Method, device, medium and product for few-shot cross-device migration diagnosis based on multi-source information fusion lightweight elastic meta-learning

Abstract

The application discloses a multi-source information fusion-based lightweight elastic meta-learning few-shot cross-device migration diagnosis method, device, medium and product, relates to the technical field of mechanical fault diagnosis, and comprises the following steps: performing feature enhancement and data reconstruction on multi-source heterogeneous signals to generate a source domain and a target data set; constructing source domain and target domain meta-tasks based on the source domain and the target data set; constructing a lightweight feature fusion meta-learner; introducing an elastic weight term, training and evaluating the lightweight feature fusion meta-learner based on the source domain and the target domain meta-tasks and the source data set and the target data set, and until a set iteration condition is reached, the trained lightweight feature fusion meta-learner is used as a cross-device migration diagnosis model, and based on real-time acquisition of cross-device to-be-tested signals, a migration diagnosis result is obtained. The application can solve the problems of limited few-shot and cross-device diagnosis performance, limited multi-source fusion information and insufficient generalization ability under inconsistent labels and time-varying working conditions.

Description

Technical Field

[0001] This application relates to the field of mechanical fault diagnosis technology, and in particular to a few-sample cross-device transfer diagnosis method, device, medium and product based on lightweight elastic element learning with multi-source information fusion. Background Technology

[0002] With the accelerated application of artificial intelligence and advanced sensing technologies in modern industrial systems, predictive and health management of complex industrial equipment has become increasingly important. Gearboxes, as a core component of many complex transmission systems, are widely used in aerospace, transportation, and energy sectors. Under harsh and variable operating conditions, gearboxes are prone to diverse and coupled failure modes, posing significant challenges to reliable condition monitoring and fault diagnosis.

[0003] Under such complex operating conditions, fault characteristics are often weak and exhibit diversity across various sensing methods, limiting the effectiveness of fault diagnosis methods based on single sensors. Therefore, intelligent fault diagnosis utilizing multi-source information fusion has attracted increasing attention. This method enables complementary utilization of multiple sensors, significantly improving diagnostic accuracy and robustness. Based on different data fusion strategies, multi-source information fusion methods are generally categorized into data-layer, feature-layer, and decision-layer fusion methods. In data-layer fusion, information processing techniques are used to effectively combine multi-source sensor signals into a single representation, achieving compact encoding with limited information loss. Feature-layer fusion employs neural networks to extract features from multi-source sensor signals and fuses these features through attention mechanisms and feature concatenation strategies to enrich feature diversity. Decision-level fusion weights and integrates the diagnostic outputs of multiple sub-networks, thereby improving diagnostic accuracy and robustness. Although intelligent diagnostic methods combining multi-source information fusion have shown good performance, some challenges remain in practical industrial applications. For example, intelligent diagnostic methods that combine multi-source information fusion with deep learning typically rely on large-scale labeled training data and assume that the data distribution of the training scenario and the test scenario are similar in order to achieve satisfactory diagnostic performance.

[0004] To alleviate these limitations, transfer learning offers an effective solution to overcome the challenges of cross-domain fault diagnosis. Transfer learning, through domain-invariant feature extraction, transfers knowledge from the source domain to the target domain, significantly improving diagnostic capabilities when data distribution shifts. Among various transfer learning methods, multi-source fusion domain adaptation, by leveraging multi-source information fusion, has attracted widespread attention for its ability to better handle complex domain differences. Furthermore, inspired by the complementary advantages of information fusion at different levels, the data layer, feature layer, and decision layer can be flexibly combined for transfer diagnosis within a unified multi-source fusion domain adaptation framework. Despite its robust and superior transfer diagnosis performance, multi-source fusion domain adaptation still faces several limitations. Specifically, existing methods typically require large amounts of source and target domain data, limiting their effectiveness in few-sample diagnostic tasks. Moreover, real-world industrial applications often present more complex target domain tasks, such as cross-device diagnosis, inconsistent label spaces between the source and target domains, and complex time-varying operating conditions. In such challenging tasks, multi-source fusion domain adaptation methods often exhibit limited diagnostic generalization, leading to a significant decline in diagnostic performance.

[0005] Meta-transfer learning holds significant potential in few-shot and generalized transfer diagnostics because it focuses on learning transferable meta-knowledge across different diagnostic tasks, enabling rapid adaptation to diagnostic tasks in the target domain using limited labeled data. Model-Agnostic Meta-Learning (MAML), a representative method in meta-transfer diagnostics, has been widely applied. Meanwhile, multi-source fusion meta-transfer diagnostic methods have also been explored, demonstrating superior diagnostic performance. However, despite their initial application in few-shot transfer diagnostics, several issues remain to be addressed. First, some studies have explored cross-device meta-transfer diagnostic tasks, but these existing methods are limited to considering stable operating conditions where the label spaces between the source and target domains are consistent. In contrast, real-world industrial applications are far more complex, exhibiting diverse failure modes across machinery under time-varying operating conditions, and new, unseen failure modes gradually emerge over long-term use. These factors often lead to inconsistent label spaces and more severe shifts in data distribution between different domains, posing greater challenges to few-shot cross-device transfer diagnostics. Second, most current multi-source fusion meta-transfer diagnostic methods focus on data-level fusion. While data-level fusion strategies can largely preserve fault information from multiple sensors, they often result in high computational costs and difficulty in accurately characterizing fine-grained discriminative features. Furthermore, relying solely on data-level fusion makes it difficult to capture deep correlations between multi-source sensor data, thus hindering the extraction of fine-grained discriminative features and limiting the effective utilization of complementary information between sources. Finally, existing multi-source fusion meta-transfer diagnostic methods are primarily designed for relatively simple transfer tasks, such as transfer diagnostics across various stable operating conditions. When faced with complex applications involving few-sample cross-device transfer diagnostics under label inconsistencies and time-varying conditions, diagnostic performance typically degrades significantly. Therefore, existing multi-source fusion meta-transfer diagnostic methods suffer from limited performance in few-sample and cross-device diagnostics under label inconsistencies and time-varying conditions, limited multi-source fusion information, and insufficient generalization ability. Summary of the Invention

[0006] To address the aforementioned problems in the existing technology, this application provides a method, device, medium, and product for few-sample cross-device migration diagnosis based on lightweight elastic meta-learning with multi-source information fusion.

[0007] To achieve the above objectives, this application provides the following solution: Firstly, this application provides a few-shot cross-device transfer diagnostic method based on lightweight elastic meta-learning with multi-source information fusion, including: Acquire multi-source heterogeneous signals across devices, and perform feature enhancement and data reconstruction on the multi-source heterogeneous signals to generate source domain datasets and target datasets; Construct source domain meta-tasks and target domain meta-tasks based on the source domain dataset and the target dataset; A lightweight feature fusion meta-learner is constructed; the lightweight feature fusion meta-learner consists of multiple convolutional-lightweight attention modules and a fully connected layer; An elastic weight term is introduced, and the lightweight feature fusion meta-learner is trained and evaluated using the source domain meta-task and the target domain meta-task, using the source domain dataset and the target dataset, until the set iteration conditions are met, and the trained lightweight feature fusion meta-learner is obtained. The trained lightweight feature fusion meta-learner is used as a cross-device transfer diagnostic model, and the transfer diagnostic results are obtained based on the real-time acquired cross-device test signals.

[0008] Secondly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the few-sample cross-device transfer diagnostic method based on multi-source information fusion lightweight elastic meta-learning provided above.

[0009] Thirdly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the aforementioned method for few-sample cross-device transfer diagnosis based on lightweight elastic meta-learning with multi-source information fusion.

[0010] Fourthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the aforementioned method for few-shot cross-device transfer diagnosis based on lightweight elastic meta-learning with multi-source information fusion.

[0011] According to the specific embodiments provided in this application, this application has the following technical effects: This application provides a few-shot cross-device transfer diagnostic method, device, medium, and product based on lightweight elastic meta-learning with multi-source information fusion. By enhancing features and reconstructing data from multi-source heterogeneous signals, source and target datasets are generated, effectively enhancing the discriminative features of the signals while improving the utilization rate of complementary information diversity. By constructing a lightweight feature fusion meta-learner consisting of multiple convolutional-lightweight attention modules and a fully connected layer, discriminative spatiotemporal features from multi-source fusion samples can be effectively integrated without introducing additional structural complexity. Local feature modeling is combined to highlight significant temporal variations, and global feature aggregation is combined to capture cross-channel spatial correlations. By introducing elastic weights during training, useful meta-level prior knowledge can be preserved and generalized parameter adjustments can be made, enhancing the training and generalization ability of the lightweight feature fusion meta-learner. This addresses the limitations of few-shot and cross-device diagnostic performance, limited multi-source fusion information, and insufficient generalization ability under inconsistent labeling and time-varying conditions. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 A flowchart illustrating a few-shot cross-device transfer diagnostic method based on lightweight elastic meta-learning with multi-source information fusion, provided as an embodiment of this application; Figure 2 This is a schematic diagram illustrating the process of data reconstruction and feature enhancement of multi-source heterogeneous signals according to an embodiment of this application; Figure 3 This is a schematic diagram of a multi-source information fusion generation process provided in an embodiment of this application; Figure 4 A schematic diagram of the structure of a lightweight feature fusion meta-learner provided in an embodiment of this application; Figure 5 This application provides a schematic diagram of the implementation process of a few-shot cross-device transfer diagnostic method based on lightweight elastic meta-learning with multi-source information fusion in one embodiment; Figure 6 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0014] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0015] Under conditions of label inconsistency and time-varying operation, this application develops a multi-source fusion meta-transfer diagnostic method with stronger diagnostic generalization ability. It develops a multi-source information fusion strategy to enhance the discriminative features of the signal and utilizes the diversity of complementary information to design a lightweight feature fusion meta-learner to integrate the discriminative temporal and spatial features from multi-source fusion samples. Furthermore, it utilizes an elastic weight meta-learning strategy to enhance the generalization ability of the lightweight feature fusion meta-learner, which is of key significance for robust cross-device transfer diagnostics with few samples.

[0016] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0017] In one exemplary embodiment, this application provides a few-shot cross-device transfer diagnostic method based on lightweight elastic meta-learning with multi-source information fusion. This method is executed by a computer device, specifically by a terminal or server alone, or by both. In this embodiment, the method is described using a server as an example. Figure 1 As shown, the method includes: Step 100: Acquire multi-source heterogeneous signals across devices, and perform feature enhancement and data reconstruction on the multi-source heterogeneous signals to generate source domain datasets and target datasets. For example, using multi-source heterogeneous sensors such as motor current clamps, accelerometers, speed and torque sensors, raw one-dimensional signals of planetary gearboxes under time-varying operating conditions and two different health states can be collected to achieve the acquisition of multi-source heterogeneous signals across devices.

[0018] Step 101: Construct source domain meta-tasks and target domain meta-tasks based on the source domain dataset and the target dataset.

[0019] Step 102: Construct a lightweight feature fusion meta-learner. The lightweight feature fusion meta-learner consists of multiple convolutional-lightweight attention modules and a fully connected layer. To balance feature extraction capability and model complexity, too few convolutional-lightweight attention modules will make it difficult to fully extract discriminative features from multi-source fusion samples; too many convolutional-lightweight attention modules will significantly increase the number of model parameters and computational complexity, easily leading to overfitting in few-sample meta-learning scenarios. Therefore, as... Figure 4As shown, this application preferably uses four convolutional-lightweight attention modules to form a lightweight feature fusion meta-learner.

[0020] Step 103: Introduce an elastic weight term. Based on the source domain meta-task and the target domain meta-task, train and evaluate the lightweight feature fusion meta-learner using the source domain dataset and the target dataset until the set iteration conditions (such as reaching the maximum number of iterations) are met, and the trained lightweight feature fusion meta-learner is obtained.

[0021] Step 104: Use the trained lightweight feature fusion meta-learner as a cross-device transfer diagnostic model, and obtain the transfer diagnostic results based on the real-time acquired cross-device test signals.

[0022] By implementing steps 100-104 above, this application can solve the problems of limited diagnostic performance with few samples and across devices, limited multi-source fusion information, and insufficient generalization ability under inconsistent labels and time-varying operating conditions.

[0023] In an exemplary embodiment of this application, to achieve multi-source information fusion of the source and target domains and further effectively enhance the discriminative features of the signals, the multi-source heterogeneous signals can first be learned through regularized dictionary projection to enhance the discriminative features in the multi-source heterogeneous sensor signals, and then data fusion can be performed. Next, a sample matrix is ​​randomly selected from the enhanced signals. Finally, the enhanced samples are mapped to three-channel color images to form multi-source fusion image samples. These samples are then used to construct multi-source fusion source and target domain datasets for designing different gearboxes and health states. Based on this, as... Figure 2 As shown, the process of performing feature enhancement and data reconstruction on multi-source heterogeneous signals to generate source domain datasets and target datasets may include: Step 1: Enhance multi-source heterogeneous signals using a regularized dictionary projection learning strategy to obtain a 3D enhanced signal. The implementation process of this step includes: Step 11: Determine the reconstruction error based on the multi-source heterogeneous signals and the projection dictionary, and update the projection dictionary with the goal of minimizing the reconstruction error until the reconstruction error is less than the tolerance error, thus obtaining the updated projection dictionary.

[0024] Step 12: Obtain the 3D augmented signal using the updated projection dictionary and multi-source heterogeneous signals.

[0025] In practical applications, the regularized dictionary projection learning strategy, by preserving the main cross-sensor structural variations while suppressing noise and redundancy, projects multi-source heterogeneous signals into a compact three-dimensional subspace, thereby generating a more discriminative representation for subsequent multi-source heterogeneous signal image construction. Based on this, we assume that the matrix of the multi-source heterogeneous signal is defined as... , This represents the length of a single sensing signal, while This represents the number of multi-source heterogeneous signals. The goal of the regularized dictionary projection learning strategy employed in this application is to learn a projection dictionary. The matrix of multi-source heterogeneous signals Mapped to a three-dimensional augmented signal array The projection dictionary is updated by minimizing the following regularization reconstruction error. : .

[0026] In the formula, It is a reconstruction error. It is the Frobenius norm regularization term, used for updating the dictionary after stable regularization. For regularization parameters, It is the Frobenius norm. T This is the matrix transpose.

[0027] In the t In the next iteration, the three-dimensional augmented signal array Defined as ,in It is the first t The projection dictionary in the next iteration. To improve computational efficiency, the Gram matrix is ​​defined as... Avoiding the complete data matrix Repeated multiplications are used to reduce the computational cost of high-dimensional datasets. Dictionary atoms. Perform the following independent updates: .

[0028] In the formula, and The first t Second and third Dictionary atoms at the next iteration. When reconstruction error... err Less than the tolerance error The iteration terminates at this point. Reconstruction error. err The definition is as follows: .

[0029] Step 2: Randomly select three-dimensional enhanced signals to form a sample matrix.

[0030] Step 3: Map the sample matrix into a three-channel color image to form a multi-source fusion sample.

[0031] In order to effectively integrate the three-dimensional enhanced signal array Leveraging the complementary diversity of information among multiple sensors, this application proposes a data mapping and fusion strategy. Unlike data-level fusion strategies based on stitching, the proposed strategy effectively captures the correlations between sensors while avoiding dimensionality expansion, thereby forming a compact and information-rich fused representation.

[0032] Randomly select a length of [value] from the three-dimensional augmented signal array. l The samples are converted into a pixel matrix, and then further converted into a two-dimensional grayscale matrix. i A two-dimensional grayscale matrix is ​​represented as Its height is Width is Therefore, the data mapping and fusion strategy is formulated as follows: .

[0033] In the formula, R , G and B These correspond to the red, green, and blue channels in the multi-source fusion sample, respectively. Indicates multi-source fusion samples, cat ( ) represents a concatenation function.

[0034] Step 4: Generate source domain datasets and target datasets based on multi-source fusion samples.

[0035] In one exemplary embodiment of this application, a source domain meta-task for meta-training and a target domain meta-task for evaluation are systematically constructed. In this embodiment, the source domain dataset and the target domain dataset are collected from two devices under time-varying operating conditions and different health states, respectively, and generated through multi-source information fusion. These datasets are then used as data inputs to a lightweight feature fusion meta-learner, such as... Figure 3 As shown.

[0036] In time-varying operating conditions, in cross-device, low-sample migration diagnostic scenarios, there are... m A cross-device migration task with different operating conditions. The first... t The dataset for each task is denoted as . ,in Represented as a multi-source fusion source domain dataset, Represented as a multi-source fusion target domain dataset, and Data were collected from two devices operating under different time-varying conditions. For the first... t Each task constructs a meta-task from source and target domain data. Each meta-task follows an "N-way K-shot" configuration. Define the first... i The dataset for the training set of each source domain element task is denoted as . This includes a source domain support set. Source domain query set , n Indicates the number of meta-tasks set in the source domain. N Indicates the number of categories in the dataset. K and M These represent the multi-source fusion samples for each class in the source domain support set and the source domain query set, respectively. The quantity, category label is The number of samples per class in the source domain support set. K Number of samples per class in the source domain query set M The total number of samples in each class is determined based on the total number of samples in the source domain training set, and the rule that the sum of the two sample numbers must not exceed the total number of samples in each class is satisfied.

[0037] In an exemplary embodiment of this application, to further effectively integrate discriminative spatiotemporal features from multi-source fusion samples without introducing additional structural complexity, the convolutional-lightweight attention module employed in this application consists of a 2-D convolution module, a local feature modeling module, and a global feature aggregation module. After multi-source information fusion in step 100, the multi-source signal samples are represented as multi-channel images, where the pixel distribution describes local temporal variations, while the channel dimension captures different feature responses and spatial information from multiple sensors. Based on this, for the acquired multi-source fusion samples, the lightweight feature fusion meta-learner uses a joint modeling method combining local feature modeling with global feature aggregation across pixel and channel dimensions, which can effectively capture discriminative fault information in the samples.

[0038] In practical applications, step 102 essentially involves integrating discriminative spatiotemporal features from MF samples. Specifically: (1) 2-D convolution module.

[0039] The 2D convolutional module can be composed of a convolutional layer with 64 3×3 convolutional kernels, a batch normalization layer, a ReLU activation function layer, and a max pooling layer stacked sequentially. This serves as the backbone network of a lightweight feature fusion meta-learner, outputting feature maps. It is used to extract hierarchical and discriminative deep features from multi-source fused images for further processing by subsequent local feature modeling and global feature aggregation modules.

[0040] (2) Local feature modeling module.

[0041] The local feature modeling module applies the pixel modulation mechanism independently to each channel, adaptively adjusting the response of each feature to highlight regions with significant temporal changes, thereby achieving fine-grained selection of discriminative local patterns.

[0042] For example, feature maps As input to the pixel modulation mechanism,C For the number of channels, H and W Representing feature maps respectively f Height and width. Pixel modulation mechanism in channel c ( c∈C Spatial location place ( h∈H, w∈W Adaptive calculation of pixel weights This reflects the representation of refined features by a single pixel. Its importance includes: .

[0043] In the formula, Indicates in the channel c spatial location Feature map at the location, It is a small constant that ensures numerical stability. and They represent feature maps respectively. f The Middle c The variance and mean of all pixels in each channel.

[0044] After calculating the pixel weights for all channels, a pixel weighted average is generated across all channel dimensions. g Detailed feature representation As shown below: .

[0045] .

[0046] In the formula, These are pixel weighting coefficients. This represents the Sigmoid activation function. It represents the Hadamaji.

[0047] (3) Global feature aggregation module.

[0048] In the global feature aggregation stage, a channel importance mechanism is introduced to adaptively reweight and refine the feature representation. The process effectively integrates multi-source features across channels to generate a fused and discriminative global feature representation. Based on this, the implementation process can be described as follows: First, the global feature aggregation module employs adaptive average pooling to compress two-dimensional information into importance scalars for all channels. This reflects the relative importance of different channels. The importance scalar for each channel... Represented as: .

[0049] In the formula, Indicates the first c Spatial position in each channel Refined feature representation at the location .

[0050] Then, a lightweight one-dimensional convolutional layer is used to capture cross-channel interactions, further generating a scalar of channel interaction importance for all channels. . No. c Channel interaction importance scalar It can be described as: .

[0051] In the formula, It is the first one-dimensional convolution. i Each core weight value, s Indicates the kernel size. Indicates the first Importance scalar in the channel z The value is zero-padded at the boundaries.

[0052] Finally, channel interaction feature representation It can be described as: .

[0053] .

[0054] In the formula, It is a scalar coefficient representing the importance of channel interaction.

[0055] By implementing collaborative interaction between local feature modeling and global feature aggregation in the convolutional-lightweight attention module, the lightweight feature fusion meta-learner can effectively fuse spatiotemporal distribution information without introducing additional structural complexity. Therefore, while maintaining a lightweight structure, the lightweight feature fusion meta-learner can produce more compact and discriminative feature representations, making it suitable for few-sample cross-device transfer diagnostics with label inconsistencies and time-varying conditions.

[0056] In an exemplary embodiment of this application, in order to enhance the generalization ability of meta-learning under label inconsistency and time-varying conditions, elastic weight meta-training and rapid adaptation can be performed. Therefore, by introducing elastic weight terms in step 103 above, useful meta-level prior knowledge can be maintained and parameter generalization adjustment can be achieved.

[0057] In practical applications, step 103 essentially involves introducing a parameter self-constraint mechanism into the inner loop update within the meta-learning framework to achieve an adaptive balance between rapidly adapting to new tasks and retaining general prior knowledge. Based on this, the implementation process of step 103 includes: (1) Elastic weight meta-training stage. An elastic weight term is introduced. Based on the source domain meta-task, the lightweight feature fusion meta-learner is trained with elastic weights using the source domain dataset and the target dataset to obtain the trained meta-parameters.

[0058] In practical applications, the training phase of elastic weights can include an inner loop based on meta-tasks and a meta-level outer loop parameter update process.

[0059] Introduce elastic weights directly into the meta-task-level inner loop. This restricts gradient-based updates while retaining some initial meta-level prior knowledge. For example, the meta-task-level inner loop learning steps are defined as follows: ei In the i The first meta-task k In the step, the inner loop parameters at the meta-task level The update process can be described as follows: .

[0060] In the formula, This represents the meta-task-level inner loop learning rate of the lightweight feature fusion meta-learner. It has a meta-task-level inner loop parameter. A lightweight feature fusion meta-learner function, It is the first i The meta-task in the first The inner loop parameters of the step's meta-task level. This indicates the use of multi-source fusion datasets. Below are parameters for the meta-task level inner loop. The loss gradient, For gradient factor, It is the weight decay parameter. Is the i-th meta-task at the th The parameters of the inner loop at the meta-task level of the step.

[0061] when At that time, the initial parameters of the inner loop at the meta-task level The parameters are set to the outer loop parameters obtained from the previous batch of meta-task level inner loops (i.e., the parameters of the lightweight feature fusion meta-learner). ).

[0062] The meta-level outer loop optimizes the parameters of the lightweight feature fusion meta-learner by minimizing the source domain query loss. This makes the learned parameters Capture knowledge unrelated to the meta-tasks within different meta-tasks. Parameters The resulting lightweight feature fusion meta-learner can be directly applied to few-sample cross-device transfer diagnostics under label inconsistencies and time-varying conditions when fully trained, or further adapted through meta-task-level inner loop updates when meta-training is insufficient. Parameters of the lightweight feature fusion meta-learner. The update process can be described as follows: .

[0063] In the formula, It is the learning rate of the outer loop at the meta level. m This represents the number of meta-tasks in the batch (i.e., the number of cross-device migration tasks with different operating conditions). Represents a multi-source fusion source domain query set The Middle i Regarding the inner loop parameters at the meta-task level The loss gradient, Represents the initial meta-parameters gradient, These represent the parameters of the updated lightweight feature fusion meta-learner. It's important to note that the parameters at the start of the first batch of training... It is initialized randomly.

[0064] Finally, the goal of elastic weighted meta-training is to minimize the average meta-task loss in a batch, as shown in the following formula: .

[0065] In the formula, Indicates the first i Source domain query set Above, a lightweight feature fusion meta-learner updated via gradient descent and weight decay is used. f The loss value obtained through evaluation.

[0066] (2) Rapid Adaptation Stage of Elastic Weights. Using the trained meta-parameters, parameter generalization adaptation is achieved in the multi-source fusion target domain support set until the set iteration conditions are met, resulting in a trained lightweight feature fusion meta-learner. The multi-source fusion target domain support set is generated based on the target dataset and the target domain meta-task.

[0067] In practical applications, elastic weights quickly adapt by utilizing well-trained meta-parameters obtained during the elastic weight meta-training phase. To achieve multi-source fusion target domain support set The parameter generalization adaptation in the model aims to achieve rapid generalization capability for small-sample cross-device migration diagnosis of label inconsistencies and time-varying operating conditions.

[0068] Quickly adapt parameters from initial elastic weights Initially, each multi-source fusion target domain meta-task was processed independently. Prediction and evaluation are performed after the subgradient update. Simultaneously, the elasticity weight term... It is also used in the rapid adaptation of elastic weights to constrain the adapted parameters. Orientation parameters This preserves the meta-level knowledge learned during the cross-domain task training phase of the elastic weights. The rapid adaptation process of elastic weights is as follows: .

[0069] In the formula, and Representing respectively in the and Parameters that adapt quickly during the step. It is a lightweight feature fusion meta-learner with parameters Related parameterized functions, To support sets in the multi-source fusion target domain Below are the parameters The loss gradient.

[0070] After completing the elastic weight meta-learning, the lightweight feature fusion meta-learner obtains well-adapted parameters. And apply it to the multi-source fusion target domain query set. It generates the final diagnostic results, which are suitable for diagnostics of small sample cross-device migration under time-varying operating conditions and label inconsistencies.

[0071] In traditional model meta-learning, meta-learning parameters are updated only along the gradient direction without regularization, which may lead to exceeding the generalization region and increasing the risk of overfitting. This application proposes a model-independent meta-learning strategy with elastic weights, which introduces an elastic weight term. As a pullback force, it limits the shift of parameters towards the initial meta-parameters. Therefore, the optimization process is guided to remain within the generalization region, closer to the optimal parameters. Thus, the elastic weight meta-learning strategy achieves effective knowledge transfer while mitigating overfitting, thereby improving cross-device transfer diagnostic robustness under label inconsistencies and time-varying conditions.

[0072] In one exemplary embodiment of this application, two planetary gearbox datasets can be used to evaluate the method provided in this application's ability to diagnose small-sample cross-device migration under label inconsistencies and time-varying operating conditions, and output the final diagnostic results. Figure 5 As shown, this application first collects vibration signals from two planetary gearboxes with inconsistent labels under time-varying operating conditions (i.e., corresponding...). Figure 5 Step 1) uses a multi-source information fusion strategy to generate multi-source fused samples from the two datasets (corresponding to...). Figure 5Step 2) is used to construct the source and target domain datasets. For each health state under all time-varying conditions in both datasets, 10 samples per category are used for source domain training, 30 samples for validation, and 30 samples for testing. Table 1 shows the settings for the cross-device transfer diagnostic task under label inconsistency and time-varying conditions. Subsequently, a lightweight feature fusion meta-learner (corresponding to...) is constructed. Figure 5 Step 3). Then, in the elastic weight meta-training phase, the parameters of the lightweight feature fusion meta-learner are iteratively updated through the meta-task-level inner loop parameters and the meta-level outer loop to learn generalized representations across different meta-tasks. In the rapid adaptation phase, the few-shot support set in the multi-source fusion target domain (i.e., the multi-source fusion target domain support set) is utilized. The trained lightweight feature fusion meta-learner is rapidly adapted. Finally, it is applied to the multi-source fusion target domain query set. This application comprehensively evaluates the diagnostic performance of small-sample cross-device migration under conditions of label inconsistency and time-varying operating conditions. It fully considers the impact of different operating conditions in industrial scenarios on diagnosis, providing a new method for fault diagnosis and offering important technical support for the safe and reliable operation of equipment.

[0073] Table 1. Cross-device migration diagnostic task setting table under label inconsistency and time-varying operating conditions

[0074] Furthermore, the effectiveness of the method provided in this application in few-shot cross-device migration diagnosis under label inconsistency and time-varying conditions is evaluated using evaluation metrics such as diagnostic accuracy. It should be noted that MF-LEMAML (Multi-source fusion lightweight elastic meta-learning) is the abbreviation for the method proposed in this application. IFMAML represents the MAML (meta-learning) method combining sparse principal component analysis and RGB fusion information enhancement; GMAML represents the MAML method incorporating multi-kernel efficient channel attention and inner loop guidance mechanisms; AMAML represents the MAML method employing spatial bilateral channel attention and adaptive meta-optimization strategies; LPCN represents the ProtoNet method fusing coordinate attention and L2 prototype correction; and WMAML++ represents the MAML++ method based on continuous wavelet transform information enhancement. The diagnostic accuracy of the proposed MF-LEMAML method in six different few-shot cross-device migration diagnosis tasks is evaluated, and the results are shown in Table 2. Clearly, among these six methods, the MF-LEMAML method achieves the best average diagnostic accuracy of 96.85% when facing label inconsistencies and time-varying operating conditions, outperforming the IFMAML method (7.97%), GMAML method (46.40%), AMAML method (51.86%), and LPCN method (49.38%). Compared to various meta-transfer diagnostic methods, the proposed MF-LEMAML method achieves superior diagnostic performance through a multi-source information fusion strategy, a lightweight meta-learner, and a flexible weighted meta-learning strategy, fully demonstrating the core advantages and practical application value of this application in few-sample cross-device transfer diagnostics under label inconsistencies and time-varying operating conditions.

[0075] Table 2. Evaluation results of diagnostic accuracy in six different few-sample cross-device transfer diagnostic tasks.

[0076] Based on the above description, compared with the prior art, this application has at least the following advantages: 1. This application provides a multi-source information fusion strategy that combines regularized dictionary projection learning with complementary information from multi-source sensor signals to achieve efficient multi-sensor signal enhancement and effectively enhance the discriminative features of the signals.

[0077] 2. The lightweight feature fusion meta-learner constructed in this application can achieve effective spatiotemporal feature fusion from multi-source fusion samples without introducing additional structural complexity. It effectively integrates discriminative spatiotemporal features from multi-source fusion samples and maintains a compact model structure. Furthermore, combining local feature modeling can highlight significant temporal changes, and combining global feature aggregation can accurately capture cross-channel spatial correlations.

[0078] 3. The elastic weight meta-learning strategy provided in this application, which combines elastic weight meta-training and elastic weight rapid adaptation, can preserve useful meta-level prior knowledge and adjust parameters for generalization, thereby enhancing the training and generalization capabilities of the lightweight feature fusion meta-learner.

[0079] 4. Experimental results on two planetary gearbox datasets demonstrate that the method provided in this application can be applied to few-sample cross-device migration diagnostics for label inconsistencies and time-varying operating conditions, and has greater generalization ability and robustness in few-sample cross-device migration diagnostics.

[0080] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 6 As shown, the computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs stored in the non-volatile storage media. The database stores few-shot cross-device transfer diagnostic data based on lightweight elastic meta-learning with multi-source information fusion. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a few-shot cross-device transfer diagnostic method based on lightweight elastic meta-learning with multi-source information fusion.

[0081] Those skilled in the art will understand that Figure 6 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer equipment to which the present application is applied. Specific computer equipment may include, for example, [the following is a list of possible additional structures]. Figure 6 The diagram shows more or fewer components, or combinations of certain components, or different component arrangements.

[0082] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0083] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0084] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0085] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0086] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (RRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0087] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0088] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0089] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A few-shot cross-device transfer learning method based on multi-source information fusion and lightweight elastic meta-learning, characterized in that, include: Acquire multi-source heterogeneous signals across devices, and perform feature enhancement and data reconstruction on the multi-source heterogeneous signals to generate source domain datasets and target datasets; Construct source domain meta-tasks and target domain meta-tasks based on the source domain dataset and the target dataset; Construct a lightweight feature fusion meta-learner; The lightweight feature fusion meta-learner consists of multiple convolutional-lightweight attention modules and a fully connected layer; An elastic weight term is introduced, and the lightweight feature fusion meta-learner is trained and evaluated using the source domain meta-task and the target domain meta-task, using the source domain dataset and the target dataset, until the set iteration conditions are met, and the trained lightweight feature fusion meta-learner is obtained. The trained lightweight feature fusion meta-learner is used as a cross-device transfer diagnostic model, and the transfer diagnostic results are obtained based on the real-time acquired cross-device test signals.

2. The method for few-shot cross-device transfer diagnostics based on lightweight elastic meta-learning with multi-source information fusion according to claim 1, characterized in that, The multi-source heterogeneous signals are subjected to feature enhancement and data reconstruction to generate source domain datasets and target datasets, including: The multi-source heterogeneous signal is enhanced using a regularized dictionary projection learning strategy to obtain a three-dimensional enhanced signal; The three-dimensional enhanced signals are randomly selected to form a sample matrix; The sample matrix is ​​mapped into a three-channel color image to form a multi-source fusion sample. The source domain dataset and the target dataset are generated based on the multi-source fusion samples.

3. The method for few-shot cross-device transfer diagnostics based on lightweight elastic meta-learning with multi-source information fusion according to claim 2, characterized in that, The multi-source heterogeneous signal is enhanced using a regularized dictionary projection learning strategy, including: Based on the multi-source heterogeneous signal and the projection dictionary, the reconstruction error is determined, and the projection dictionary is updated with the goal of minimizing the reconstruction error until the reconstruction error is less than the tolerance error, thus obtaining the updated projection dictionary. The three-dimensional enhancement signal is obtained using the updated projection dictionary and the multi-source heterogeneous signal.

4. The method for few-shot cross-device transfer diagnostics based on lightweight elastic meta-learning with multi-source information fusion according to claim 2, characterized in that, Randomly selecting the three-dimensional enhanced signal to form a sample matrix includes: A set length of samples is randomly selected from the three-dimensional enhanced signal, and the selected samples are converted into a pixel matrix; The pixel matrix is ​​converted into a two-dimensional grayscale matrix to obtain the sample matrix.

5. The method for few-shot cross-device transfer diagnostics based on lightweight elastic meta-learning with multi-source information fusion according to claim 2, characterized in that, The convolution-lightweight attention module consists of a 2D convolution module, a local feature modeling module, and a global feature aggregation module. The 2-D convolutional module is composed of convolutional layers, batch normalization layers, ReLU activation function layers, and max pooling layers stacked sequentially; the 2-D convolutional module is used to extract hierarchical and discriminative features from the multi-source fusion samples. The local feature modeling module applies the pixel modulation mechanism independently to each channel to adaptively adjust the response of each feature output by the 2-D convolution module. The global feature aggregation module introduces a channel importance mechanism to adaptively reweight the feature channels of the feature representation output by the local feature modeling module.

6. The method for few-shot cross-device transfer diagnostics based on lightweight elastic meta-learning with multi-source information fusion according to claim 1, characterized in that, Introducing elastic weights, the lightweight feature fusion meta-learner is trained and evaluated using the source domain meta-task and the target domain meta-task, employing the source domain dataset and the target dataset, until a set iteration condition is met, resulting in the trained lightweight feature fusion meta-learner, including: An elastic weight term is introduced, and the lightweight feature fusion meta-learner is trained with elastic weights based on the source domain meta-task and the target dataset to obtain trained meta-parameters. The trained meta-parameters are used to achieve parameter generalization adaptation in the multi-source fusion target domain support set until the set iteration conditions are met, thus obtaining the trained lightweight feature fusion meta-learner; the multi-source fusion target domain support set is generated based on the target dataset and the target domain meta-task.

7. The method for few-shot cross-device transfer diagnostics based on lightweight elastic meta-learning with multi-source information fusion according to claim 6, characterized in that, The elastic weight meta-training aims to minimize the average meta-task loss, as expressed as: ； In the formula, Indicates the initial meta-parameters. Represents a multi-source fusion source domain query set The Middle i Regarding the inner loop parameters at the meta-task level The loss gradient, Indicates the first i Meta-task level inner loop parameters for each meta-task Indicates the first i A dataset for a meta-task in a source domain training set. This indicates the inner loop parameter with meta-task level. A lightweight feature fusion meta-learner function. This indicates the number of meta-tasks in the batch. This represents the meta-task-level inner loop learning rate of the lightweight feature fusion meta-learner. Represents the loss gradient factor. Represents a multi-source fusion source domain query set The Middle i Regarding the inner loop parameters at the meta-task level The loss gradient, Indicates the inner loop parameter with meta-task level. A lightweight feature fusion meta-learner function. This represents the weight decay parameter. Indicates the first i The meta-task in the first The inner loop parameters of the step's meta-task level. Indicates the first i Source domain query set Above, a lightweight feature fusion meta-learner updated via gradient descent and weight decay is used. f The loss value obtained through evaluation.

8. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that the processor executes the computer program to implement the few-sample cross-device migration diagnostic method based on lightweight elastic meta-learning with multi-source information fusion as described in any one of claims 1-7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the few-sample cross-device transfer diagnostic method based on lightweight elastic meta-learning with multi-source information fusion as described in any one of claims 1-7.

10. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the few-sample cross-device transfer diagnostic method based on lightweight elastic meta-learning with multi-source information fusion as described in any one of claims 1-7.

Images