Feature-label collaborative decoupling oltc compound fault diagnosis method and device

Abstract

This invention discloses a feature-label collaborative decoupling method and apparatus for OLTC composite fault diagnosis, belonging to the field of machine learning technology. Specifically, it includes collecting raw data samples of multiple views of the on-load tap changer (OLTC) of a converter transformer; extracting shared and private features of each view from the raw data samples using a constructed dual-feature encoder; constructing a feature-label collaborative decoupling model to divide the labels into shared and private labels, and after consistency alignment, generating shared and private label sets; concatenating these sets along the label dimension to generate the final predicted labels for the raw data samples; and outputting the composite fault diagnosis results for the OLTC. This invention achieves accurate composite fault diagnosis in scenarios with missing and sparse multi-label data in OLTC multi-views through a four-level process of feature-label collaborative decoupling, hierarchical contrastive learning, dual consistency alignment, and dual-branch classification fusion.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of power system equipment fault diagnosis and machine learning, specifically involving an OLTC composite fault diagnosis method and device with feature-label collaborative decoupling. Background Technology

[0002] The on-load tap changer (OLTC) of the converter transformer is a core component in ultra-high voltage direct current (UHVDC) transmission systems, regulating DC voltage and balancing transmission power. Its operating status directly determines the reliability of the power grid. OLTCs require frequent tap switching. Over long-term operation, due to factors such as mechanical wear, electrical aging, and insulation deterioration, faults often exhibit multi-type coupled composite characteristics—rather than single, independent faults. For example, mechanical system abnormalities can induce electrical faults, which in turn exacerbate insulation deterioration, forming a mechanical-electrical-insulation chain-like composite fault. If such composite faults are not diagnosed promptly and accurately, they can easily lead to OLTC switching jamming, contact welding, or even converter transformer shutdown, causing a sharp drop in DC transmission power and significant losses in cross-regional energy allocation.

[0003] Existing methods for diagnosing composite faults in converter transformer on-load tap changers (OLTCs) still have three prominent problems in practical applications: First, composite faults often simultaneously contain multiple types of coupled features, including mechanical, electrical, and insulation faults. The coupling between features is high and the boundaries are blurred, making it difficult for traditional methods to accurately distinguish different types of fault information. Second, OLTC monitoring data comes from diverse sources, including multi-view signals such as electrical quantities, mechanical quantities, physical quantities, and state quantities. However, under actual operating conditions, some views are often missing, resulting in incomplete features. Third, some fault combinations in composite faults occur with extremely low frequency, forming rare labels. Existing multi-label classification methods lack a dedicated modeling and identification mechanism for rare labels, leading to a high rate of missed detection for rare labels. Summary of the Invention

[0004] To address the aforementioned technical issues, this invention proposes a feature-label collaborative decoupling method and apparatus for OLTC composite fault diagnosis. Through a four-level process of feature-label collaborative decoupling, hierarchical comparative learning, dual consistency alignment, and dual-branch classification fusion, it achieves accurate composite fault diagnosis for OLTC multi-view data missing and multi-label sparse and rare scenarios.

[0005] In a first aspect, this invention proposes a feature-label collaborative decoupling method for OLTC composite fault diagnosis, specifically including the following steps:

[0006] The raw data samples of the on-load tap changer (OLTC) of the converter transformer are collected, and the shared features and private features of each view are extracted from the raw data samples by the constructed dual-feature encoder.

[0007] A feature-label collaborative decoupling model based on shared features, private features, and labels is constructed. Labels are divided into shared labels and private labels. After consistent alignment of shared labels and shared features, and private labels and private features, shared label sets and private label sets are generated.

[0008] The dual-label prediction model outputs the prediction probabilities of the shared label set and the private label set respectively. After concatenation along the label dimension, the final predicted label of the original data sample is generated, and the composite fault diagnosis result of OLTC is output.

[0009] Furthermore, the constructed dual-feature encoder is specifically composed of two different stacked encoders, represented as follows:

[0010] ;

[0011] In the above formula, v represents a specific view, and m represents the total number of views. This is the original data sample. To extract the first Shared features of views To extract the first Private characteristics of each view For the first Shared encoder for each view, No. A private encoder for each view.

[0012] Furthermore, after constructing the dual-feature encoder, it also includes assigning a feature reconstruction decoder to each view:

[0013] ;

[0014] via decoder network Obtain the Reconstructed data for each view ;

[0015] The loss from reconstructing the original data from multiple views is as follows:

[0016] ;

[0017] In the above formula, Represents the potential spatial dimension. Let i represent the original features of the i-th sample. This represents the reconstructed feature of the i-th sample. This represents the missing view index value of the v-th view of the i-th sample.

[0018] Furthermore, the step of extracting shared and private features of each view from the original data samples by constructing a dual-feature encoder also includes fusing an instance-level contrastive learning loss function and a category-level contrastive learning loss function in the dual-feature encoder.

[0019] The instance-level contrastive learning loss function consists of a positive sample pair loss and a negative sample pair loss function. The positive sample pair loss is generated by positive sample similarity based on the shared features of each view and the shared features of other views of the same sample instance. The negative sample loss is generated by negative sample similarity based on private features.

[0020] The category-level contrastive learning loss function consists of the correlation matrix of multiple views and label embedding information.

[0021] Furthermore, the specific steps of constructing a feature-label collaborative decoupling model based on shared features, private features, and labels to divide labels into shared labels and private labels include:

[0022] For each label Define a learnable semantic embedding For each view and each tag , view Shared and private feature vectors are concatenated with label semantic embeddings to generate joint-aware labels:

[0023] ;

[0024] Extracting nonlinear semantic features using MLP networks:

[0025] ;

[0026] We construct label-view weights to obtain the sharing probability of each label in different views and the variance term that measures cross-view stability, in order to generate a label sharing evaluation model;

[0027] Based on the tag sharing evaluation model and the pre-built threshold relationship, tags are divided into shared tags and private tags;

[0028] The sharing probability is expressed as:

[0029] ;

[0030] The variance term is represented as:

[0031] ;

[0032] Shared evaluation model The meaning is "sample". tags Probability of belonging to a shared tag:

[0033] ;

[0034] In the above formula, Indicates the first The first sample The first label in the The label vector under each view, and That is, the potential feature dimension is 3 times that of the previous dimension. This indicates that feature concatenation is performed in the last dimension. Indicates the first The first sample The first label in the Shared feature vectors across views Indicates the first The first sample The first label in the Private feature vectors under each view , indicating that the i-th sample corresponds to the i-th sample. The label is in the Intensity values ​​under each view It is an activation function. and This represents the weight matrix of the MLP network. , This represents the network bias term, corresponding to the offset parameters of the two-layer linear transformation. For balance coefficient, Indicates the first The first sample The variance term for each label.

[0035] Furthermore, the steps for generating shared tag sets and private tag sets after consistently aligning shared tags and shared features, and private tags and private features specifically include:

[0036] For the Shared features of each sample after weighted fusion and private characteristics They are represented as follows:

[0037] ;

[0038] Obtain a consistent representation of multi-view label vectors:

[0039] ;

[0040] Based on projection network This embeds shared features, private features, and labels into the same latent space:

[0041] ;

[0042] After calculating the label-feature residuals of the shared and private pathways to achieve consistent alignment, a shared label set and a private label set are generated.

[0043] The label-feature residuals of shared and private pathways are represented as follows:

[0044] ;

[0045] in, For the sample tags The probability of belonging to a shared tag.

[0046] Furthermore, the steps to generate the final predicted labels for the original data samples by concatenating the predicted probabilities of the shared label set and the private label set, respectively, using a dual-label prediction model, specifically include:

[0047] The shared features of each view are input into the shared classifier of the two-branch label prediction model. This yields the predicted probability for the shared tag set. ;

[0048] Private features of each view Independent private classifiers in a two-branched label prediction model Make predictions for private tag sets. ;

[0049] The prediction results of private features are weighted and fused to obtain:

[0050] ;

[0051] Based on the generated results of the shared and private label sets, the labels are concatenated along the label dimension to form the final predicted label for the sample.

[0052] Furthermore, the step of generating the final predicted label of the original data sample by concatenating the predicted probabilities of the shared label set and the private label set after outputting the predicted probabilities of the shared label set and the private label set through the dual-label prediction model and concatenating them along the label dimension also includes calculating the classification error of the shared label set and the private label set separately using binary cross-entropy loss.

[0053] Secondly, the present invention also proposes a feature-tag collaborative decoupling OLTC composite fault diagnosis device, applied to perform the OLTC composite fault diagnosis method as described in the first aspect, specifically including:

[0054] The data acquisition module is used to collect raw data samples from the multi-view of the on-load tap changer (OLTC) of the converter transformer.

[0055] The feature extraction module is used to extract the shared and private features of each view from the original data sample through a constructed dual-feature encoder;

[0056] The label generation module is used to construct a feature-label collaborative decoupling model based on shared features, private features, and labels. It divides labels into shared labels and private labels, and after aligning the shared labels and shared features, and the private labels and private features, it generates a shared label set and a private label set.

[0057] The prediction output module is used to output the prediction probabilities of the shared label set and the private label set respectively through the dual-label prediction model. After concatenation along the label dimension, the final predicted label of the original data sample is generated, and the composite fault diagnosis result of OLTC is output.

[0058] Thirdly, the present invention also proposes a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in the first aspect.

[0059] This invention proposes a feature-label collaborative decoupling OLTC composite fault diagnosis method and device. Specifically, it includes introducing a dual encoder structure on the feature side to distinguish and extract shared and private features from multi-view data, mitigating information loss caused by missing views; designing a label semantic modeling and shared-private partitioning mechanism on the label side, dividing labels into shared and private labels through learnable label embedding and view visibility modeling, realizing a unified process from feature-label interaction to label attribute discrimination; combining a hierarchical contrastive learning strategy to enhance feature discriminativeness and semantic consistency, and introducing a correlation-guided double consistency alignment mechanism to maintain the consistency between label semantic structure and feature geometric space; finally, using a shared-private dual-branch classifier to predict shared and private labels respectively and fusing them at the decision layer, thereby realizing feature-label collaborative modeling, semantic alignment, and result fusion under a unified framework, significantly improving the identification accuracy and robustness of composite and rare faults. Attached Figure Description

[0060] Figure 1 This is a flowchart of the feature-tag collaborative decoupling OLTC composite fault diagnosis method proposed in this invention;

[0061] Figure 2 This is a schematic diagram of the feature-label collaborative decoupling model framework in the OLTC composite fault diagnosis method with feature-label collaborative decoupling proposed in this invention.

[0062] Figure 3This is a schematic diagram of category-level contrastive learning in the feature-label collaborative decoupling OLTC composite fault diagnosis method proposed in this invention.

[0063] Figure 4 This is a framework diagram of the OLTC composite fault diagnosis device with feature-tag collaborative decoupling proposed in this invention. Detailed Implementation

[0064] The present application will now be described in further detail with reference to the accompanying drawings, but the scope of protection of the present application is not limited to the specific embodiments.

[0065] While current OLTC fault diagnosis technology covers multi-dimensional monitoring, it still faces the challenge of diagnosing complex faults due to their complex characteristics. The complexity of OLTC complex fault characteristics manifests in two aspects: Firstly, OLTC complex fault diagnosis relies on four types of multi-view data—electrical quantities, mechanical quantities, physical quantities, and equipment status—but in real-world scenarios, view gaps frequently occur, resulting in incomplete multi-view states. Secondly, existing multi-view-based... Figure 1 Diagnostic methods based on consistency modeling overemphasize common information between different views, obscuring the specific information of each view. This leads to the dilution of the characteristics of a single fault type in a complex fault, making it impossible to accurately locate the key causes in the fault chain.

[0066] On the other hand, the sample size of compound faults is far less than that of single faults, and some fault combinations become rare labels due to their low probability of occurrence, resulting in sparse label data. Existing diagnostic methods do not have a dedicated identification mechanism for rare labels and rely solely on conventional multi-label classification logic for prediction, making it difficult to capture the specific features corresponding to rare labels. This leads to a high rate of missed diagnoses for rare compound faults and fails to meet the diagnostic needs of OLTC for full-scenario fault coverage.

[0067] Based on this, this embodiment provides a feature-label collaborative decoupling method for OLTC composite fault diagnosis, such as... Figure 1-2 As shown, the specific steps include the following:

[0068] S101. Collect raw data samples of the on-load tap changer (OLTC) multi-view of the converter transformer, and extract the shared features and private features of each view from the raw data samples using a constructed dual-feature encoder.

[0069] In this embodiment, the raw data samples collected from the on-load tap changer (OLTC) multi-view of the converter transformer mainly include multi-view data such as electrical quantities, mechanical quantities, physical quantities, and equipment status. To compensate for the unavoidable loss of view data during the OLTC data extraction stage and to preserve the individuality and complementarity between views, this embodiment utilizes a dual-feature extraction encoder to extract the shared and private features of each view. Specifically, this embodiment sets up two different stacked encoders, as shown below:

[0070] ;

[0071] in For a certain view, For the total number of views, The original data, To extract the first Each view shares information. This is private information. For the first Shared encoder for each view, No. Each view has its own private encoder. Subsequently, to improve the encoder's view feature extraction capabilities, a feature reconstruction decoder is assigned to each view:

[0072] ;

[0073] in To reconstruct view information, through the decoder network. Obtain the Reconstructed data for each view In addition, to eliminate the negative impact of missing views on model training, a missing view index matrix was introduced. Finally, the loss for reconstructing the original data from multiple views is as follows:

[0074] ;

[0075] In the above formula, Represents the potential spatial dimension. Let i represent the original features of the i-th sample. This represents the reconstructed feature of the i-th sample. This represents the missing view index value of the v-th view of the i-th sample.

[0076] In the above dual feature extraction process, the common view features are obtained. and private view features To maintain the consistency of shared view features within the same sample, that is, to obtain the shared embedding features of a sample across different views. To align features in the latent space and effectively distinguish between shared and private view features, this embodiment employs a two-level cross-channel contrastive loss function, which integrates instance-level and class-level contrastive learning loss functions to enhance the shared feature encoder. and private feature encoder Its feature extraction and discrimination capabilities.

[0077] Instance-level contrastive loss: In the latent feature space learned by the encoder, contrastive learning is used to maximize consistency between different views. It assumes that each sample is an independent instance, bringing positive sample pairs closer together and pushing negative samples further apart. The loss function for general contrastive learning can be expressed as:

[0078] ;

[0079] in For anchor samples, As a positive sample, For negative samples, For negative sample set, (,) represents the method for calculating the similarity between two samples.

[0080] This embodiment requires a shared feature encoder to attempt to mine common features among views, preserving the basic attributes of the samples held by all views, so each view... Common features Anchor samples are the samples that provide the anchor data, while positive samples are other views of the same sample instance. Common features This allows us to obtain the similarity of positive sample pairs. .in The loss for positive sample pairs is obtained as follows:

[0081] ;

[0082] in This indicates the set of indexes that indicate the existence of the sample view. Indicates the existence of the first The first sample A view.

[0083] To separate common features from private features, let the negative sample be... ,in Meanwhile, to ensure the uniqueness and distinctiveness of the private features of multiple views of the same sample, and to make them more discriminative during classification, the private features of multiple views are treated as negative sample pairs, which can be written as... ,in This yields the loss function for negative sample pairs:

[0084] ;

[0085] To calculate the similarity between two samples, the cosine similarity method is used, i.e.

[0086] ;

[0087] Therefore, the final instance-level contrastive loss function can be expressed as:

[0088] ;

[0089] Category-level contrastive loss: To fully consider the correlation between samples of different categories, obtain semantic information across samples, and improve the discriminative ability of shared information, this embodiment introduces prior label correlation to guide the shared view features of similar instances to move closer together. Simultaneously, introducing label correlation helps avoid simply treating multi-label classification as multiple binary classification problems. The similarity matrix between samples in the incomplete multi-view space is defined as follows: in, Let Y represent the label vector, which is the similarity matrix of n samples. A missing label indicator matrix G is introduced to help reduce the negative impact of missing labels. Meanwhile, the relevance matrix of the multi-view dataset is defined as follows:

[0090] ;

[0091] Will Viewed as a similarity matrix with strong supervised information, and serving as a label structure graph, this prior graph can promote consistency between the spatial geometry of different samples and the label embedding spatial structure, thus facilitating the learning of shared information among similar categories. Figure 3 As shown, the following definitions will guide the learning of views and labels:

[0092] ;

[0093] in This represents the number of valid instance pairs. The loss is used to embed information using labels. guide The learning process allows the category relevance in the shared information space to continuously approach the similarity of the true labels. Compared to previous contrastive learning methods, this embodiment constructs a similarity matrix based on the shared features of multiple views for learning, rather than fusing consistency information across multiple views. This effectively avoids over-eliminating the characteristics of each view. Furthermore, this embodiment no longer simply pushes away negative sample pairs, but guides samples to maintain an appropriate relative distance through label geometry. This ensures that cross-view similarity is maintained. Figure 1 While maintaining consistency, it avoids disrupting the inherent relationships between samples and prevents the incorrect identification of non-negative samples as negative samples, thus avoiding class conflicts during training.

[0094] In summary, the final two-level contrast loss can be expressed as follows:

[0095] ;

[0096] in This is the corresponding penalty function.

[0097] S102. Construct a feature-label collaborative decoupling model based on shared features, private features, and labels. Divide the labels into shared labels and private labels, and after aligning the shared labels and shared features, and the private labels and private features in a consistent manner, generate a shared label set and a private label set.

[0098] In this embodiment, relying solely on feature separation ignores the separability and modal dependencies of different labels. Private labels on the view are forced to depend on the geometry and consistency of shared paths, which can easily lead to label-level mismatches. Therefore, this embodiment, based on feature separation, performs semantic modeling and attribute segmentation on the labels, dividing them into shared labels and private labels.

[0099] In this embodiment, for each label Define a learnable semantic embedding For each view and each tag The shared and private feature vectors of the view are concatenated with the embedded label semantics to achieve interactive modeling of label semantics and multi-view features, generating joint-aware labels:

[0100] ;

[0101] in Indicates the first The first sample The first label in the The label vector under each view, and That is, the potential feature dimension is 3 times that of the previous dimension. This indicates that feature concatenation is performed in the last dimension. Indicates the first The first sample The first label in the Shared feature vectors across views Indicates the first The first sample The first label in the Private feature vectors under each view. And non-linear semantic features are extracted using an MLP network:

[0102] ;

[0103] It is an activation function. and This represents the weight matrix of the MLP network. , This represents the network bias term, corresponding to the offset parameters of the two-layer linear transformation. To measure the visibility of a label in different views, a label-view weight is introduced, whose output value represents the strength of the label's presence in that view.

[0104] ;

[0105] in For the Sigmoid function, Indicates the network bias term. , indicating that the i-th sample corresponds to the i-th sample. The label is in the Intensity value under each view, when The larger the value, the higher the value of the label. In the The more significant the label is across different views, the more significant it becomes. Furthermore, the mean salience of each label across different views is defined as the shared probability of the label:

[0106] ;

[0107] To avoid noise interference, the first The first sample The variance term of each label is used to measure cross-view stability. Finally, a label-sharing evaluation model is generated by combining the significance mean and stability. In this embodiment, the evaluation model uses... The meaning of "sample" is: tags The probability of belonging to a shared tag is expressed as follows:

[0108] ;

[0109] in For balance coefficient, The larger the value, the more likely the label belongs to a shared view. Set a threshold. Utilizing the obtained probability The threshold is used to classify tags into shared tags. and private tags .

[0110] In this embodiment, step S102 realizes the process from feature-label interaction to label attribute partitioning by semantic reconstruction of label features and joint modeling of feature space, which enhances the model's ability to understand the semantic structure of labels and effectively improves label distinguishability and cross-modal consistency in multi-view scenarios.

[0111] In this embodiment, after label segmentation, consistency alignment is performed on shared labels and shared features, and private labels and private features, respectively. Simultaneous optimization of both pathways is achieved through a dual consistency loss guided by label relevance. This network introduces co-occurrence correlation between labels and constructs a label similarity graph based on this, thereby explicitly modeling the consistency of the label structure during the shared and private alignment process. For the first... Shared features of each sample after weighted fusion and private characteristics They are represented as follows:

[0112] ;

[0113] Obtain a consistent representation of multi-view label vectors:

[0114] ;

[0115] Based on projection network This embeds shared features, private features, and labels into the same latent space:

[0116] ;

[0117] It should be noted that, in this embodiment, the projection network It means to The mapping function for the parameters belongs to the commonly used mapping technology field in this invention, and this embodiment does not specifically limit it. This projection unifies multi-view features and label features into the same embedding space, enabling direct comparison and alignment of information from different sources, thereby achieving shared semantics at the geometric level. To characterize the potential dependencies between labels, a label similarity matrix is ​​defined based on the label co-occurrence probability:

[0118] ;

[0119] in Indicates label The number of times it appears in the training samples. Indicates label and The number of times they occur simultaneously. And through the degree matrix. The normalized graph Laplacian matrix is ​​obtained. This matrix reflects the global semantic relationships in the label space, providing structural priors for subsequent alignment constraints.

[0120] To achieve geometric consistency between shared and private labels in their respective attribute spaces, this embodiment calculates the label-feature residuals for the shared and private paths, respectively. The label-feature residuals for the shared and private paths are expressed as follows:

[0121] ;

[0122] ;

[0123] in For the sample tags The probability of belonging to a shared label. By using residual matrix transformation, the errors of shared and private channels can be measured holistically, while soft assignment is achieved using probability weights. To simultaneously ensure the consistency of label-view alignment and the relevance of label structure, this embodiment proposes a label relevance-guided dual consistency loss:

[0124] ;

[0125] This alignment loss aims to incorporate the statistical correlation between labels into the shared and private alignment process, enabling related labels and views to converge collaboratively in the latent space, thereby maintaining both the geometric consistency of the label-view pair and the semantic consistency of the label structure.

[0126] S103. The prediction probabilities of the shared label set and the private label set are output separately through the dual-label prediction model. After concatenation in the label dimension, the final predicted label of the original data sample is generated, and the composite fault diagnosis result of OLTC is output.

[0127] In this embodiment, the tag set has been divided into a shared tag set through the aforementioned steps. With private tag sets To achieve accurate prediction of these two types of labels without losing some view-private or rare label information, this embodiment designs a dual-branch label prediction mechanism, which independently classifies the labels using fused shared features and multi-view private features, and obtains the final prediction result through label-level fusion.

[0128] Specifically, the shared features of each view are input into the shared classifier of the two-branch label prediction model. This yields the predicted probability for the shared tag set. ;

[0129] Private features of each view Independent private classifiers in a two-branched label prediction model Make predictions for private tag sets. ;

[0130] The prediction results of private features are weighted and fused to obtain:

[0131] ;

[0132] Based on the generated results of the shared and private label sets, the labels are concatenated along the label dimension to form the final predicted label for the sample:

[0133] ;

[0134] The classifier described above can simultaneously consider both the sharing of view labels and the unique characteristics of each view label. Given that the labels are independent in multi-label tasks, this embodiment uses binary cross-entropy loss to calculate the classification error for shared labels and private labels separately. The loss for the shared label channel is defined as:

[0135] ;

[0136] in Let be the true value of the shared label for the sample. Similarly, the loss function for the private label channel is:

[0137] ;

[0138] Combining the two types of losses, we obtain the bi-branch classification loss:

[0139] ;

[0140] In summary, the above reconstruction losses Layered contrast loss Double consistency loss Integration with final decision The total loss equation obtained is:

[0141] ;

[0142] in This is the corresponding non-negative penalty parameter.

[0143] To more clearly illustrate the technical solution of this embodiment, four state-of-the-art and most representative incomplete multi-view multi-label classification methods were selected for comparison, including LMVCAT, DICNet, MTD, and UPDGD-net. In this experiment, all models were tested with optimal parameters and underwent multiple trials to obtain results, avoiding the influence of random factors. Common datasets such as corel5k, iaprtc12, and mirrlicker were chosen. To ensure the accuracy of the tests, this embodiment used five metrics—AP, 1-RL, AUC, 1-OE, and 1-Cov—to comprehensively evaluate the models. The specific results are shown in the table below:

[0144] Table 1. Performance comparison of different models on three datasets

[0145] ;

[0146] Compared to the other five methods, the method proposed in this embodiment demonstrates superior classification performance across all datasets, ranking first in most metrics. Specifically, it improves mean accuracy (AP) by approximately 1%-2%.

[0147] This embodiment also conducted experiments to quantify the model's classification ability for rare categories. Based on the imbalance ratio... (Label The ratio of negative samples to positive samples (50 < 100) is used to classify rare labels into three levels: rare1 (50 < 100) and rare2 (50 < 100) and rare3 (50 < 100) and rare4 (50 < 100) and rare5 (50 < 100) and rare3 (50 < 100) and rare4 (50 < 100) and rare3 ... 150), rare2 (150< 250) and rare3 ( The model's ability to classify rare labels was quantified experimentally on the Core L5K dataset (>250). The experimental results are shown in Table 2, demonstrating that this model outperforms other models in handling rare categories.

[0148] Table 2 Comparison of rare label classification performance of different models on the core5k dataset.

[0149] ;

[0150] According to the purpose of this application embodiment, this embodiment also provides a feature-tag collaborative decoupling OLTC composite fault diagnosis device 40, which is applied to perform the OLTC composite fault diagnosis method as described above, such as Figure 4 As shown, it specifically includes:

[0151] The data acquisition module 401 is used to acquire raw data samples from the multi-view of the on-load tap changer (OLTC) of the converter transformer.

[0152] The feature extraction module 402 is used to extract the shared features and private features of each view from the original data sample through the constructed dual feature encoder;

[0153] The label generation module 403 is used to construct a feature-label collaborative decoupling model based on shared features, private features, and labels. The model divides the labels into shared labels and private labels, and after aligning the shared labels and shared features, and the private labels and private features, it generates a shared label set and a private label set.

[0154] The prediction output module 404 is used to output the prediction probability of the shared label set and the prediction probability of the private label set respectively through the dual-label prediction model. After concatenation in the label dimension, the final prediction label of the original data sample is generated, and the composite fault diagnosis result of OLTC is output.

[0155] This application provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the methods described in the above embodiments.

[0156] This application may be a system, apparatus, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of this application.

[0157] The various embodiments of this application have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical applications, or technical improvements to the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein. The scope of this application is defined by the appended claims.

Claims

1. A feature-label collaborative decoupling method for OLTC composite fault diagnosis, characterized in that, Specifically, the steps include the following: The raw data samples of the on-load tap changer (OLTC) of the converter transformer are collected, and the shared features and private features of each view are extracted from the raw data samples by the constructed dual-feature encoder. A feature-label collaborative decoupling model based on shared features, private features, and labels is constructed. Labels are divided into shared labels and private labels. After aligning shared labels with shared features and private labels with private features, a shared label set and a private label set are generated. Specifically, this includes: For each label Define a learnable semantic embedding For each view and each mark , view Shared and private feature vectors are concatenated with label semantic embeddings to generate joint-aware labels: ； Extracting nonlinear semantic features using MLP networks: ； We construct label-view weights to obtain the sharing probability of each label in different views and the variance term that measures cross-view stability, in order to generate a label sharing evaluation model; Based on the tag sharing evaluation model and the pre-built threshold relationship, tags are divided into shared tags and private tags; The sharing probability is expressed as: ； The variance term is represented as: ； Shared evaluation model The meaning is "sample". tags Probability of belonging to a shared tag: ； In the above formula, Indicates the first The first sample The first label in the The label vector under each view, and That is, the potential feature dimension is 3 times that of the previous dimension. This indicates that feature concatenation is performed in the last dimension. Indicates the first The first sample The first label in the Shared feature vectors across views Indicates the first The first sample The first label in the Private feature vectors under each view , indicating the first The sample corresponds to the first The label is in the Intensity values ​​under each view It is an activation function. and express Network weight matrix, This represents the network bias term, corresponding to the offset parameters of the two-layer linear transformation. For balance coefficient, Indicates the first The first sample The variance term for each label, Indicates the first The first sample Missing view index value for a view; The dual-label prediction model outputs the prediction probabilities of the shared label set and the private label set respectively. After concatenation along the label dimension, the final predicted label of the original data sample is generated, and the composite fault diagnosis result of OLTC is output.

2. The OLTC composite fault diagnosis method with feature-tag collaborative decoupling according to claim 1, characterized in that, The constructed dual-feature encoder is specifically two different stacked encoders, represented as follows: ； In the above formula, For a certain view, For the total number of views, This is the original data sample. To extract the first Shared features of views To extract the first Private characteristics of each view For the first Shared encoder for each view, No. A private encoder for each view.

3. The OLTC composite fault diagnosis method with feature-tag collaborative decoupling according to claim 2, characterized in that, After constructing the dual-feature encoder, it also includes assigning a feature reconstruction decoder to each view: ； via decoder network Obtain the Reconstructed data for each view ; The loss from reconstructing the original data from multiple views is as follows: ； In the above formula, Represents the potential spatial dimension. Indicates the first The original features of each sample Indicates the first Reconstructed features of each sample Indicates the first The first sample Missing view index value for each view.

4. The OLTC composite fault diagnosis method with feature-tag collaborative decoupling according to claim 3, characterized in that, The step of extracting shared and private features of each view from the original data samples by constructing a dual-feature encoder also includes fusing an instance-level contrastive learning loss function and a category-level contrastive learning loss function in the dual-feature encoder. The instance-level contrastive learning loss function consists of a positive sample pair loss and a negative sample pair loss function. The positive sample pair loss is generated by positive sample similarity based on the shared features of each view and the shared features of other views of the same sample instance. The negative sample loss is generated by negative sample similarity based on private features. The category-level contrastive learning loss function consists of the correlation matrix of multiple views and label embedding information.

5. The OLTC composite fault diagnosis method with feature-tag collaborative decoupling according to claim 4, characterized in that, The specific steps for generating shared label sets and private label sets after aligning shared labels and shared features, and private labels and private features include: For the Shared features of each sample after weighted fusion and private characteristics They are represented as follows: ； Obtain a consistent representation of multi-view label vectors: ； Based on projection network This embeds shared features, private features, and labels into the same latent space: ； After calculating the label-feature residuals of the shared and private pathways to achieve consistent alignment, a shared label set and a private label set are generated. The label-feature residuals of shared and private pathways are represented as follows: ； in, For the sample tags The probability of belonging to a shared tag.

6. The OLTC composite fault diagnosis method with feature-tag collaborative decoupling according to claim 5, characterized in that, The steps for generating the final predicted labels for the original data samples by concatenating the predicted probabilities of the shared label set and the private label set, respectively, using a dual-label prediction model, specifically include: The shared features of each view are input into the shared classifier of the two-branch label prediction model. This yields the predicted probability for the shared tag set. ; Private features of each view Independent private classifiers in a two-branched label prediction model Make predictions for private tag sets. ; The prediction results of private features are weighted and fused to obtain: ； Based on the generated results of the shared and private label sets, the labels are concatenated along the label dimension to form the final predicted label for the sample.

7. The OLTC composite fault diagnosis method with feature-tag collaborative decoupling according to claim 6, characterized in that, The step of generating the final predicted label of the original data sample by concatenating the predicted probabilities of the shared label set and the private label set through the dual-label prediction model and then using binary cross-entropy loss to calculate the classification error of the shared label set and the private label set respectively.

8. A feature-tag collaborative decoupling OLTC composite fault diagnosis device, applied to perform the OLTC composite fault diagnosis method as described in any one of claims 1-7, characterized in that, Specifically, it includes: The data acquisition module is used to collect raw data samples from the multi-view of the on-load tap changer (OLTC) of the converter transformer. The feature extraction module is used to extract the shared and private features of each view from the original data sample through a constructed dual-feature encoder; The label generation module is used to construct a feature-label collaborative decoupling model based on shared features, private features, and labels. It divides labels into shared labels and private labels, and after aligning the shared labels and shared features, and the private labels and private features, it generates a shared label set and a private label set. The prediction output module is used to output the prediction probabilities of the shared label set and the private label set respectively through the dual-label prediction model. After concatenation along the label dimension, the final predicted label of the original data sample is generated, and the composite fault diagnosis result of OLTC is output.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method described in any one of claims 1 to 7.

Images