A method for insulation fault diagnosis of power equipment based on unsupervised domain adaptation
By employing an unsupervised domain adaptive method and utilizing cluster prototype vector matching and cluster contrastive learning mechanisms, the adaptive problem of power equipment insulation fault diagnosis models in unknown scenarios is solved, improving the model's accuracy and robustness, and achieving effective transfer of fine-grained class discriminative features.
Patent Information
- Application Number
- CN202410360755.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-03-27
- Publication Date
- 2024-11-08
- Estimated Expiration
- 2044-03-27
AI Technical Summary
Existing methods for diagnosing insulation faults in power equipment cannot adaptively update model parameters in unknown scenarios, resulting in poor model generalization performance. They also cannot extract class discrimination features from samples with missing label information in unknown scenarios, and their transfer learning strategies are coarse, affecting the robustness of the model.
We employ a pseudo-label generator based on cluster prototype vector matching, a cluster contrastive learning mechanism, and a cluster subdomain feature matcher. Through unsupervised domain adaptation, we update model parameters using unlabeled data from unseen scenarios, thereby enhancing the model's robustness and generalization ability in new scenarios.
It improves the accuracy and robustness of the power equipment insulation fault diagnosis model in practical applications, effectively transfers fine-grained class discrimination information, and enhances the model's diagnostic capabilities in unknown scenarios.
Smart Images

Figure CN118445676B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of partial discharge fault diagnosis of electric equipment, and specifically to an insulation fault diagnosis method of electric equipment based on unsupervised domain adaptation. Background Art
[0002] As an important part of the power grid, the safe and stable operation of power equipment directly affects the reliability of power transmission. Among them, insulation performance is a key indicator affecting the operational reliability of power equipment. However, typical insulation defects often occur during the production, transportation, installation, mechanical operation and other processes, which in turn induce different forms of partial discharge (Partial Charge, PD) inside the equipment. Research by the International Conference on Large Electric Systems working group shows that different forms of partial discharge have different effects on the insulation performance of damaged equipment. Therefore, accurately diagnosing the types of insulation defects in power equipment and identifying different forms of partial discharge are of great significance to ensuring the safe operation of power equipment and improving the stability of the power grid.
[0003] With the on-site deployment of PD monitoring sensors and the development of intelligent algorithms, data-driven intelligent diagnosis methods for insulation defects have brought new breakthroughs in improving the perception of the internal insulation status of power equipment. However, affected by different insulation severity and working conditions, the potential assumptions of the intelligent model and the consistency of the distribution of training and test set features are difficult to meet in actual application scenarios. In addition, the lack of data label information in unknown scenarios brings new challenges to the intelligent diagnosis model's adaptive adaptation to new scenario data features. Y. Wang et al. used domain adversarial transfer technology to achieve the migration of the insulation defect diagnosis model trained with simulation data to actual application scenarios (A Small Sample GIS Insulation Defect Intelligent Diagnosis Method Based on Domain Adversarial Transfer Convolutional Neural Network Wang Yanxin Yan Jing Wang Jianhua Geng Yingsan Liu Zhiyuan National Key Laboratory of Electrical Insulation of Power Equipment (Xi'an Jiaotong University) Xi'an 710049), but this migration process requires accurate and reliable label information to be provided in the actual scenario.
[0004] Based on this, the existing technology has the following technical problems: (1) Many existing methods cannot adaptively update the parameters of the intelligent diagnosis model according to the data distribution characteristics in the unknown scene, and thus can only be applied to scenes with similar data distribution characteristics to the training set, resulting in poor model generalization performance; (2) The existing intelligent diagnosis model cannot extract the class discrimination feature information of samples with missing label information in the unknown scene; (3) In the existing transfer learning strategy, the feature alignment strategy is too coarse, resulting in the model being unable to perceive fine-grained class discrimination features, making it impossible to effectively transfer class discrimination information from the source domain to the target domain, affecting the robustness of the model. Summary of the invention
[0005] In order to overcome the shortcomings and deficiencies of the prior art, the present invention proposes a method for diagnosing insulation faults of power equipment based on unsupervised domain adaptation, including a pseudo-label generator based on cluster prototype vector matching, a cluster contrast learning mechanism, and a cluster subdomain feature matcher, which can better solve the problem of difficulty in obtaining data labels in actual unseen scenarios and the damage to the robustness of the model caused by the difference between the feature distribution of the sample set and the training set, and comprehensively improve the accuracy and robustness of the intelligent diagnosis model for insulation defects in practical applications. The present invention can directly update the model parameters using PD data collected in new scenarios that do not contain label information, so that the intelligent diagnosis model can adapt to new scenarios and improve the robustness and generalization ability of the model.
[0006] The present invention is achieved by at least one of the following technical solutions.
[0007] A method for diagnosing insulation faults of power equipment based on unsupervised domain adaptation comprises the following steps:
[0008] S1. Collect UHF data of partial discharge of power equipment;
[0009] S2. Build an insulation defect diagnosis model based on deep learning, including a time series feature preprocessor, a feature extractor, and a classifier;
[0010] S3, using a dataset containing label information in a known scenario as a source domain, and pre-training an insulation fault diagnosis model based on it, and retaining the model parameters;
[0011] S4, constructing a target domain pseudo label generator matching the cluster prototype vector in the high-dimensional feature space output by the feature extractor;
[0012] S5, taking the unlabeled dataset in the unseen scene as the target domain, and using the unsupervised clustering algorithm to construct cluster subdomains of the source domain and the target domain in the feature space;
[0013] S6. Taking the cluster prototype vector as the anchor point and the similar samples in the cluster subdomain as the positive sample pairs, a cluster contrast learning mechanism is constructed to strengthen the machine granularity of the discriminable features in the domain;
[0014] S7. Use Gaussian radial basis function to map the cluster subdomain to high-dimensional Hilbert space, construct a cluster subdomain feature matcher, align the cluster subdomain features in the source domain and the target domain, and realize the migration of class-discriminable feature information from the source domain to the target domain;
[0015] S8, combining cluster contrast learning mechanism and cluster subdomain feature matcher to update feature extractor parameters;
[0016] S9. Use the source domain dataset containing label information to fine-tune the classifier parameters to achieve insulation defect fault diagnosis.
[0017] Furthermore, in step S1, the insulation defect types of the power equipment include foreign matter attachment, metal tips, metal particles and insulating air gaps. The insulation defects of the power equipment induce local discharge inside the equipment, and a corresponding discharge signal is collected using a UHF sensor.
[0018] Furthermore, four types of partial discharge ultra-high frequency signal data under different working conditions are collected, and different data sets are constructed according to the differences in working conditions; in the data sets, N containing label information are S The dataset of samples is used as the source domain x i represents the i-th partial discharge UHF signal in the source domain, y i Indicates the insulation fault type corresponding to the i-th sample; N without label information T The target domain is a dataset of samples x j Represents the jth partial discharge UHF signal in the target domain.
[0019] Furthermore, in step S2, an insulation defect diagnosis model based on deep learning is constructed, including:
[0020] The time series feature preprocessor h(·) includes using a one-dimensional convolutional neural network to slide the convolution kernel along the time series direction to directly extract the original partial discharge UHF data, and construct a two-dimensional time series feature map by feature splicing;
[0021] The feature extractor H(·) includes a Res2Net-18 skeleton network with a grouped convolutional structure, which is used to deeply mine the nonlinear mapping relationship between the two-dimensional time series feature map output by the time series feature preprocessor and the target insulation defect type;
[0022] The classifier f(·), including a fully connected neural network, is used to learn the classification feature information of the high-dimensional features output by the feature extractor.
[0023] Furthermore, the target domain pseudo label generator for cluster prototype vector matching includes:
[0024] The source domain and target domain datasets are respectively input into the insulation defect diagnosis model to obtain the high-dimensional feature representation of the source domain and target domain samples:
[0025]
[0026] In the formula and Represent the source domain samples and target domain samples The high-dimensional feature representation of , φ is the high-dimensional feature mapping function;
[0027] In the high-dimensional feature space output by the feature extractor, K-means++ is used to identify the cluster prototype vectors of source domain samples;
[0028]
[0029] In the formula It indicates that K-means++ identifies the kth cluster set of high-dimensional feature representation of source domain samples. is the number of samples in the k-th cluster set, that is, the k-th cluster subdomain in the source domain feature space; is the cluster prototype vector of the kth cluster in the source domain;
[0030] Based on the Euclidean distance matching, the similarity between the target domain sample feature representation and the cluster prototype vector is generated, and the confidence of the cluster prototype vector belonging to the class is combined to generate the target domain sample pseudo label, where the confidence of the cluster prototype vector belonging to the class is determined according to the number of samples of each category in the cluster set in the source domain:
[0031] Where W k is the logical vector of the confidence that the cluster prototype vector belongs to the class; express Contains the number of samples of type m; The target domain sample Pseudo label vector of ; express and The Euclidean distance of Represents a high-dimensional feature function based on Output feature representation; N C is the total number of clusters. Further, in step S6, the cluster prototype vector is used as an anchor point, and the same samples in the cluster subdomain are used as positive sample pairs, and other samples in the domain are used as negative sample pairs:
[0032] In the formula, the source domain positive and negative sample pairs are constructed based on the high-dimensional features of the source domain dataset, where is the cluster prototype vector of the kth cluster in the source domain; and Represent the positive and negative sample pairs in the source domain respectively; represents the kth cluster subdomain in the source domain; c represents the label attribute; Indicates subdomain The number of samples with label attribute c in ; c k is the class label with the largest number of samples in the cluster; u i and v j Respectively indicate belonging to and Elements of Because N C Cluster subdomain The source domain feature space constructed; For I S Medium Similarly, to construct the positive and negative sample pairs in the target domain, the target domain feature space is expressed as in In turn, they represent the cluster subdomains in the target domain; according to the above method, a positive and negative sample pair of the target domain can also be constructed, wherein the target domain label information comes from the pseudo label generated by the target domain pseudo label generator. Further, in step S6, the contrastive learning mechanism is as follows:
[0033] In the formula Indicated by N C Cluster subdomain The source domain feature space, σ(r 1 ,r 2 )=(r 1 T ·r 2 ) / (||r 1 ||·||r 1 ||) indicates the relationship between r 1 、r 2 The similarity measurement function, r 1 、r 2 denote the two variables of the input of the function σ(·), and Represent the positive and negative sample pairs in the source domain, is the cluster prototype vector of the kth cluster in the source domain, τ is the temperature hyper-parameter, which determines the model's attention to positive and negative sample pairs; represents the impairment term of cluster contrastive learning mechanism in the source domain;
[0034] Similarly, cluster contrastive learning is implemented in the target domain:
[0035]
[0036] In the formula is the cluster prototype vector in the kth cluster of the target domain.
[0037] Furthermore, in step S7, the cluster subdomains are mapped to the high-dimensional regenerated Hilbert space using the Gaussian radial basis function, and the maximum mean difference between the source domain and the target domain features at the cluster subdomain level is solved by the kernel method:
[0038]
[0039] In the formula and Respectively represent the set of cluster subdomains with the largest number of m-type samples in the source domain and the target domain, and Represent the source domain samples and target domain samples The high-dimensional feature representation of express Different from Elements of express Different from Elements of Respectively represent the confidence that the i-th cluster in the source domain and the target domain belongs to the m-th class sample, which can be obtained according to the confidence logic vector of the cluster prototype vector belonging to the class in claim 6; M is the total number of insulation fault types; represents Gaussian radial basis function; is the maximum mean difference between the source domain and the target domain at the cluster subdomain level.
[0040] Further, in step S8, the source domain cluster contrast learning impairment term is jointly optimized Target domain cluster contrast learning impairment and the maximum mean difference between the source domain and the target domain at the cluster subdomain level The back-propagation algorithm is used to update the model parameters of the temporal feature preprocessor h(·) and the feature extractor H(·).
[0041] Furthermore, in step S9, the source domain dataset containing label information is Input into the model to predict the insulation defect type, where x i represents the i-th partial discharge UHF signal in the source domain, y i represents the actual insulation fault type corresponding to the i-th sample, and the classification task damage item is:
[0042]
[0043] In the formula represents the operator for obtaining the expectation; The source domain dataset Sample elements in ; Classify the damage items in the source domain and optimize And the back-propagation algorithm is used to update the classifier f(·) parameters.
[0044] Compared with the existing technology, the embodiments of the present invention have the following beneficial effects:
[0045] The present invention provides a method for diagnosing insulation faults of power equipment based on unsupervised domain adaptation. The invention solves the problem that the robustness of the insulation fault diagnosis model is threatened by the difference between the training data and the test data in the new scenario application and the missing label information. The present invention constructs a cluster contrast learning mechanism to compress the distance between the cluster prototype vector and the same samples in the cluster in the feature space of the source domain and the target domain and push away other samples, thereby enhancing the mutual information between the cluster subdomain and the class fine-grained features and the discriminable features of the class in the domain; in addition, by matching the source domain cluster prototype vector to generate the target domain pseudo-label, reliable supervision information is provided for the cluster subdomain feature matching. It is worth noting that the present invention aligns the source domain and target domain features at the cluster level, which can better transfer the fine-grained class discrimination information in the source domain to the target domain, thereby improving the robustness and generalization of the model. PD ultra-high frequency data under different insulation defect conditions are collected in the gas-insulated metal-enclosed switchgear experimental platform, and the method described in the present invention is used for on-site diagnosis, showing good accuracy and robustness. BRIEF DESCRIPTION OF THE DRAWINGS
[0046] Figure 1 It is a block diagram of a method for diagnosing insulation faults of power equipment based on unsupervised domain adaptation according to an embodiment of the present invention;
[0047] Figure 2 1 is a structural diagram of a temporal feature preprocessor based on a one-dimensional convolutional network according to an embodiment of the present invention;
[0048] Figure 3 is a block diagram of a feature extractor according to an embodiment of the present invention;
[0049] Figure 4 It is a flowchart of target domain pseudo label generation and cluster subdomain identification process according to an embodiment of the present invention;
[0050] Figure 5 It is a block diagram of the cluster contrast learning mechanism and cluster sub-domain feature alignment principle of an embodiment of the present invention;
[0051] Figure 6 1 is a comparison diagram of GIS insulation fault diagnosis results in an unseen scenario without labels (target domain) according to an embodiment of the present invention;
[0052] Figure 7 It is a visualization diagram of the model feature distribution of an embodiment of the present invention before and after using the method of the present invention. DETAILED DESCRIPTION
[0053] For ease of understanding, the present invention is further described in detail below in conjunction with embodiments and drawings, but the embodiments of the present invention are not limited thereto.
[0054] like Figure 1As shown, an embodiment of the present invention provides a method for diagnosing insulation faults of power equipment based on unsupervised domain adaptation, including using a source domain pre-trained insulation fault diagnosis model containing label information and an unsupervised domain adaptive model parameter fine-tuning framework, wherein the insulation fault diagnosis model includes a time series feature preprocessor, a feature extractor, and a classifier. The method includes the following steps:
[0055] S1. Collect 4 kinds of partial discharge UHF signal data under different working conditions, and construct different data sets according to the differences in working conditions; the data sets contain label information, a total of N S The dataset of samples is used as the source domain x i represents the i-th partial discharge UHF signal in the source domain, y i Indicates the insulation fault type corresponding to the i-th sample, without label information, totaling N T The target domain is a dataset of samples (x j Represents the jth partial discharge UHF signal in the target domain.
[0056] As an example, this example conducts experiments on PD induced by four types of insulation defects based on a 100kV gas-insulated metal-enclosed switchgear (GIS) experimental platform, and uses a UHF sensor to detect the UHF signal during the discharge process. In order to construct a PD UHF dataset under different working conditions, this example conducts PD experiments on various defects under the conditions of discharge starting voltage and 120% discharge starting voltage, and the datasets are recorded as χ 0 , χ 1 .
[0057] S2. Build an insulation defect diagnosis model based on deep learning. Its structure includes a time series feature preprocessor h(·), a feature extractor H(·) and a classifier f(·). The interconnection method is as follows: Figure 1 The pre-trained insulation diagnosis model is shown in Figure 2. The structural design of the model includes the following steps:
[0058] like Figure 2 As shown, the time series feature preprocessor h(·) is realized by directly extracting the original partial discharge UHF data through a one-dimensional convolutional neural network by sliding the convolution kernel along the time series direction, and constructing a two-dimensional time series feature map using feature splicing technology;
[0059] The feature extractor H(·) is composed of a Res2Net-18 skeleton network with a grouped convolutional structure, which is used to deeply mine the nonlinear mapping relationship between the two-dimensional time series feature map output by the time series feature preprocessor and the target insulation defect type. Its structural design and corresponding parameters are as follows Figure 3 Shown
[0060] The classifier f(·) is composed of a fully connected neural network and is used to learn the classification feature information of the high-dimensional features output by the feature extractor;
[0061] S3, determining a data set containing label information collected in step S1 as the source domain, and another data set without label information as the target domain. Use the source domain data set to train the insulation diagnosis model built in step S2, and retain the model parameters;
[0062] S4. Input the source domain and target domain data determined in step S3 into the pre-trained insulation diagnosis model to obtain high-dimensional feature representations of the source domain and target domain samples:
[0063]
[0064] In the formula and Represent the source domain samples and target domain samples The high-dimensional feature representation of is, and φ is recorded as the high-dimensional feature mapping function. Figure 4 As shown, firstly, K-means++ is used to identify the cluster subdomains and cluster prototype vectors of the source domain in the high-dimensional feature space:
[0065]
[0066] In the formula It indicates that K-means++ identifies the kth cluster subdomain represented by the high-dimensional features of the source domain samples. is the number of samples in the k-th cluster set, that is, the k-th cluster subdomain in the source domain feature space; is the cluster prototype vector of the kth cluster in the source domain.
[0067] Furthermore, for classification tasks, the essence of intelligent algorithms is to predict sample categories based on sample feature similarity. Figure 4 As shown in the figure, this method matches the similarity between the feature representation of the target domain sample and the cluster prototype vector based on the Euclidean distance, and combines the confidence that the cluster prototype vector belongs to the class to generate the pseudo label of the target domain sample. Since the cluster prototype vector is the geometric center of the cluster subdomain sample, it has good noise resistance. In addition, the confidence that the cluster prototype vector belongs to the class is determined according to the number of samples of each category in the cluster set of the source domain:
[0068]
[0069] Where W k is the logical vector of the confidence that the cluster prototype vector belongs to the class; express The number of samples of type q is contained; The target domain sample Pseudo label vector of ; express and The Euclidean distance of Represents a high-dimensional feature function based on Output feature representation; N C is the total number of clusters.
[0070] The cluster subdomains in the source and target domains output by the feature extractor are identified by K-means++; similarly, the cluster subdomains and cluster prototype vectors of the target domain in the high-dimensional feature space are identified as follows:
[0071]
[0072] In the formula It indicates that K-means++ identifies the kth cluster subdomain of the high-dimensional feature representation of the target domain samples. is the number of samples in the k-th cluster set, that is, the k-th cluster subdomain in the target domain feature space; is the cluster prototype vector of the kth cluster in the target domain.
[0073] S5. Taking the above cluster prototype vector as the anchor point, the same samples in the corresponding cluster subdomain are taken as positive sample pairs, and other samples in the domain are taken as negative sample pairs. Taking the source domain features as an example, the construction method of positive and negative sample pairs is described as follows:
[0074]
[0075] In the formula Because N C Cluster subdomain The source domain feature space constructed; and Represent the positive and negative sample pairs in the source domain respectively; represents the kth cluster subdomain in the source domain; c represents the label attribute; Indicates subdomain The number of samples with label attribute c in ; c k is the class label with the largest number of samples in the cluster; u i and v j Respectively indicate belonging to and Elements of For I S Remove The subsequent sample set.
[0076] According to the same construction method, the construction of positive and negative sample pairs in the target domain is realized, where the target domain label information comes from the pseudo label generated by the target domain pseudo label generator:
[0077]
[0078] Where the target domain feature space The label information comes from the pseudo-label generated in S4; represents the target domain feature space, where are cluster subdomains in the target domain; u i and v j Respectively indicate belonging to and Elements; c represents the tag attribute; Indicates subdomain The number of samples with label attribute c; is the class label with the largest number of samples in the cluster in the target domain.
[0079] S6, such as Figure 5 As shown in the figure, in the feature spaces of the source domain and the target domain, cluster contrast learning is used to compress the distance between similar sample pairs in the cluster subdomain, expand the gap with other sample feature representations, and enhance the intra-domain class discriminant features. The specific description is as follows:
[0080]
[0081] Where σ(r 1 ,r 2 )=(r 1 T ·r 2 ) / (||r 1 ||·||r 1 ||) represents the similarity metric function about r1 and r2 (r 1 、r 2 denote the two variables of the input of the function σ(·), and Represent the positive and negative sample pairs in the source domain, is the cluster prototype vector of the kth cluster in the source domain, and τ is the temperature hyper-parameter, which determines the attention paid by the contrastive learning mechanism to the positive and negative sample pairs. Represents the impairment term of the cluster contrastive learning mechanism in the source domain.
[0082] Similarly, cluster contrastive learning also uses the impairment term in the target domain It is expressed as:
[0083]
[0084] In the formula is the cluster prototype vector in the kth cluster of the target domain.
[0085] Furthermore, the contrastive learning mechanism comprehensive damage is described as follows:
[0086]
[0087] In the formula I=I S ∪I T Represents a dataset of synthetic source and target domains.
[0088] S7. To facilitate efficient calculation of the insulation diagnosis model, this embodiment uses Gaussian radial basis function to map the cluster subdomains to
[0089] The kernel method is used to solve the relationship between the source domain and the target domain features at the cluster subdomain level.
[0090] Maximum mean difference:
[0091]
[0092] In the formula and Respectively represent the set of cluster subdomains with the largest number of m-type samples in the source domain and the target domain, and They are and The two different elements in it represent the high-dimensional features of the source domain samples and the target domain samples; Respectively represent the confidence that the i-th cluster in the source domain and the target domain belongs to the m-th class sample, which can be obtained according to the confidence logic vector of the cluster prototype vector belonging to the class in claim 6; M is the total number of insulation fault types; represents Gaussian radial basis function; is the maximum mean difference between the source domain and the target domain at the cluster subdomain level.
[0093] As a preferred embodiment, cluster subdomain feature alignment can better refine class discriminant features, avoid the problem of blurred target domain class boundaries caused by large-scale distribution of class features during feature alignment, and ensure that fine-grained class discriminant information is effectively transferred from the source domain to the target domain.
[0094] S8. Jointly optimize source domain cluster contrastive learning impairment term Target domain cluster contrast learning impairment Align the damage term with cluster subdomain features The back-propagation algorithm is used to update the model parameters of the temporal feature preprocessor h(·) and the feature extractor H(·);
[0095] S9: Source domain dataset containing label information Input into the model to predict the insulation defect type, where x i represents the i-th partial discharge UHF signal in the source domain, y irepresents the actual insulation fault type corresponding to the i-th sample, and the classification task damage item is:
[0096]
[0097] In the formula represents the operator for obtaining the expectation; The source domain dataset Sample elements in ; Classify the damage items in the source domain and optimize And the back-propagation algorithm is used to update the classifier f(·) parameters.
[0098] S10, input the target domain data set without label information into the model, and update the model parameters according to the steps S8 and S9. After the insulation fault diagnosis model is pre-trained with data with known label information, a new data set without label information is input in a new unknown scenario to fine-tune the model parameters, so that the model can diagnose insulation faults in the new unknown scenario and output the insulation fault type.
[0099] Figure 6 The results of the insulation diagnosis model's accuracy in identifying various defects in the target domain before and after using the unsupervised domain adaptive framework in the present invention. The results in the figure show that in this embodiment, the unsupervised domain adaptive framework proposed in the present invention can significantly improve the model's accuracy in diagnosing insulation faults in unseen scenarios.
[0100] Figure 7 Visualization of high-dimensional features of the target domain based on t-distributed random neighbor embedding. Figure 7 (a) is the feature distribution diagram of the target domain data set directly identified by the insulation defect diagnosis model, and (b) is the feature distribution diagram of the target domain data set identified by the diagnosis model after being processed by the method of the present invention. The results show that the method of the present invention enables the insulation diagnosis model to have good intra-class aggregation and clear inter-class boundaries at the feature level, thereby greatly improving the robustness of the model in the target domain.
[0101] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not violate the spirit and principles of the present invention should be equivalent replacement methods and are included in the protection scope of the present invention.
Claims
1. A method for diagnosing insulation faults of power equipment based on unsupervised domain adaptation, characterized in that: The following steps are involved: S1. Collect UHF data of partial discharge of power equipment; S2. Build an insulation defect diagnosis model based on deep learning, including a time series feature preprocessor, a feature extractor, and a classifier; S3, using a dataset containing label information in a known scenario as a source domain, and pre-training an insulation fault diagnosis model based on it, and retaining the model parameters; S4, constructing a target domain pseudo label generator matching the cluster prototype vector in the high-dimensional feature space output by the feature extractor; S5, taking the unlabeled dataset in the unseen scene as the target domain, and using the unsupervised clustering algorithm to construct cluster subdomains of the source domain and the target domain in the feature space; S6. Taking the cluster prototype vector as the anchor point and the similar samples in the cluster subdomain as the positive sample pairs, a cluster contrast learning mechanism is constructed to strengthen the machine granularity of the discriminable features in the domain; S7. Use Gaussian radial basis function to map the cluster subdomain to high-dimensional Hilbert space, construct a cluster subdomain feature matcher, align the cluster subdomain features in the source domain and the target domain, and realize the migration of class-discriminable feature information from the source domain to the target domain; S8, combining cluster contrast learning mechanism and cluster subdomain feature matcher to update feature extractor parameters; S9. Use the source domain dataset containing label information to fine-tune the classifier parameters to achieve insulation defect fault diagnosis.
2. The method for diagnosing insulation faults of power equipment based on unsupervised domain adaptation according to claim 1 is characterized in that: In step S1, the types of insulation defects of the power equipment include foreign matter attachment, metal tips, metal particles and insulating air gaps. The insulation defects of the power equipment induce local discharge inside the equipment, and the corresponding discharge signals are collected using ultra-high frequency sensors.
3. The method for diagnosing insulation faults of power equipment based on unsupervised domain adaptation according to claim 1 is characterized in that: Collect four types of partial discharge UHF signal data under different working conditions, and construct different data sets according to the differences in working conditions; In the data set, N containing label information S The dataset of samples is used as the source domain x i represents the i-th partial discharge UHF signal in the source domain, y i Indicates the insulation fault type corresponding to the i-th sample; N without label information T The target domain is a dataset of samples x j Represents the jth partial discharge UHF signal in the target domain.
4. The method for diagnosing insulation faults of power equipment based on unsupervised domain adaptation according to claim 1, characterized in that: In step S2, an insulation defect diagnosis model based on deep learning is constructed, including: The time series feature preprocessor h(·) includes using a one-dimensional convolutional neural network to slide the convolution kernel along the time series direction to directly extract the original partial discharge UHF data, and construct a two-dimensional time series feature map by feature splicing; The feature extractor H(·) includes a Res2Net-18 skeleton network with a grouped convolutional structure, which is used to deeply mine the nonlinear mapping relationship between the two-dimensional time series feature map output by the time series feature preprocessor and the target insulation defect type; The classifier f(·), including a fully connected neural network, is used to learn the classification feature information of the high-dimensional features output by the feature extractor.
5. The method for diagnosing insulation faults of power equipment based on unsupervised domain adaptation according to claim 1, characterized in that: The target domain pseudo-label generator for cluster prototype vector matching includes: The source domain and target domain datasets are respectively input into the insulation defect diagnosis model to obtain the high-dimensional feature representation of the source domain and target domain samples: In the formula and Represent the source domain samples and target domain samples The high-dimensional feature representation of , φ is the high-dimensional feature mapping function; In the high-dimensional feature space output by the feature extractor, K-means++ is used to identify the cluster prototype vectors of source domain samples; In the formula It indicates that K-means++ identifies the kth cluster set of high-dimensional feature representation of source domain samples. is the number of samples in the k-th cluster set, that is, the k-th cluster subdomain in the source domain feature space; is the cluster prototype vector of the kth cluster in the source domain; Based on the Euclidean distance matching, the similarity between the target domain sample feature representation and the cluster prototype vector is generated, and the confidence of the cluster prototype vector belonging to the class is combined to generate the target domain sample pseudo label, where the confidence of the cluster prototype vector belonging to the class is determined according to the number of samples of each category in the cluster set in the source domain: Where W k is the logical vector of the confidence that the cluster prototype vector belongs to the class; express Contains the number of samples of type m; The target domain sample Pseudo label vector of ; express and The Euclidean distance of Represents a high-dimensional feature function based on Output feature representation; N C is the total number of clusters.
6. The method for diagnosing insulation faults of power equipment based on unsupervised domain adaptation according to claim 1, characterized in that: In step S6, the cluster prototype vector is used as an anchor point, and the same type of samples in the cluster subdomain are used as positive sample pairs, and other samples in the domain are used as negative sample pairs: In the formula, the source domain positive and negative sample pairs are constructed based on the high-dimensional features of the source domain dataset, where is the cluster prototype vector of the kth cluster in the source domain; and Represent the positive and negative sample pairs in the source domain respectively; represents the kth cluster subdomain in the source domain; c represents the label attribute; Indicates subdomain The number of samples with label attribute c in ; c k is the class label with the largest number of samples in the cluster; u i and v j Respectively indicate belonging to and Elements of Because N C Cluster subdomain The source domain feature space constructed; For I S Medium Similarly, to construct positive and negative sample pairs in the target domain, the target domain feature space is expressed as in represent the cluster subdomains in the target domain in turn; according to the above method, the target domain positive and negative sample pairs can also be constructed, where the target domain label information comes from the pseudo label generated by the target domain pseudo label generator.
7. The method for diagnosing insulation faults of power equipment based on unsupervised domain adaptation according to claim 1, characterized in that: In step S6, the contrastive learning mechanism is as follows: In the formula Indicated by N C Cluster subdomain The source domain feature space is constructed, σ(r1,r2)=(r1T·r2) / (||r1||·||r1||) represents the similarity measurement function about r1 and r2, r1 and r2 represent the two variables input by the function σ(·), and Represent the positive and negative sample pairs in the source domain, is the cluster prototype vector of the kth cluster in the source domain, τ is the temperature hyper-parameter, which determines the model's attention to positive and negative sample pairs; represents the impairment term of cluster contrast learning mechanism in the source domain; Similarly, cluster contrastive learning is implemented in the target domain: In the formula is the cluster prototype vector in the kth cluster of the target domain.
8. The method for diagnosing insulation faults of power equipment based on unsupervised domain adaptation according to claim 1, characterized in that: In step S7, the cluster subdomains are mapped to the high-dimensional regenerated Hilbert space using the Gaussian radial basis function, and the maximum mean difference between the source domain and the target domain features at the cluster subdomain level is solved by the kernel method: In the formula and Respectively represent the set of cluster subdomains with the largest number of m-type samples in the source domain and the target domain, and Represent the source domain samples and target domain samples The high-dimensional feature representation of express Different from Elements of express Different from Elements of Respectively represent the confidence that the i-th cluster in the source domain and the target domain belongs to the m-th class sample, which can be obtained according to the confidence logic vector of the cluster prototype vector belonging to the class in claim 6; M is the total number of insulation fault types; represents Gaussian radial basis function; is the maximum mean difference between the source domain and the target domain at the cluster subdomain level.
9. The method for diagnosing insulation faults of power equipment based on unsupervised domain adaptation according to claim 1, characterized in that: In step S8, the source domain cluster contrast learning impairment term is jointly optimized Target domain cluster contrast learning impairment and the maximum mean difference between the source domain and the target domain at the cluster subdomain level The back-propagation algorithm is used to update the model parameters of the temporal feature preprocessor h(·) and the feature extractor H(·).
10. The method for diagnosing insulation faults of power equipment based on unsupervised domain adaptation according to any one of claims 1 to 9, characterized in that: In step S9, the source domain dataset containing label information is Input into the model to predict the insulation defect type, where x i represents the i-th partial discharge UHF signal in the source domain, y i represents the actual insulation fault type corresponding to the i-th sample, and the classification task damage item is: In the formula represents the operator for obtaining the expectation; The source domain dataset Sample elements in ; Classify the damage items in the source domain and optimize And the back-propagation algorithm is used to update the classifier f(·) parameters.
Citation Information
Patent Citations
Rolling bearing fault diagnosis method based on self-supervised learning and clustering
CN113792758A
Face tampering video detection method and system based on double-flow contrast learning model
CN115116108A
Abnormal multimedia resource identification method and device, electronic equipment and storage medium
CN116563679A
Method, apparatus, and system for wireless object tracking
US20200191943A1
