Power equipment output abnormal defect diagnosis method and device based on transfer learning, computer equipment, storage medium and computer program product

Abstract

The invention relates to a power equipment output abnormal defect diagnosis method and device based on transfer learning, computer equipment, a storage medium and a computer program product. The method comprises the steps of collecting source domain data and target domain data, and constructing a source domain pre-training model; obtaining a target domain adaptation model; performing abnormity judgment by calling the historical maintenance record of the to-be-diagnosed equipment to obtain a final diagnosis result; and updating the source domain pre-training model and the target domain adaptation model. By adopting the method, the problem of target domain data scarcity in power equipment fault diagnosis can be effectively solved by combining transfer learning with adversarial training and semi-supervised fine tuning, cross-equipment high-precision anomaly detection and defect classification are realized, meanwhile, a diagnosis result is optimized by utilizing historical maintenance records, and continuous iterative optimization of a model is supported.

Description

Technical Field

[0001] This application relates to the field of anomaly detection technology, and in particular to a method, apparatus, computer equipment, storage medium, and computer program product for diagnosing anomalies and defects in power equipment output based on transfer learning. Background Technology

[0002] With the continuous expansion of power system scale and the increasing complexity of equipment, the diagnosis of abnormal output defects in power equipment faces severe challenges. Traditional diagnostic methods rely heavily on human experience and threshold rules, making it difficult to cope with real-world problems such as equipment heterogeneity, variable operating conditions, and the scarcity of fault samples. Especially in the construction of new power systems, the high proportion of new energy integration leads to more complex equipment operating conditions, and fault characteristics exhibit strong nonlinearity, time-varying nature, and coupling.

[0003] Most current power equipment anomaly diagnosis methods rely on monitoring data from a single device to train models, typically requiring a large amount of labeled data for the target device. However, fault samples are scarce in actual industrial scenarios, leading to insufficient generalization ability of the models under small sample conditions. Secondly, existing methods generally use fixed thresholds or static models, making it difficult to adapt to the differences in feature distributions between different device models, resulting in a sharp decline in performance when migrating across devices. Thirdly, most solutions only use single-modal data or simple time-series models, failing to fully explore the spatiotemporal correlation features in equipment operation data. Fourthly, diagnostic results often lack linkage analysis with equipment maintenance history, resulting in a high false alarm rate. Finally, traditional systems typically employ closed architectures, making it impossible to continuously optimize the model with new cases, requiring retraining when faced with new defects. In comparison, these methods have significant shortcomings in cross-device adaptability, small sample learning ability, multimodal feature fusion, and system evolvability. Summary of the Invention

[0004] Therefore, it is necessary to provide a method, device, computer equipment, storage medium, and computer program product for diagnosing power equipment output anomalies and defects by combining transfer learning with adversarial training and semi-supervised fine-tuning. This method can effectively solve the problem of scarce target domain data in power equipment fault diagnosis, achieve high-precision anomaly detection and defect classification across devices, optimize diagnostic results by utilizing historical maintenance records, and support continuous iterative optimization of the model.

[0005] In a first aspect, this application provides a method for diagnosing abnormal output defects in power equipment based on transfer learning, the method comprising:

[0006] Collect source domain data and target domain data. The source domain data includes historical operating data of at least two power devices, and the target domain data includes real-time operating data of the power device to be diagnosed.

[0007] The source domain data and the target domain data are sequentially subjected to noise suppression, data alignment and normalization processes;

[0008] A basic model based on convolutional neural networks and long short-term memory networks is constructed, and the source domain model is trained based on the processed source domain data to obtain the source domain pre-trained model.

[0009] A domain discriminator module is added to the source domain pre-trained model. The distribution difference between the source domain and the target domain features is minimized through adversarial training. The maximum mean difference between the source domain and the target domain features is calculated in real time to obtain a domain adaptive transfer model.

[0010] The labeled data in the processed target domain data is selected as the target domain fine-tuning set, and the remaining unlabeled data is used as the semi-supervised auxiliary set. The semi-supervised auxiliary set is pseudo-labeled using the label propagation algorithm. After merging the labeled data and pseudo-label data, the domain adaptive transfer model is fine-tuned to obtain the target domain adaptation model.

[0011] The real-time standardized data of the power equipment to be diagnosed is input into the target domain adaptation model to obtain the anomaly probability value and defect type. The anomaly is judged by retrieving the historical maintenance records of the equipment to be diagnosed, and the final diagnosis result is obtained.

[0012] Collect diagnostic feedback data from the target domain adaptation model and new operational data from the devices to be diagnosed. Filter out cases with error rates greater than the error threshold and newly emerging defect types, and supplement them to the source domain dataset and / or the target domain dataset. Update the source domain pre-trained model and the target domain adaptation model.

[0013] In some embodiments of the method, the sequential execution of noise suppression, data alignment, and normalization processes on the source domain data and the target domain data includes:

[0014] A combination of wavelet thresholding and Kalman filtering is used. First, the source domain data and target domain data are decomposed into three levels using the db4 wavelet basis. The high-frequency wavelet coefficients are processed by an adaptive threshold based on Stein unbiased risk estimation, while the low-frequency coefficients remain unchanged. After wavelet reconstruction, initial denoising is performed.

[0015] Establish the state equation and observation equation for Kalman filtering, and recursively calculate and eliminate low-frequency sensor drift noise by setting the process noise covariance matrix and the observation noise covariance matrix.

[0016] Timestamp synchronization is performed using linear interpolation, and missing features are filled in using the mean value during feature dimension matching.

[0017] The min-max algorithm is used to map the aligned data to the target interval to obtain a standardized dataset.

[0018] In some embodiments of the method, the construction of a base model based on convolutional neural networks and long short-term memory networks, and the training of the source domain model based on the processed source domain data to obtain a source domain pre-trained model, includes:

[0019] The convolutional neural network includes three convolutional layers, three normalization layers, three ReLU activation function layers, and two max pooling layers. The convolutional neural network is used to extract local features from power equipment operation data.

[0020] Long Short-Term Memory (LSTM) networks consist of two layers of bidirectional LSM units and one output layer. The LSM networks are used to extract temporal correlation features from data.

[0021] The standardized source domain dataset is divided into source domain training set, source domain validation set and source domain test set according to the proportion. The basic model is pre-trained using AdamW optimizer and cross-entropy loss function.

[0022] By employing an early stopping strategy, training is stopped when the loss value of the source domain validation set does not decrease for several consecutive training cycles, thus obtaining the source domain pre-trained model.

[0023] In some embodiments of the method, the step of adding a domain discriminator module to the source domain pre-trained model, minimizing the distribution difference between the source and target domain features through adversarial training, and calculating the maximum mean difference between the source and target domain features in real time to obtain a domain adaptive transfer model includes:

[0024] A domain adversarial neural network structure is introduced into the source domain pre-trained model to build a domain discriminator consisting of three fully connected layers, and the output layer uses the Sigmoid activation function.

[0025] Determine the total loss function of the model, use cross-entropy loss for source domain classification loss, and adopt binary cross-entropy loss for domain adversarial loss;

[0026] The vector is concatenated based on the local features output by the convolutional neural network module and the temporal features output by the long short-term memory network module, and the maximum mean difference between the features of the source domain and the target domain is calculated using a Gaussian kernel function.

[0027] The domain adaptation weights are adjusted based on the magnitude of the maximum mean difference, and the basic model parameters and domain discriminator parameters are updated alternately. The maximum mean difference is calculated once every few batches in each iteration.

[0028] When the maximum mean difference is less than a preset threshold or the number of training iterations reaches a preset threshold, adversarial training stops, cross-domain feature transfer is completed, and a domain adaptive transfer model is obtained.

[0029] In some embodiments of the method, the step of selecting labeled data from the processed target domain data as the target domain fine-tuning set, and the remaining unlabeled data as the semi-supervised auxiliary set, performing pseudo-labeling on the semi-supervised auxiliary set using a label propagation algorithm, merging the labeled data and pseudo-label data, and then fine-tuning the domain adaptive transfer model to obtain the target domain adaptation model includes:

[0030] Annotated data containing normal labels and at least three types of abnormal labels are selected from the standardized target domain dataset as the target domain fine-tuning set, and the remaining unlabeled data are classified as the semi-supervised auxiliary set;

[0031] The cosine similarity between each unlabeled sample in the semi-supervised auxiliary set and the labeled samples in the fine-tuning set is calculated using the label propagation algorithm. The category with the largest sum of similarity weights is used as the pseudo-label of the unlabeled sample, thus completing the labeling of the semi-supervised auxiliary set.

[0032] Merge the target domain fine-tuning set with the semi-supervised auxiliary set after pseudo-labeling, fine-tune the input domain adaptive transfer model, set the number of training epochs, evaluate the accuracy on the target domain validation set divided from the fine-tuning set every certain number of training epochs, stop training if there is no improvement in accuracy for a certain number of consecutive training epochs, and obtain the target domain adapted model.

[0033] In some embodiments of the method, the step of inputting real-time standardized data of the power equipment to be diagnosed into the target domain adaptation model, obtaining anomaly probability values ​​and defect types, and performing anomaly judgment by retrieving historical maintenance records of the equipment to be diagnosed to obtain the final diagnostic result includes:

[0034] The real-time standardized data of the power equipment to be diagnosed is input into the target domain adaptation model, and the model outputs the anomaly probability value and the matched defect type.

[0035] Obtain the historical maintenance records of the device to be diagnosed, and extract data on maintenance time, type of replaced parts, and fault repair records.

[0036] Based on the combination of abnormal probability value and maintenance record, if the abnormal probability is within the target range and there is a record of replacing insulating components within a preset time, the probability is reduced by the first preset range. If the abnormal probability is within another range and there is no maintenance record in the recent period, the probability is increased by the second preset range. If the probability is less than the target value, it is considered normal. If it is greater than or equal to the target value, it is considered high risk and emergency maintenance is triggered.

[0037] Generate structured data results that include diagnosis time, device identification, post-correction anomaly probability, defect type, confidence level, correction basis, and maintenance recommendations.

[0038] According to a second aspect of the present disclosure, a power equipment output anomaly defect diagnosis device based on transfer learning is provided, the device comprising:

[0039] The first module is used to collect source domain data and target domain data. The source domain data includes historical operating data of at least two power devices, and the target domain data includes real-time operating data of the power device to be diagnosed.

[0040] The second module is used to sequentially perform noise suppression, data alignment and normalization processing on the source domain data and the target domain data;

[0041] The third module is used to build a basic model based on convolutional neural networks and long short-term memory networks. It trains the source domain model based on the processed source domain data to obtain the source domain pre-trained model.

[0042] The fourth module is used to add a domain discriminator module to the source domain pre-trained model. It minimizes the distribution difference of features between the source domain and the target domain through adversarial training, calculates the maximum mean difference of features between the source domain and the target domain in real time, and obtains a domain adaptive transfer model.

[0043] The fifth module is used to select labeled data from the processed target domain data as the target domain fine-tuning set, and the remaining unlabeled data as the semi-supervised auxiliary set. The semi-supervised auxiliary set is pseudo-labeled using the label propagation algorithm. After merging the labeled data and pseudo-label data, the domain adaptive transfer model is fine-tuned to obtain the target domain adaptation model.

[0044] The sixth module is used to input the real-time standardized data of the power equipment to be diagnosed into the target domain adaptation model, obtain the anomaly probability value and defect type, and make anomaly judgment by retrieving the historical maintenance records of the equipment to be diagnosed, and obtain the final diagnosis result.

[0045] The seventh module is used to collect diagnostic feedback data from the target domain adaptation model and new operational data from the devices to be diagnosed, filter out cases with error rates greater than the error threshold and newly emerging defect type cases, supplement them to the source domain dataset and / or the target domain dataset, and update the source domain pre-trained model and the target domain adaptation model.

[0046] According to a third aspect of the present disclosure, a computer device is provided. The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement a method for diagnosing abnormal defects in power equipment output based on transfer learning.

[0047] According to a fourth aspect of the present disclosure, a computer-readable storage medium is provided. The computer-readable storage medium stores a computer program thereon, which, when executed by a processor, implements a method for diagnosing abnormal defects in power equipment output based on transfer learning.

[0048] According to a fifth aspect of the present disclosure, a computer program product is provided. The computer program product includes a computer program that, when executed by a processor, implements a method for diagnosing abnormal defects in power equipment output based on transfer learning.

[0049] The power equipment output anomaly defect diagnosis method based on transfer learning provided in this application can effectively solve the model generalization problem under the condition of scarce target domain data by combining multi-source domain data pre-training with adversarial domain adaptation technology, and significantly improve the diagnostic accuracy in small sample scenarios. It employs a hybrid network to capture spatiotemporal features and combines maximum mean difference metric to achieve dynamic domain adaptation, enabling the model to automatically align with the feature distribution differences of different devices. A semi-supervised label propagation algorithm is introduced to complete target domain adaptation, significantly reducing annotation costs. Historical maintenance records are used to correct model output, and domain knowledge is integrated to improve diagnostic credibility. Furthermore, the system has continuous evolution capabilities, achieving iterative model optimization through error case feedback and new defect type identification. Compared with traditional methods, this scheme has outstanding advantages in cross-device transfer adaptability, small sample learning efficiency, and multi-defect identification accuracy, and is particularly suitable for industrial scenarios with diverse power equipment models and insufficient fault samples.

[0050] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0051] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure, and are not intended to unduly limit this disclosure.

[0052] Figure 1 This is a flowchart illustrating a method for diagnosing abnormal output defects in power equipment based on transfer learning, according to an exemplary embodiment.

[0053] Figure 2 This is a flowchart illustrating the steps of obtaining a domain adaptive migration model according to an exemplary embodiment;

[0054] Figure 3 This is a structural block diagram of a power equipment output anomaly defect diagnosis device based on transfer learning, according to an exemplary embodiment.

[0055] Figure 4 This is a diagram illustrating the internal structure of a computer device according to an exemplary embodiment. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0057] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe the order or sequence of objects. It should be understood that such data can be interchanged where appropriate so that the embodiments of this disclosure described herein can be implemented in orders other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure. The terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, product, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, product, or apparatus. Without further limitation, the presence of other identical or equivalent elements in a process, method, product, or apparatus that includes said elements is not excluded. For example, the use of terms such as "first," "second," etc., to denote names does not indicate any order of objects.

[0058] In some embodiments of this disclosure, a method for diagnosing abnormal output defects in power equipment based on transfer learning is provided, such as... Figure 1 As shown, it includes the following steps:

[0059] S20. Collect source domain data and target domain data. The source domain data includes historical operating data of at least two power devices, and the target domain data includes real-time operating data of the power device to be diagnosed.

[0060] Data is collected from both the source and target domains. The source domain data consists of historical operating data from at least two electrical devices of the same or similar models, while the target domain data consists of real-time operating data from the electrical device to be diagnosed. Both historical and real-time operating data include electrical parameters, environmental parameters, and operating status parameters. The source domain data includes labeled normal operation tags and at least five types of output anomaly / defect tags. The anomaly / defect tags include insulation aging, poor contact, component wear, electromagnetic interference, and load imbalance.

[0061] S21. Perform noise suppression, data alignment and normalization processing on the source domain data and the target domain data in sequence.

[0062] Noise suppression removes interfering signals from the data, preventing noise from affecting the accuracy of subsequent model training and diagnostic results. It can be combined with wavelet thresholding and Kalman filtering: first, the data is decomposed using a target wavelet basis to handle high-frequency interference; then, a filtering model is used to eliminate low-frequency noise such as sensor drift, ultimately obtaining cleaner, more effective data. Data alignment synchronizes timestamps from different devices through linear interpolation and fills in missing feature dimensions with the mean, ensuring consistent data structure. Normalization can use a target algorithm to map aligned data to a fixed interval, eliminating the influence of differences in units and numerical ranges of different parameters (such as voltage and temperature), making the data comparable.

[0063] S22. Construct a basic model based on convolutional neural networks and long short-term memory networks. Train the source domain model based on the processed source domain data to obtain the source domain pre-trained model.

[0064] The model structure consists of a convolutional neural network and a long short-term memory network, used to extract local features and temporal correlation features from the data, respectively. The training process uses source domain data pre-labeled with normal and / or abnormal tags, and adjusts parameters through an optimizer and loss function to ultimately obtain a model capable of initially identifying equipment defects.

[0065] S23. Add a domain discriminator module to the source domain pre-trained model, minimize the distribution difference between the source domain and target domain features through adversarial training, calculate the maximum mean difference between the source domain and target domain features in real time, and obtain the domain adaptive transfer model.

[0066] Domain adaptive transfer models address the data distribution mismatch problem by optimizing the model to reduce the difference between historical data and the data to be diagnosed. This can be achieved by introducing a domain adversarial neural network and adding a domain discriminator module. Through adversarial training, the model learns the common features of the source and target domains, thus reducing the distribution differences between them.

[0067] S24. Select the labeled data in the processed target domain data as the target domain fine-tuning set, and the remaining unlabeled data as the semi-supervised auxiliary set. Use the label propagation algorithm to perform pseudo-labeling on the semi-supervised auxiliary set. After merging the labeled data and pseudo-label data, fine-tune the domain adaptive transfer model to obtain the target domain adaptation model.

[0068] During training, a small amount of labeled data (fine-tuning set) from the device to be diagnosed and a large amount of unlabeled data (semi-supervised auxiliary set) can be combined. A label propagation algorithm is used to add pseudo-labels to the unlabeled data, thereby increasing the amount of training data. The merged dataset is then input into the domain adaptive transfer model for fine-tuning. Some irrelevant parameters are frozen, and the parameters of key layers are updated, ultimately resulting in a model that can accurately diagnose the defects of the device.

[0069] S25. Input the real-time standardized data of the power equipment to be diagnosed into the target domain adaptation model, obtain the anomaly probability value and defect type, and make anomaly judgment by retrieving the historical maintenance records of the equipment to be diagnosed, and obtain the final diagnosis result.

[0070] The final diagnostic result is a structured and actionable diagnostic conclusion formed by combining model output with historical equipment information. First, the target domain adaptation model outputs the anomaly probability and defect type of the equipment to be diagnosed. Then, the probability value is corrected by combining the equipment's maintenance records in recent years (such as component replacement and fault repair). Finally, a complete result is generated, which includes the diagnosis time, equipment identification, corrected probability, defect type, confidence level, correction basis, and operation and maintenance suggestions, directly guiding the operation and maintenance work.

[0071] S26. Collect diagnostic feedback data of the target domain adaptation model and newly added operating data of the device to be diagnosed, screen out cases with error rates greater than the error threshold and newly emerging defect type cases, supplement them to the source domain dataset and / or target domain dataset, and update the source domain pre-trained model and target domain adaptation model.

[0072] In some implementations, diagnostic feedback data from the target domain adaptation model and newly added operational data from the devices to be diagnosed can be collected. Cases with an error rate >5% and newly added defect cases not included in the initial labels can be filtered. The filtered data is supplemented according to domain attributes; cases of other devices of the same type are assigned to the source domain dataset, and cases of the devices to be diagnosed are assigned to the target domain dataset. The previous version of the target domain adaptation model is loaded, the parameters of the first two layers of the CNN module and the first layer of the LSTM module are frozen, and only the parameters of the third layer of the CNN module, the second layer of the LSTM module, and the fully connected layer are updated. The dynamic domain adaptation strategy and few-sample fine-tuning of parameters are used, and incremental training is performed with a set of training epochs. The performance of the model after iteration is evaluated using the newly added test set in the target domain. If the accuracy improves by ≥3% compared to the previous version and the F1 value for each defect type is ≥0.85, the original model is replaced. If the target domain is not met, the training parameters are adjusted and incremental training is performed again.

[0073] In some embodiments of this disclosure, S21 includes:

[0074] A combination of wavelet thresholding and Kalman filtering is used. First, the source domain data and target domain data are decomposed into three levels using the db4 wavelet basis. The high-frequency wavelet coefficients are processed by an adaptive threshold based on Stein unbiased risk estimation, while the low-frequency coefficients remain unchanged. After wavelet reconstruction, initial denoising is performed.

[0075] Establish the state equation and observation equation for Kalman filtering, and recursively calculate and eliminate low-frequency sensor drift noise by setting the process noise covariance matrix and the observation noise covariance matrix.

[0076] Timestamp synchronization is performed using linear interpolation, and missing features are filled in using the mean value during feature dimension matching.

[0077] The min-max algorithm is used to map the aligned data to the target interval to obtain a standardized dataset.

[0078] Specifically, threshold denoising using the db4 wavelet basis involves selecting and decomposing the wavelet basis for both the source and target domain data. The db4 wavelet basis is used as the decomposition basis, and three levels of wavelet decomposition are performed on each type of data. The decomposed data is split into high-frequency and low-frequency wavelet coefficients. For high-frequency coefficients, an adaptive thresholding method based on Stein's unbiased risk estimation is used. This method automatically calculates the optimal threshold based on the distribution characteristics of the high-frequency coefficients, filtering out high-frequency noise coefficients exceeding the threshold and retaining the effective signal components in the high-frequency part. For low-frequency wavelet coefficients, since the effective signal proportion is high, no processing is performed, and the original values ​​are directly retained. Preliminary wavelet reconstruction denoising involves recombining the processed high-frequency wavelet coefficients with the unprocessed low-frequency wavelet coefficients and performing wavelet reconstruction to obtain the source and target domain data after preliminary high-frequency noise removal.

[0079] In some embodiments of this disclosure, S22 includes:

[0080] The convolutional neural network includes three convolutional layers, three normalization layers, three ReLU activation function layers, and two max pooling layers. The convolutional neural network is used to extract local features from power equipment operation data.

[0081] Long Short-Term Memory (LSTM) networks consist of two layers of bidirectional LSM units and one output layer. The LSM networks are used to extract temporal correlation features from data.

[0082] The standardized source domain dataset is divided into source domain training set, source domain validation set and source domain test set according to the proportion. The basic model is pre-trained using AdamW optimizer and cross-entropy loss function.

[0083] By employing an early stopping strategy, training is stopped when the loss value of the source domain validation set does not decrease for several consecutive training cycles, thus obtaining the source domain pre-trained model.

[0084] Specifically, in source domain training, the source domain training set is input into the base model in batches, and the cross-entropy loss function is minimized through the AdamW optimizer. The formula for the cross-entropy loss function can be found in the following equation (1):

[0085]

[0086] In equation (1), This represents the batch sample size. For the number of categories, For the true label of sample i, Predict the probability that sample i belongs to class c for the model.

[0087] The CNN module is constructed by building three convolutional blocks with convolutional layers and supporting layers. Each convolutional block is connected in the order of convolutional layer, batch normalization layer, and ReLU activation function layer. Max pooling layers are added to the above convolutional blocks. Typically, one layer is added between the first and second convolutional blocks, and another between the second and third convolutional blocks, to compress the feature data dimension and retain key local features. After the output CNN features are processed by the three convolutional blocks and the two pooling layers, the output local feature data is flattened and used as the input of the subsequent LSTM module. The LSTM module is constructed by building a two-layer bidirectional LSTM unit. The first layer of the bidirectional LSTM receives the flattened feature sequence output by the CNN module and extracts temporal features from both forward and reverse temporal sequences. The second layer of the bidirectional LSTM further captures more complex long-term temporal correlation features based on the output of the first layer. A Dropout layer is added after the two-layer bidirectional LSTM unit to reduce the risk of overfitting and enhance the model's generalization ability. The complete basic model is formed by connecting the output of the LSTM module to the output layer and combining the CNN module and the LSTM module to form a complete CNN-LSTM basic model.

[0088] In some embodiments of this disclosure, reference is made to Figure 2 S23 includes:

[0089] S231. Introduce a domain adversarial neural network structure into the source domain pre-trained model, build a domain discriminator consisting of three fully connected layers, and use the Sigmoid activation function for the output layer.

[0090] S232. Determine the total loss function of the model. The source domain classification loss uses the cross-entropy loss, and the domain adversarial loss uses the binary cross-entropy loss.

[0091] S233. The vector is concatenated based on the local features output by the convolutional neural network module and the temporal features output by the long short-term memory network module, and the maximum mean difference between the features of the source domain and the target domain is calculated using the Gaussian kernel function.

[0092] S234. Adjust the domain adaptation weights based on the maximum mean difference, alternately update the basic model parameters and the domain discriminator parameters, and calculate the maximum mean difference once every few batches in each iteration.

[0093] S235. When the maximum mean difference is less than a preset threshold or the number of training iterations reaches a preset threshold, stop adversarial training, complete cross-domain feature transfer, and obtain a domain adaptive transfer model.

[0094] Specifically, the training objective of the domain discriminator is to distinguish whether the input features come from the source domain or the target domain. Its loss function adopts the binary cross-entropy loss, as shown in the following equation (2):

[0095]

[0096] In equation (2), The number of feature samples, For domain labels, source domain Target domain , The intermediate layer of the model outputs features. Let be the predicted probability of the domain discriminator. The total loss function of the model is... , The dynamic adjustment is based on the MMD value.

[0097] In some examples, the maximum mean difference of feature distribution is calculated, and local and temporal features are concatenated. Two types of core features are extracted from the base model: local features output by the CNN module and temporal correlation features output by the LSTM module. The two types of features are concatenated dimensionally to form a feature concatenation vector. The MMD is calculated based on the Gaussian kernel function. The Gaussian kernel function is selected as the kernel for calculating the MMD, and appropriate kernel parameters are set according to the distribution characteristics of the feature data. The inter-domain feature difference is calculated by inputting the source domain feature concatenation vector and the target domain feature concatenation vector respectively, and calculating the MMD value of the two through the Gaussian kernel function. The larger the MMD value, the greater the difference in feature distribution between the source domain and the target domain, and vice versa.

[0098] Based on MMD, the domain adaptation weights are adjusted. The MMD calculation frequency is set to calculate the current MMD value once every 50 training batches. The adversarial loss weights are dynamically adjusted according to the MMD value, adjusting the weight ratio of domain adversarial loss in the total loss: if the MMD value is large, the weight of adversarial loss is increased to strengthen the learning of feature distribution convergence by the base model; if the MMD value is small, the weight of adversarial loss is decreased to avoid excessive adversarialism leading to a decline in source domain task performance. The model parameters are alternately updated. The parameters of the base model are fixed by updating the parameters of the domain discriminator to minimize the domain adversarial loss, and the weights and biases of the fully connected layers of the domain discriminator are updated to allow the discriminator to more accurately distinguish the source of features. The parameters of the base model are fixed by updating the parameters of the domain discriminator to minimize the total loss, and the parameters of the base model are updated to make the source and target domain features output by the base model more difficult for the discriminator to distinguish, realizing inter-domain feature transfer. The process of alternatingly updating the discriminator and updating the base model is repeated until the training stopping condition is met.

[0099] In some embodiments of this disclosure, S24 includes:

[0100] Annotated data containing normal labels and at least three types of abnormal labels are selected from the standardized target domain dataset as the target domain fine-tuning set, and the remaining unlabeled data are classified as the semi-supervised auxiliary set;

[0101] The cosine similarity between each unlabeled sample in the semi-supervised auxiliary set and the labeled samples in the fine-tuning set is calculated using the label propagation algorithm. The category with the largest sum of similarity weights is used as the pseudo-label of the unlabeled sample, thus completing the labeling of the semi-supervised auxiliary set.

[0102] Merge the target domain fine-tuning set with the semi-supervised auxiliary set after pseudo-labeling, fine-tune the input domain adaptive transfer model, set the number of training epochs, evaluate the accuracy on the target domain validation set divided from the fine-tuning set every certain number of training epochs, stop training if there is no improvement in accuracy for a certain number of consecutive training epochs, and obtain the target domain adapted model.

[0103] Specifically, the core feature vectors of the labeled samples in the target domain fine-tuning set and the unlabeled samples in the semi-supervised auxiliary set are extracted. Similarity is calculated sample-by-sample. For each unlabeled sample in the semi-supervised auxiliary set, cosine similarity is calculated with all labeled samples in the target domain fine-tuning set. The closer the cosine similarity value is to 1, the more similar the features of the two samples are; the closer it is to 0, the greater the feature difference. A weighted sum is calculated by category. For each unlabeled sample, the labeled samples in the fine-tuning set are first grouped by label category. Then, for each category, the cosine similarity between the unlabeled sample and all labeled samples in that category is summed to obtain the total similarity weight for each category. Pseudo-labels are assigned. The total similarity weight for the unlabeled sample across all categories is compared, and the category with the largest weight is selected as the pseudo-label for the unlabeled sample. The auxiliary set labeling process is repeated for all unlabeled samples in the semi-supervised auxiliary set, resulting in a pseudo-labeled semi-supervised auxiliary set.

[0104] In some embodiments of this disclosure, S25 includes:

[0105] The real-time standardized data of the power equipment to be diagnosed is input into the target domain adaptation model, and the model outputs the anomaly probability value and the matched defect type.

[0106] Obtain the historical maintenance records of the device to be diagnosed, and extract data on maintenance time, type of replaced parts, and fault repair records.

[0107] Based on the combination of abnormal probability value and maintenance record, if the abnormal probability is within the target range and there is a record of replacing insulating components within a preset time, the probability is reduced by the first preset range. If the abnormal probability is within another range and there is no maintenance record in the recent period, the probability is increased by the second preset range. If the probability is less than the target value, it is considered normal. If it is greater than or equal to the target value, it is considered high risk and emergency maintenance is triggered.

[0108] Generate structured data results that include diagnosis time, device identification, post-correction anomaly probability, defect type, confidence level, correction basis, and maintenance recommendations.

[0109] In some examples, historical maintenance records for the equipment under diagnosis over the past three years can be obtained, extracting maintenance time, types of replaced parts, and fault repair records. Based on a combination of anomaly probability values ​​and maintenance records, if the anomaly probability is 0.6-0.8 and there are records of insulation component replacement within the past six months, the probability is lowered to 0.15-0.2; if the anomaly probability is 0.5-0.7 and there are no maintenance records within the past year, the probability is increased to 0.1-0.15; if the probability is less than 0.5, it is directly judged as normal; if it is greater than or equal to 0.8, it is judged as high risk and emergency maintenance is triggered. Structured data results are generated, including diagnosis time, equipment ID, corrected anomaly probability, defect type, confidence level, correction basis, and maintenance recommendations.

[0110] Specifically, the model loads all parameters of the target domain adaptation model currently in use, including the weights and biases of the CNN module, LSTM module, and domain discriminator, ensuring that incremental training continues to optimize based on existing effective features and avoiding training from scratch. The general feature extraction layer freezes the parameters of the first two layers of the CNN module and the first layer of the LSTM module: these two layers are mainly responsible for extracting general local features and basic temporal features of the device operation data. These features are highly universal when the device type remains unchanged and do not require repeated updates; freezing them reduces the risk of overfitting. The target optimization layer only opens the parameters of the third layer of the CNN module, the second layer of the LSTM module, and the fully connected layer: the third layer of the CNN is responsible for extracting detailed local features, and the second layer of the LSTM is responsible for capturing complex temporal correlations; the fully connected layer directly affects defect type judgment and anomaly probability calculation. Opening these layers allows the model to adjust key decision parameters for new cases and new defects.

[0111] The dynamic domain adaptation strategy is retained and continues to be used—the weights of the domain adversarial loss are dynamically adjusted by calculating the maximum mean difference between the source domain supplementary data and the target domain supplementary data, ensuring that the model maintains the consistency of the distribution of features between the source and target domains during the optimization process; the parameters are fine-tuned using small samples: a small training batch size and a low learning rate are set to ensure that the model learns new data features slowly and stably; the number of incremental training epochs is set according to the amount of incremental data to ensure that the model can fully learn the features of new cases without over-iteration; incremental training is performed by inputting the source domain supplementary data and the target domain supplementary training data into the model and starting incremental training: the model only updates the parameters of the open layers, and optimizes the ability to judge high error rate defects and newly added defects by minimizing the classification loss and the domain adversarial loss.

[0112] This disclosure presents several methods for diagnosing abnormal output defects in power equipment based on transfer learning. These methods effectively address the model generalization problem under conditions of scarce target domain data by combining multi-source domain data pre-training with adversarial domain adaptation techniques, significantly improving diagnostic accuracy in small-sample scenarios. A hybrid network is employed to capture spatiotemporal features, combined with a maximum mean difference metric to achieve dynamic domain adaptation, enabling the model to automatically align with the feature distribution differences of different devices. A semi-supervised label propagation algorithm is introduced to complete target domain adaptation, significantly reducing annotation costs. Historical maintenance records are used to correct model output, and domain knowledge is integrated to enhance diagnostic credibility. Furthermore, the system possesses continuous evolution capabilities, achieving iterative model optimization through error case feedback and new defect type identification. Compared to traditional methods, this approach exhibits significant advantages in cross-device transfer adaptability, small-sample learning efficiency, and multi-defect identification accuracy, making it particularly suitable for industrial scenarios with diverse power equipment models and insufficient fault samples.

[0113] It is understood that the various embodiments of the methods described in this specification are presented in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. Related details can be found in the descriptions of other method embodiments.

[0114] It should be understood that although the steps in the flowcharts shown in the accompanying drawings are displayed sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some of the steps in the accompanying drawings may include multiple steps or stages, which are not necessarily completed at the same time, but may be executed at different times, and the execution order of these steps or stages is not necessarily sequential, but may be performed alternately or in turn with other steps or at least a portion of the steps or stages of other steps.

[0115] Based on the description of the above-described embodiments of the power equipment output anomaly defect diagnosis method based on transfer learning, this disclosure also provides a power equipment output anomaly defect diagnosis device based on transfer learning for implementing the aforementioned power equipment output anomaly defect diagnosis method based on transfer learning. The device may include a system (including a distributed system), software (application), module, component, controller, server, terminal, etc., using the method described in the embodiments of this specification, combined with necessary hardware implementation. Based on the same innovative concept, the devices in one or more embodiments provided in this disclosure are as described in the following embodiments. Since the implementation schemes and methods for solving the problem by the device are similar, the implementation of the specific device in the embodiments of this specification can refer to the implementation of the aforementioned method, and repeated details will not be repeated. As used below, the terms "unit" or "module" can refer to a combination of software and / or hardware that implements a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0116] Figure 3 This is a schematic block diagram illustrating a power equipment output anomaly defect diagnosis device based on transfer learning, according to an exemplary embodiment. The device can be the aforementioned terminal, a server, or a module, component, device, control unit, etc., integrated into the terminal. For details, please refer to... Figure 3The device 100 may include: a first module 110, a second module 120, a third module 130, a fourth module 140, a fifth module 150, a sixth module 160, and a seventh module 170. Specifically, the first module 110 is used to collect source domain data and target domain data, wherein the source domain data includes historical operating data of at least two power devices, and the target domain data includes real-time operating data of the power device to be diagnosed; the second module 120 is used to sequentially perform noise suppression, data alignment, and normalization processing on the source domain data and the target domain data; the third module 130 is used to construct a basic model based on convolutional neural networks and long short-term memory networks, and train the source domain model based on the processed source domain data to obtain a source domain pre-trained model; the fourth module 140 is used to add a domain discriminator module to the source domain pre-trained model, minimize the distribution difference of source and target domain features through adversarial training, and calculate the maximum mean difference of source and target domain features in real time to obtain a domain adaptive transfer model; the fifth module 150 is used to select the processed target... The labeled data in the domain data serves as the target domain fine-tuning set, while the remaining unlabeled data serves as the semi-supervised auxiliary set. A pseudo-labeling algorithm is used to label the semi-supervised auxiliary set. After merging the labeled and pseudo-labeled data, the domain adaptive transfer model is fine-tuned to obtain the target domain adaptation model. Module 160 is used to input real-time standardized data of the power equipment to be diagnosed into the target domain adaptation model to obtain anomaly probability values ​​and defect types. Anomaly judgment is performed by retrieving historical maintenance records of the equipment to be diagnosed, and the final diagnostic result is obtained. Module 170 is used to collect diagnostic feedback data from the target domain adaptation model and newly added operational data of the equipment to be diagnosed. Cases with error rates exceeding the error threshold and newly emerging defect type cases are selected and added to the source domain dataset and / or target domain dataset to update the source domain pre-trained model and the target domain adaptation model.

[0117] In some embodiments of the device, the second module 120 is further configured to employ a combination of wavelet threshold denoising and Kalman filtering. First, the source and target domain data are decomposed into three levels using a db4 wavelet basis. High-frequency wavelet coefficients are processed with an adaptive threshold based on Stein's unbiased risk estimation, while low-frequency coefficients remain unchanged. Preliminary denoising is achieved after wavelet reconstruction. The state equation and observation equation for Kalman filtering are established. Low-frequency sensor drift noise is eliminated by recursively calculating the process noise covariance matrix and observation noise covariance matrix. Timestamp synchronization is performed using linear interpolation. When matching feature dimensions, missing features are filled in using the mean. The min-max algorithm is used to map the aligned data to the target interval to obtain a standardized dataset.

[0118] In some embodiments of the device, the third module 130 is further configured to include a convolutional neural network comprising three convolutional layers, three normalization layers, three ReLU activation function layers, and two max pooling layers, wherein the convolutional neural network is used to extract local features from power equipment operation data; a long short-term memory network comprising two bidirectional long short-term memory units and one output layer, wherein the long short-term memory network is used to extract temporal correlation features from the data; the standardized source domain dataset is proportionally divided into a source domain training set, a source domain validation set, and a source domain test set, and the basic model is pre-trained using the AdamW optimizer and the cross-entropy loss function; and training is stopped when the loss value of the source domain validation set has not decreased for several consecutive training cycles using an early stopping strategy, thereby obtaining the source domain pre-trained model.

[0119] In some embodiments of the device, the fourth module 140 is further used to introduce a domain adversarial neural network structure into the source domain pre-trained model, build a domain discriminator composed of three fully connected layers, and use the Sigmoid activation function for the output layer; determine the total loss function of the model, with the source domain classification loss using cross-entropy loss and the domain adversarial loss using binary cross-entropy loss; concatenate the vectors of the local features output by the convolutional neural network module and the temporal features output by the long short-term memory network module, and calculate the maximum mean difference between the features of the source domain and the target domain using a Gaussian kernel function; adjust the domain adaptation weights based on the magnitude of the maximum mean difference, alternately update the basic model parameters and the domain discriminator parameters, and calculate the maximum mean difference once every several batches of iterations; when the maximum mean difference is less than a preset threshold or the number of training iterations reaches a preset threshold, stop the adversarial training, complete the cross-domain feature transfer, and obtain the domain adaptive transfer model.

[0120] In some embodiments of the device, the fifth module 150 is further configured to select labeled data containing normal labels and at least three types of abnormal labels from the standardized target domain dataset as the target domain fine-tuning set, and classify the remaining unlabeled data as a semi-supervised auxiliary set; calculate the cosine similarity between each unlabeled sample in the semi-supervised auxiliary set and the labeled sample in the fine-tuning set through the label propagation algorithm, and take the category with the largest sum of similarity weights as the pseudo-label of the unlabeled sample to complete the labeling of the semi-supervised auxiliary set; merge the target domain fine-tuning set and the semi-supervised auxiliary set after pseudo-label labeling, input the domain adaptive transfer model for fine-tuning, set the number of training cycles, evaluate the accuracy on a partial target domain validation set divided from the fine-tuning set every few training cycles, stop training if the accuracy does not improve for several consecutive training cycles, and obtain the target domain adapted model.

[0121] In some embodiments of the device, the sixth module 160 is further configured to input real-time standardized data of the power equipment to be diagnosed into the target domain adaptation model, and the model outputs anomaly probability values ​​and matching defect types; acquire historical maintenance records of the equipment to be diagnosed, and extract maintenance time, replacement component type, and fault repair record data; based on the combination of anomaly probability values ​​and maintenance records, if the anomaly probability is within the target range and there are insulation component replacement records within a preset time, the probability is lowered by a first preset range; if the anomaly probability is within another range and there are no maintenance records in the recent period, the probability is increased by a second preset range; if the probability is less than the target value, it is considered normal; if it is greater than or equal to the target value, it is considered high risk and emergency maintenance is triggered; generate structured data results including diagnosis time, equipment identification, corrected anomaly probability, defect type, confidence level, correction basis, and maintenance suggestions.

[0122] Each module in the aforementioned power equipment output anomaly defect diagnosis device based on transfer learning can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the computer device's memory as software, so that the processor can call and execute the corresponding operations of each module.

[0123] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 4 As shown, the computer device includes a processor, memory, and a network interface connected via a system bus. The processor provides computing 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 an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements a method for diagnosing abnormal output defects in power equipment based on transfer learning.

[0124] Those skilled in the art will understand that Figure 4 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 device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0125] Based on the foregoing description of the relevant methods and apparatus embodiments, this disclosure also provides a computer device, including a memory and a processor. The memory stores a computer program, which, when executed by the processor, implements the power equipment output anomaly defect diagnosis method based on transfer learning as described in any embodiment of this specification.

[0126] Based on the foregoing description of the relevant methods and apparatus embodiments, this disclosure also provides a computer-readable storage medium that, when the instructions in the computer-readable storage medium are executed by the processor of a computer device, enables the computer device to implement the power equipment output anomaly defect diagnosis method based on transfer learning as described in any embodiment of this disclosure.

[0127] Based on the foregoing description of the relevant methods and apparatus embodiments, this disclosure also provides a computer program product, including a computer program that, when executed by a processor, implements the power equipment output anomaly defect diagnosis method based on transfer learning as described in any embodiment of this specification.

[0128] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on its differences from other embodiments. In particular, hardware + program embodiments are relatively simple in description because they are fundamentally similar to method embodiments; relevant parts can be referred to the descriptions in the method embodiments.

[0129] 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 (ReRAM), 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). 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.

[0130] It should be noted that the apparatus, computer equipment, storage medium, and computer program products described above may also include other implementation methods according to the description of the method embodiments. Specific implementation methods can be found in the description of the relevant method embodiments. Furthermore, new embodiments formed by combinations of features from various methods, apparatuses, devices, and server embodiments still fall within the scope of this disclosure and will not be elaborated upon here.

[0131] For ease of description, the above devices are described in terms of function, divided into various modules. Of course, when implementing one or more of these specifications, the functions of each module can be implemented in the same or different software and / or hardware, or a module that performs the same function can be implemented by a combination of multiple sub-modules or sub-units. The device embodiments described above are merely illustrative. For example, the division of modules or units is only a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling and communication connections between the devices or units shown or described can be implemented through direct and / or indirect coupling / connection, through standard or custom interfaces or protocols, and can be implemented electrically, mechanically, or in other forms.

[0132] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This disclosure is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the following claims.

[0133] It should be understood that this disclosure is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope.

Claims

1. A method for diagnosing abnormal output defects in power equipment based on transfer learning, characterized in that, The method includes: Collect source domain data and target domain data. The source domain data includes historical operating data of at least two power devices, and the target domain data includes real-time operating data of the power device to be diagnosed. The source domain data and the target domain data are sequentially subjected to noise suppression, data alignment and normalization processes; A basic model based on convolutional neural networks and long short-term memory networks is constructed, and the source domain model is trained based on the processed source domain data to obtain the source domain pre-trained model. A domain discriminator module is added to the source domain pre-trained model. The distribution difference between the source domain and the target domain features is minimized through adversarial training. The maximum mean difference between the source domain and the target domain features is calculated in real time to obtain a domain adaptive transfer model. The labeled data in the processed target domain data is selected as the target domain fine-tuning set, and the remaining unlabeled data is used as the semi-supervised auxiliary set. The semi-supervised auxiliary set is pseudo-labeled using the label propagation algorithm. After merging the labeled data and pseudo-label data, the domain adaptive transfer model is fine-tuned to obtain the target domain adaptation model. The real-time standardized data of the power equipment to be diagnosed is input into the target domain adaptation model to obtain the anomaly probability value and defect type. The anomaly is judged by retrieving the historical maintenance records of the equipment to be diagnosed, and the final diagnosis result is obtained. Collect diagnostic feedback data from the target domain adaptation model and new operational data from the devices to be diagnosed. Filter out cases with error rates greater than the error threshold and newly emerging defect types, and supplement them to the source domain dataset and / or the target domain dataset. Update the source domain pre-trained model and the target domain adaptation model.

2. The method according to claim 1, characterized in that, The step of sequentially performing noise suppression, data alignment, and normalization processing on the source domain data and the target domain data includes: A combination of wavelet thresholding and Kalman filtering is used. First, the source domain data and target domain data are decomposed into three levels using the db4 wavelet basis. The high-frequency wavelet coefficients are processed by an adaptive threshold based on Stein unbiased risk estimation, while the low-frequency coefficients remain unchanged. After wavelet reconstruction, initial denoising is performed. Establish the state equation and observation equation for Kalman filtering, and recursively calculate and eliminate low-frequency sensor drift noise by setting the process noise covariance matrix and the observation noise covariance matrix. Timestamp synchronization is performed using linear interpolation, and missing features are filled in using the mean value during feature dimension matching. The min-max algorithm is used to map the aligned data to the target interval to obtain a standardized dataset.

3. The method according to claim 1, characterized in that, The construction of a basic model based on convolutional neural networks and long short-term memory networks, and the training of the source domain model based on the processed source domain data, to obtain a source domain pre-trained model, includes: The convolutional neural network includes three convolutional layers, three normalization layers, three ReLU activation function layers, and two max pooling layers. The convolutional neural network is used to extract local features from power equipment operation data. Long Short-Term Memory (LSTM) networks consist of two layers of bidirectional LSM units and one output layer. The LSM networks are used to extract temporal correlation features from data. The standardized source domain dataset is divided into source domain training set, source domain validation set and source domain test set according to the proportion. The basic model is pre-trained using AdamW optimizer and cross-entropy loss function. By employing an early stopping strategy, training is stopped when the loss value of the source domain validation set does not decrease for several consecutive training cycles, thus obtaining the source domain pre-trained model.

4. The method according to claim 1, characterized in that, The process of adding a domain discriminator module to the source domain pre-trained model, minimizing the distribution difference between the source and target domain features through adversarial training, and calculating the maximum mean difference between the source and target domain features in real time, yields a domain adaptive transfer model, including: A domain adversarial neural network structure is introduced into the source domain pre-trained model to build a domain discriminator consisting of three fully connected layers, and the output layer uses the Sigmoid activation function. Determine the total loss function of the model, use cross-entropy loss for source domain classification loss, and adopt binary cross-entropy loss for domain adversarial loss; The vector is concatenated based on the local features output by the convolutional neural network module and the temporal features output by the long short-term memory network module, and the maximum mean difference between the features of the source domain and the target domain is calculated using a Gaussian kernel function. The domain adaptation weights are adjusted based on the magnitude of the maximum mean difference, and the basic model parameters and domain discriminator parameters are updated alternately. The maximum mean difference is calculated once every few batches in each iteration. When the maximum mean difference is less than a preset threshold or the number of training iterations reaches a preset threshold, adversarial training stops, cross-domain feature transfer is completed, and a domain adaptive transfer model is obtained.

5. The method according to claim 1, characterized in that, The process involves selecting labeled data from the processed target domain data as the target domain fine-tuning set, and the remaining unlabeled data as the semi-supervised auxiliary set. A pseudo-labeling algorithm is used to label the semi-supervised auxiliary set. The labeled data and pseudo-labeled data are then merged to fine-tune the domain adaptive transfer model, resulting in a target domain-fitting model. This includes: Annotated data containing normal labels and at least three types of abnormal labels are selected from the standardized target domain dataset as the target domain fine-tuning set, and the remaining unlabeled data are classified as the semi-supervised auxiliary set; The cosine similarity between each unlabeled sample in the semi-supervised auxiliary set and the labeled samples in the fine-tuning set is calculated using the label propagation algorithm. The category with the largest sum of similarity weights is used as the pseudo-label of the unlabeled sample, thus completing the labeling of the semi-supervised auxiliary set. Merge the target domain fine-tuning set with the semi-supervised auxiliary set after pseudo-labeling, fine-tune the input domain adaptive transfer model, set the number of training epochs, evaluate the accuracy on the target domain validation set divided from the fine-tuning set every certain number of training epochs, stop training if there is no improvement in accuracy for a certain number of consecutive training epochs, and obtain the target domain adapted model.

6. The method according to claim 1, characterized in that, The process involves inputting real-time standardized data of the power equipment to be diagnosed into the target domain adaptation model to obtain anomaly probability values ​​and defect types. Anomaly detection is performed by retrieving historical maintenance records of the equipment to be diagnosed, and the final diagnostic result is obtained, including: The real-time standardized data of the power equipment to be diagnosed is input into the target domain adaptation model, and the model outputs the anomaly probability value and the matched defect type. Obtain the historical maintenance records of the device to be diagnosed, and extract data on maintenance time, type of replaced parts, and fault repair records. Based on the combination of abnormal probability value and maintenance record, if the abnormal probability is within the target range and there is a record of replacing insulating components within a preset time, the probability is reduced by the first preset range. If the abnormal probability is within another range and there is no maintenance record in the recent period, the probability is increased by the second preset range. If the probability is less than the target value, it is considered normal. If it is greater than or equal to the target value, it is considered high risk and emergency maintenance is triggered. Generate structured data results that include diagnosis time, device identification, post-correction anomaly probability, defect type, confidence level, correction basis, and maintenance recommendations.

7. A power equipment output anomaly defect diagnosis device based on transfer learning, characterized in that, The device includes: The first module is used to collect source domain data and target domain data. The source domain data includes historical operating data of at least two power devices, and the target domain data includes real-time operating data of the power device to be diagnosed. The second module is used to sequentially perform noise suppression, data alignment and normalization processing on the source domain data and the target domain data; The third module is used to build a basic model based on convolutional neural networks and long short-term memory networks. It trains the source domain model based on the processed source domain data to obtain the source domain pre-trained model. The fourth module is used to add a domain discriminator module to the source domain pre-trained model. It minimizes the distribution difference of features between the source domain and the target domain through adversarial training, calculates the maximum mean difference of features between the source domain and the target domain in real time, and obtains a domain adaptive transfer model. The fifth module is used to select labeled data from the processed target domain data as the target domain fine-tuning set, and the remaining unlabeled data as the semi-supervised auxiliary set. The semi-supervised auxiliary set is pseudo-labeled using the label propagation algorithm. After merging the labeled data and pseudo-label data, the domain adaptive transfer model is fine-tuned to obtain the target domain adaptation model. The sixth module is used to input the real-time standardized data of the power equipment to be diagnosed into the target domain adaptation model, obtain the anomaly probability value and defect type, and make anomaly judgment by retrieving the historical maintenance records of the equipment to be diagnosed, and obtain the final diagnosis result. The seventh module is used to collect diagnostic feedback data from the target domain adaptation model and new operational data from the devices to be diagnosed, filter out cases with error rates greater than the error threshold and newly emerging defect type cases, supplement them to the source domain dataset and / or the target domain dataset, and update the source domain pre-trained model and the target domain adaptation model.

8. A computer device, characterized in that, It includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, It stores a computer program thereon, which, when executed by a processor, implements the steps of the method according to any one of claims 1 to 6.

10. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, implements the steps of the method according to any one of claims 1 to 6.

Images