key factor affecting the temperature field distribution. At the same time, the material softening and deformation caused by high temperature also have a reaction effect on the interface morphology. Therefore, there is a significant thermo-mechanical coupling relationship between the interface morphology of the friction element and the temperature field. Ultimately, the experiment extracted hundreds of sets of interface morphology data and dozens of sets of temperature field data, providing a sample of heterogeneous physical quantities corresponding to interface morphology data and temperature field data for cross-domain transfer modeling.

[0033] In step 2) above, the Transformer-LSTM-Adaboost hybrid neural network model achieves multi-scale feature extraction through a parallel heterogeneous architecture.

[0034] Specifically, in the Transformer-LSTM-AdaBoost hybrid neural network model constructed by inputting samples corresponding to heterogeneous physical quantities, the Transformer-LSTM-AdaBoost hybrid neural network model is pre-trained. Specifically, the measured rotational speed, pressure, sliding time, oil temperature, degradation stage and initial interface morphology are used as input features, and the interface morphology corresponding to the same working condition under different degradation stages is used as the prediction target to construct the Transformer-LSTM-AdaBoost hybrid neural network model.

[0035] The Transformer employs a multivariate self-attention mechanism to globally encode the interface morphology data of the friction element, constructing a high-dimensional contextual relevance matrix to capture nonlocal dependencies and spatial mapping relationships. The LSTM introduces a dynamic temporal information flow through cell states and gating mechanisms to model the evolution of the interface morphology with friction time and degradation stages, learning the temporal dependencies and state transition laws in wear dynamics. As a meta-learner, AdaBoost can hierarchically integrate multiple weak learners and adaptively adjust sample weights to enhance the learning ability for complex samples under complex working conditions. These three components complement each other in the feature space, temporal domain, and model structure, enabling high-precision prediction of the temperature field of wet friction elements under zero-sample conditions, while ensuring strong robustness and physical consistency.

[0036] In step 3) above, the surface of the friction element is composed of micro-protrusions of varying heights. The higher the micro-protrusion, the greater the local contact pressure and the larger the contact area, resulting in an increase in temperature in the corresponding surface region. Conversely, the smaller the micro-protrusions on the surface of the friction element, the lower the pressure it experiences during the bonding process, and the lower the temperature in the corresponding region. Therefore, the interface morphology of the friction element has a significant impact on the temperature field. Based on the correlation that the higher the surface micro-protrusions, the higher the temperature value in the corresponding region, this invention employs a cross-domain migration model to extract common features from the two sets of parameters to predict the temperature distribution of the friction element.

[0037] Specifically, common feature extraction includes the following steps: 3.1) Use a feature extractor to extract high-dimensional abstract features from the interface morphology data and temperature field data samples respectively.

[0038] 3.2) The extracted high-dimensional abstract features are simultaneously input to two downstream modules: one is a label predictor, whose task is to regress the temperature field value based on the features; the other is a domain discriminator, which is trained to distinguish which domain the features actually come from.

[0039] 3.3) To achieve effective knowledge transfer, the DGDAN cross-domain transfer model introduces a gradient reversal layer during training. By reversing the gradient lost by the domain discriminator during backpropagation, the feature extractor is forced to generate domain-invariant features, thereby reducing the distribution difference between interface morphology data and temperature field data. At the same time, the DGDAN cross-domain transfer model explicitly minimizes the distance between the feature distributions of the two domains through the maximum mean difference module, thereby strengthening feature alignment.

[0040] 3.4) By combining adversarial training, common physical features related to the heating mechanism can be extracted from the morphological data and transferred for the prediction of the temperature field of friction elements. Finally, high-precision and strong generalization temperature field reconstruction can be achieved in the case of zero samples of the temperature field of wet friction elements.

[0041] In step 3) above, the domain adversarial mechanism is a method used to solve the domain offset problem. It improves the model's generalization ability on temperature field data by learning a domain discriminator, enabling the model to minimize the data distribution differences between interface topography data and temperature field data when performing feature extraction. The domain adversarial mechanism includes a feature extractor. G f Tag predictor G p Domain discriminant G d Among them, feature extractor G f That is, a trained Transformer-LSTM-AdaBoost hybrid neural network model.

[0042] The principle of domain adversarial mechanism is as follows Figure 2 As shown. Feature extractor G f Learn invariant feature representations from input data including interface morphology and temperature field data; label predictor G p Predicting interface morphology data; Domain discriminator G d Attempts were made to distinguish between samples of interface topography data and temperature field data; domain adversarial training utilized a GAN-based mini-maximum game to reduce the distributional differences between interface topography data and temperature field data.

[0043] Specifically, let { x s , y s}={( x i s , y i s )}ns i=1 , representing interface shape data D s . { x t}={( x i t )} nt i=1 Representing temperature field data D t . n s This indicates the number of labeled samples in the interface topography data. n t This indicates the number of unlabeled samples in the temperature field data.

[0044] (1) (2) Given input x and its predicted labels y Feature extractor G f Generate high-level features G f ( x i , θ f ), the parameters are θ f Subsequently, these features were fed to... G p and G d In the middle, the parameters are obtained as θ p Advanced features G p ( G f ( x i , θ f ), θ p ) and parameters are θ d Advanced features G d ( G f ( x i , θ f ), θ d Finally, the target loss function is constructed.

[0045] In this embodiment, the target loss function used during domain adversarial training is: (3) in, n s This indicates the number of labeled samples in the interface topography data; n t This indicates the number of unlabeled samples in the temperature field data; L p express G p The predicted loss; L d express G d Cross-entropy loss; G f ( x i , θ f ) represents the feature extractor G f High-level features generated; Let represent the true value of the temperature field corresponding to the i-th interface morphology sample in the source domain Ds; D s This represents the interface shape data; D t Represents temperature field data; The domain label is a binary label (D) used to mark the source domain of the i-th sample. s Sample d i =1,D t Sample d i =0); This represents the set of all trainable parameters of the feature extractor Gf; The set of all trainable parameters for the domain discriminator Gd; This represents the set of all trainable parameters for the label predictor Gp.

[0047] To achieve this by solving the defined loss function E To obtain domain-invariant feature representations, adjust the parameters accordingly. θ d and θ p To minimize G d and G p At the same time, for parameters θ f Optimize to make G d Maximize this, thereby promoting the convergence of the two feature distributions.

[0048] In step 3) above, a gradient reversal layer (GRL) is used during backpropagation to facilitate the optimization process. The gradient reversal layer (GRL) ensures that the gradient flowing through the feature extractor is negative, enabling the DGDAN cross-domain transfer model to simultaneously minimize the prediction loss and maximize the domain adversarial loss, and uses stochastic gradient descent to iteratively update the parameters and minimize the overall objective loss.

[0049] (4) (5) G f High-dimensional abstract features are extracted using a Transformer and then fed into a multi-head attention module. The multi-head attention module helps to enhance attention to different parts of the input temperature data. The output of the multi-head attention module is further fed into an LSTM network. G p It consists of LSTMs that process time series data.

[0050] In step 3) above, in Domain Adaptation (DA), MMD is a widely used distance measurement method. It helps align marginal distributions and probability distributions, improving DA capabilities.

[0051] Specifically, the domain adaptation mechanism guided by the mean of the maximum mean difference includes: During the domain adaptation phase, the interface morphology data D s The sample is defined as { x s 1, x s 2, …, x s ns}, n s This indicates the number of interface topography data samples and temperature field data. D t The sample is defined as { x t 1, x t 2, …, x t nt}, n t This indicates the number of temperature field data samples. The maximum mean difference is used to measure the difference in distribution between samples from two domains. (6) In the formula, x s ns Indicates the first n s One sample of interface morphology data; x t nt Indicates the first n t A temperature field data sample; p represents the probability of the characteristic distribution of the interface morphology data; q represents the probability of the characteristic distribution of the temperature field data. Represents the feature mapping function; Represents the regenerated nucleus Hilbert space; This indicates that the interface topography data sample is in Expected value (mean) under the following conditions; Indicates temperature field data samples in The expected value (mean) under the given conditions.

[0052] A larger MMD value indicates a greater difference between the two distributions, while a smaller value indicates a closer similarity between them.

[0053] Domain adversarial mechanisms learn domain-invariant feature representations through adversarial training, independent of the labeling information of temperature field data. This makes them applicable to scenarios where labeled temperature field data is limited or unavailable. Therefore, DGDAN... G f The final layer uses MMD. Simultaneously, samples from both domains are... G f Extract as f s and f t and provide G d Further adaptation is then performed. By combining MMD with domain-invariant feature representation learning, the domain shift problem can be better addressed, resulting in better generalization performance.

[0054] The core function of the DGDAN cross-domain migration model is to align and map the common features between the temperature field, which is difficult to measure comprehensively, and the interface morphology, which can be accurately acquired offline. Then, the temperature field distribution can be inferred through the interface morphology data. This approach can overcome the physical limitations of sensor deployment and provide basic support for dynamic perception and prediction of relevant scenarios.

[0055] In summary, the DGDAN cross-domain transfer model of this invention is used to achieve cross-physical quantity feature transfer from interface morphology to temperature field, thereby solving the problem of zero-sample temperature field data. Since the temperature field of friction components operating in a sealed environment is difficult to measure directly and comprehensively, resulting in zero-sample temperature field data, while interface morphology data is relatively easy to obtain. This model is based on the physical mechanism that higher surface protrusions lead to greater contact pressure and higher local temperatures. It utilizes a distance-guided domain adversarial network to extract deep features common to the temperature field data in the target domain from the interface morphology data (source domain), thus achieving high-precision prediction under zero-sample temperature field conditions.

[0056] In one embodiment of the present invention, a wet friction element temperature field prediction system based on a cross-domain transfer learning model is provided, comprising: The data acquisition module uses a wet clutch friction and wear test bench to acquire temperature field data and interface morphology data of friction elements, forming samples corresponding to heterogeneous physical quantities; among them, the amount of interface morphology data is much larger than the amount of temperature field data. The pre-training module uses the samples corresponding to heterogeneous physical quantities as inputs to construct the Transformer-LSTM-AdaBoost hybrid neural network model, and pre-trains the Transformer-LSTM-AdaBoost hybrid neural network model. The trained Transformer-LSTM-AdaBoost hybrid neural network model is used as the feature extractor of the DGDAN cross-domain transfer model. The DGDAN cross-domain transfer module, based on a feature extractor, a domain adversarial mechanism, and a domain adaptation mechanism guided by the maximum mean difference, constructs a DGDAN cross-domain transfer model. Interface topography data is used as the source domain, and temperature field data as the target domain, both input into the DGDAN cross-domain transfer model for training. The feature extractor initially extracts common features from the two sets of data. A gradient inversion layer is used to conduct adversarial training between the feature extractor and the domain discriminator in the domain adversarial mechanism, driving the DGDAN cross-domain transfer model to learn domain-invariant features. Simultaneously, the maximum mean difference metric is used to explicitly align the feature distributions of the two domains, obtaining the final common features to predict the temperature distribution of the friction element. The prediction output module integrates the trained DGDAN cross-domain transfer model into the digital twin system, and infers the full surface temperature field distribution online through real-time input interface morphology data to achieve dynamic state synchronization between the physical entity and the virtual model.

[0057] In the above embodiments, the Transformer-LSTM-AdaBoost hybrid neural network model constructed by inputting samples corresponding to heterogeneous physical quantities includes pre-training the Transformer-LSTM-AdaBoost hybrid neural network model, which involves: The measured rotational speed, pressure, sliding time, oil temperature, degradation stage, and initial interface morphology are used as input features, and the interface morphology corresponding to the same working condition under different degradation stages is used as the prediction target to construct a Transformer-LSTM-AdaBoost hybrid neural network model. Transformer employs a multivariate self-attention mechanism to globally encode the interface morphology data of friction elements, constructing a high-dimensional contextual relevance matrix to capture non-local dependencies and spatial mapping relationships. LSTM introduces a dynamic temporal information flow through cell state and gating mechanism to model the evolution of interface morphology with friction time and degradation stage, and learns the time dependence and state transition law in wear dynamics. As a meta-learner, AdaBoost can hierarchically integrate multiple weak learners and adaptively adjust sample weights to enhance the learning ability of complex samples under complex conditions.

[0058] In the above embodiments, common feature extraction includes: High-dimensional abstract features were extracted from interface morphology data and temperature field data samples using a feature extractor. The extracted high-dimensional abstract features are simultaneously input into two downstream modules: one is a label predictor, whose task is to regress the temperature field value based on the features; the other is a domain discriminator, which is trained to distinguish which domain the features come from. The DGDAN cross-domain transfer model introduces a gradient inversion layer during training. By inverting the gradient lost by the domain discriminator during backpropagation, it forces the feature extractor to generate domain-invariant features, thereby reducing the distribution difference between interface topography data and temperature field data. At the same time, the DGDAN cross-domain transfer model explicitly minimizes the distance between the feature distributions of the two domains through the maximum mean difference module, thus strengthening feature alignment. By combining adversarial training, common physical features related to the heating mechanism can be extracted from morphological data and transferred for the prediction of the temperature field of friction elements.

[0059] In the above embodiments, the domain adversarial mechanism includes a feature extractor. G f Tag predictor G p Domain discriminant G d ; Feature extractor G f Learn invariant feature representations from input data of interface morphology data and temperature field data; Tag predictor G p Predicting interface morphology data; Domain discriminator G d Attempt to distinguish between samples of interface morphology data and temperature field data; Domain adversarial training utilizes GAN-based mini-maximum games to reduce the distributional differences between interface topography data and temperature field data.

[0060] In the above embodiments, the objective loss function used during domain adversarial training is:

[0061] in, n s This indicates the number of labeled samples in the interface topography data; n t This indicates the number of unlabeled samples in the temperature field data; L p express G p The predicted loss; L d express G d Cross-entropy loss; G f ( x i , θ f ) represents the feature extractor G f High-level features generated; Let represent the true temperature field value corresponding to the i-th interface morphology sample in the source domain Ds; Ds represents the interface morphology data; Dt represents the temperature field data. Indicates a field label; This represents the set of all trainable parameters of the feature extractor Gf; The set of all trainable parameters for the domain discriminator Gd; This represents the set of all trainable parameters for the label predictor Gp.

[0062] In the above embodiments, the gradient inversion layer ensures that the gradient flowing through the feature extractor is negative, enabling the DGDAN cross-domain transfer model to simultaneously minimize the prediction loss and maximize the domain adversarial loss, and to use stochastic gradient descent to iteratively update the parameters and minimize the overall objective loss.

[0063] In the above embodiments, the domain adaptation mechanism guided by the mean of the maximum average difference includes: During the domain adaptation phase, the interface morphology data D s The sample is defined as { x s 1, xs 2, …, x s ns}, n s This indicates the number of interface topography data samples and temperature field data. D t The sample is defined as { x t 1, x t 2, …, x t nt}, n t This indicates the number of temperature field data samples. The maximum mean difference is used to measure the difference in distribution between samples from two domains.

[0064] In the formula, x s ns Indicates the first n s One sample of interface morphology data; x t nt Indicates the first n t A temperature field data sample; p represents the probability of the characteristic distribution of the interface morphology data; q represents the probability of the characteristic distribution of the temperature field data. Represents the feature mapping function; Represents the regenerated nucleus Hilbert space; This indicates that the interface topography data sample is in The following expectations; Indicates temperature field data samples in The expectations below.

[0065] The system provided in this embodiment is used to execute the above-described method embodiments. For specific processes and details, please refer to the above embodiments, which will not be repeated here.

[0066] In summary, this invention proposes a cross-domain transfer learning model that achieves cross-physical quantity feature transfer from interface morphology to temperature field, overcoming the zero-sample limitation of temperature field. By constructing a Transformer-LSTM-AdaBoost hybrid neural network model, combining global attention and temporal modeling capabilities, the physical consistency and prediction accuracy of feature extraction are improved. Furthermore, this invention uses the cross-domain transfer model as the core engine of digital twins, achieving synchronous updates and dynamic predictions of virtual and physical states, supporting real-time online inference and closed-loop optimization.

[0067] In one embodiment of the present invention, a computing device is provided. This computing device can be a terminal and may include a processor, a communication interface, memory, a display screen, and an input device. The processor, communication interface, and memory communicate with each other via a communication bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. When the computer programs are executed by the processor, they implement the methods described in the above embodiments. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The communication interface is used for wired or wireless communication with external terminals. Wireless communication can be achieved through Wi-Fi, a management network, NFC (Near Field Communication), or other technologies. The display screen can be a liquid crystal display (LCD) or an e-ink display. The input device can be a touch layer covering the display screen, or buttons, a trackball, or a touchpad mounted on the casing of the computing device, or an external keyboard, touchpad, or mouse. The processor can call logical instructions stored in the memory.

[0068] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and sold or used as independent products, and can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0069] In one embodiment of the present invention, a computer program product is provided, the computer program product including a computer program stored on a non-transitory computer-readable storage medium, the computer program including program instructions, and when the program instructions are executed by a computer, the computer is able to perform the methods provided in the above-described method embodiments.

[0070] In one embodiment of the present invention, a non-transitory computer-readable storage medium is provided, which stores server instructions that cause a computer to perform the methods provided in the above embodiments.

[0071] The computer-readable storage medium provided in the above embodiments has a similar implementation principle and technical effect to the above method embodiments, and will not be described again here.

[0072] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0073] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0074] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0075] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting the temperature field of wet friction elements based on a cross-domain transfer learning model, characterized in that, include: Temperature field data and interface morphology data of friction elements were obtained using a wet clutch friction and wear test bench, forming a sample of heterogeneous physical quantities; among them, the amount of interface morphology data was much larger than the amount of temperature field data. In the Transformer-LSTM-AdaBoost hybrid neural network model constructed by inputting samples corresponding to heterogeneous physical quantities, the Transformer-LSTM-AdaBoost hybrid neural network model is pre-trained, and the trained Transformer-LSTM-AdaBoost hybrid neural network model is used as the feature extractor of the DGDAN cross-domain transfer model. Based on a feature extractor, a domain adversarial mechanism, and a domain adaptation mechanism guided by the maximum mean difference, a DGDAN cross-domain transfer model is constructed. Interface topography data is used as the source domain, and temperature field data is used as the target domain. Both are input into the DGDAN cross-domain transfer model for training. The feature extractor initially extracts common features from the two sets of data. A gradient reversal layer is used to conduct adversarial training between the feature extractor and the domain discriminator in the domain adversarial mechanism, driving the DGDAN cross-domain transfer model to learn domain-invariant features. At the same time, the feature distributions of the two domains are explicitly aligned using the maximum mean difference metric to obtain the final common features for predicting the temperature distribution of friction elements. The trained DGDAN cross-domain transfer model is integrated into the digital twin system, and the temperature field distribution of the entire surface is inferred online through real-time input interface morphology data, so as to achieve dynamic state synchronization between the physical entity and the virtual model.

2. The method for predicting the temperature field of wet friction elements based on a cross-domain transfer learning model as described in claim 1, characterized in that, In the Transformer-LSTM-AdaBoost hybrid neural network model constructed by inputting samples corresponding to heterogeneous physical quantities, the pre-training of the Transformer-LSTM-AdaBoost hybrid neural network model includes: The measured rotational speed, pressure, sliding time, oil temperature, degradation stage, and initial interface morphology are used as input features, and the interface morphology corresponding to the same working condition under different degradation stages is used as the prediction target to construct a Transformer-LSTM-AdaBoost hybrid neural network model. Transformer employs a multivariate self-attention mechanism to globally encode the interface morphology data of friction elements, constructing a high-dimensional contextual relevance matrix to capture non-local dependencies and spatial mapping relationships. LSTM introduces a dynamic temporal information flow through cell state and gating mechanism to model the evolution of interface morphology with friction time and degradation stage, and learns the time dependence and state transition law in wear dynamics. As a meta-learner, AdaBoost can hierarchically integrate multiple weak learners and adaptively adjust sample weights to enhance the learning ability of complex samples under complex conditions.

3. The method for predicting the temperature field of wet friction elements based on a cross-domain transfer learning model as described in claim 1, characterized in that, Common feature extraction includes: High-dimensional abstract features were extracted from interface morphology data and temperature field data samples using a feature extractor. The extracted high-dimensional abstract features are simultaneously input into two downstream modules: one is a label predictor, whose task is to regress the temperature field value based on the features; the other is a domain discriminator, which is trained to distinguish which domain the features come from. The DGDAN cross-domain transfer model introduces a gradient inversion layer during training. By inverting the gradient lost by the domain discriminator during backpropagation, it forces the feature extractor to generate domain-invariant features, thereby reducing the distribution difference between interface topography data and temperature field data. At the same time, the DGDAN cross-domain transfer model explicitly minimizes the distance between the feature distributions of the two domains through the maximum mean difference module, thus strengthening feature alignment. By combining adversarial training, common physical features related to the heating mechanism can be extracted from morphological data and transferred for the prediction of the temperature field of friction elements.

4. The method for predicting the temperature field of wet friction elements based on a cross-domain transfer learning model as described in claim 1, characterized in that, Domain adversarial mechanisms include feature extractors G f Tag predictor G p Domain discriminant G d ; Feature extractor G f Learn invariant feature representations from input data of interface morphology data and temperature field data; Tag predictor G p Predicting interface morphology data; Domain discriminator G d Attempt to distinguish between samples of interface morphology data and temperature field data; Domain adversarial training utilizes GAN-based mini-maximum games to reduce the distributional differences between interface topography data and temperature field data.

5. The method for predicting the temperature field of wet friction elements based on a cross-domain transfer learning model as described in claim 4, characterized in that, During domain adversarial training, the objective loss function used is: in, n s This indicates the number of labeled samples in the interface topography data; n t This indicates the number of unlabeled samples in the temperature field data; L p express G p The predicted loss; L d express G d Cross-entropy loss; G f ( x i , θ f ) represents the feature extractor G f High-level features generated; Let represent the true temperature field value corresponding to the i-th interface morphology sample in the source domain Ds; Ds represents the interface morphology data; Dt represents the temperature field data. Indicates a field label; This represents the set of all trainable parameters of the feature extractor Gf; The set of all trainable parameters for the domain discriminator Gd; This represents the set of all trainable parameters for the label predictor Gp.

6. The method for predicting the temperature field of wet friction elements based on a cross-domain transfer learning model as described in claim 1, characterized in that, The gradient inversion layer ensures that the gradient flowing through the feature extractor is negative, enabling the DGDAN cross-domain transfer model to simultaneously minimize the prediction loss and maximize the domain adversarial loss. It also employs stochastic gradient descent to iteratively update parameters and minimize the overall objective loss.

7. The method for predicting the temperature field of wet friction elements based on a cross-domain transfer learning model as described in claim 1, characterized in that, Domain adaptation mechanisms guided by the mean of the maximum mean difference include: During the domain adaptation phase, the interface morphology data D s The sample is defined as { x s 1, x s 2, …, x s ns }, n s This indicates the number of interface topography data samples and temperature field data. D t The sample is defined as { x t 1, x t 2, …, x t nt }, n t This indicates the number of temperature field data samples. The maximum mean difference is used to measure the difference in distribution between samples from two domains. In the formula, x s ns Indicates the first n s One sample of interface morphology data; x t nt Indicates the first n t A temperature field data sample; p represents the probability of the characteristic distribution of the interface morphology data; q represents the probability of the characteristic distribution of the temperature field data. Represents the feature mapping function; Represents the regenerated nucleus Hilbert space; This indicates that the interface topography data sample is in The following expectations; Indicates temperature field data samples in The expectations below.

8. A system for predicting the temperature field of wet friction elements based on a cross-domain transfer learning model, characterized in that, include: The data acquisition module uses a wet clutch friction and wear test bench to acquire temperature field data and interface morphology data of friction elements, forming samples corresponding to heterogeneous physical quantities; among them, the amount of interface morphology data is much larger than the amount of temperature field data. The pre-training module uses the samples corresponding to heterogeneous physical quantities as inputs to construct the Transformer-LSTM-AdaBoost hybrid neural network model, and pre-trains the Transformer-LSTM-AdaBoost hybrid neural network model. The trained Transformer-LSTM-AdaBoost hybrid neural network model is used as the feature extractor of the DGDAN cross-domain transfer model. The DGDAN cross-domain transfer module, based on a feature extractor, a domain adversarial mechanism, and a domain adaptation mechanism guided by the maximum mean difference, constructs a DGDAN cross-domain transfer model. Interface topography data is used as the source domain, and temperature field data as the target domain, both input into the DGDAN cross-domain transfer model for training. The feature extractor initially extracts common features from the two sets of data. A gradient inversion layer is used to conduct adversarial training between the feature extractor and the domain discriminator in the domain adversarial mechanism, driving the DGDAN cross-domain transfer model to learn domain-invariant features. Simultaneously, the maximum mean difference metric is used to explicitly align the feature distributions of the two domains, obtaining the final common features to predict the temperature distribution of the friction element. The prediction output module integrates the trained DGDAN cross-domain transfer model into the digital twin system, and infers the full surface temperature field distribution online through real-time input interface morphology data to achieve dynamic state synchronization between the physical entity and the virtual model.

9. A computer-readable storage medium for storing one or more programs, characterized in that, The one or more programs include instructions that, when executed by a computing device, cause the computing device to perform any of the methods described in claims 1 to 7.

10. A computing device, characterized in that, include: One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the one or more programs including instructions for performing any of the methods described in claims 1 to 7.

Images