A two-stage semi-supervised transformer fault diagnosis method with active learning

Abstract

The present invention provides a two-stage semi-supervised transformer fault diagnosis method incorporating active learning, including DGA data feature space selection, DGA data preprocessing, DGA data set division, one-to-many training of n FLDA two classifiers on an initial labeled sample set SI, and m Second and m 0+ k Iterations. The present invention integrates active learning into semi-supervised learning, updates the model by manually labeling a small number of noise samples with large uncertainty that are easy to be mislabeled, and cooperates with semi-supervised learning to enhance the robustness of the initial model, so that the initial classification model is improved. The collaborative training in semi-supervised learning is then used to fully explore the implicit information in the remaining unlabeled data, that is, the model is trained with samples with high confidence that are consistent with the distribution of labeled samples, which further improves the overall performance of transformer fault diagnosis on the basis of reducing the cost of manual labeling.

Description

Technical Field

[0001] The invention relates to the technical field of transformer fault diagnosis, and in particular to a two-stage semi-supervised transformer fault diagnosis method incorporating active learning. Background Art

[0002] Power transformers are important power transmission and transformation equipment in power systems. Whether they operate safely or not will directly affect the reliable supply of electric energy. Dissolved Gas Analysis (DGA) is the most important method for power transformer fault diagnosis. It can monitor various gas data in oil without power outages. The proportion and content of characteristic gas components can reflect the fault type and severity. In recent years, with the extensive development of artificial intelligence technology in various fields, artificial intelligence technologies including support vector machines, decision trees, neural networks, clustering, fuzzy logic reasoning, etc. have been widely used in transformer fault diagnosis and have achieved good results. However, a large amount of DGA data with labeled fault types is required to establish models. Models established with too little labeled data often lead to poor generalization performance of the learner and reduced fault diagnosis rate. In fact, the reliable fault samples available on site are very limited. First, there is less data on the on-site fault status and the actual fault data is not accurately determined. Dismantling is required to confirm the true fault type of the transformer and obtain reliable fault category samples. However, since the transformer bears the important task of power supply and the dismantling process consumes a lot of manpower, material and financial resources, the transformer will only be dismantled when the on-site judgment and assessment risk is high, and the dismantling result may not be consistent with the on-site judgment. Second, the transformer fault data is in an opaque state, resulting in extremely limited and distorted transformer fault data that can be used. This means that traditional artificial intelligence algorithms cannot effectively diagnose transformer faults.

[0003] Although DGA fault data is very limited, unlabeled DGA data is easily available. Semi-supervised learning uses a small amount of DGA labeled data as guidance and a large amount of DGA unlabeled data to improve learning performance, which can solve the transformer fault diagnosis problem when fault labeled samples are scarce.

[0004] In transformer fault diagnosis, semi-supervised learning usually assumes that the distribution of unlabeled DGA data is consistent with that of labeled DGA data, that is, it assumes that the predicted unlabeled data category is accurate. In fact, when labeled samples are very scarce, erroneous pseudo-labeled data are usually introduced during the iteration process. These data noises accumulate during iterations and eventually affect the performance of the classifier. These incorrectly labeled data are often the samples with the greatest uncertainty in the classification results, which makes semi-supervised learning have limited improvement in the accuracy of transformer fault diagnosis. Summary of the invention

[0005] The present invention provides a two-stage semi-supervised transformer fault diagnosis method incorporating active learning to solve the problem that transformer DGA labeled fault samples are limited and semi-supervised iteration easily introduces unknown distribution noise.

[0006] This specification discloses a two-stage semi-supervised transformer fault diagnosis method incorporating active learning, comprising:

[0007] Step 1: Get DGA data;

[0008] Step 2: Preprocess the DGA data;

[0009] Step 3: Divide the preprocessed DGA data into training set S T and validation set S V , the training set is divided into the initial labeled sample set S I and the unlabeled sample set S U ;

[0010] Step 4: Use one-to-many to initialize the sample set S I On training n FLDA two classifiers, obtain the unlabeled sample set S U All samples in n discriminant scores and predict the unlabeled sample set S U and validation set S V The categories of all samples in S V The accuracy rate of

[0011] Step 5: m In the iteration, from S U Select samples U with large uncertainty m And manually mark the real label and add it to S I In the middle, the sample C with high confidence m Add the predicted labels to S I In the new labeled training set S L ( m );

[0012] Step 6: Use S L ( m ) Re-establish the FLDA model and re-predict the updated S U All samples in n discriminant scores and predict the updated unlabeled sample set S U and validation set S V The categories of all samples in S U and S V The accuracy rate of

[0013] Step 7: Observe the m The iteration and m -1 iteration in S V Are the samples marked by the above predictions consistent? If so, the iteration stops. Otherwise, the iteration stops until the m The iteration ends if the upper limit of the value is reached; otherwise, return to step 5 to perform the m +1 iteration; the number of iteration stops is counted as m 0;

[0014] Step 8: m 0+ k The training set S is updated in the last iteration of the initial startup iteration. I A collaborative semi-supervised model is established. By establishing the FLDA model, the inter-class discreteness is maximized and the intra-class discreteness is minimized respectively, and different discrimination rules are formed to construct two different classifiers; that is, two identical initial labeled training sets, unlabeled training set data and validation set S are initialized. V ; S I1 =S I2 =S I , S U1 =S U2 =S U , S V1 =S V2 =S V ; Using S I1 and S I2 Two different classifiers C1 and C2 derived from FLDA are established respectively, and the unlabeled sample set S is obtained. U1 and S U2 All samples in n discriminant scores and predict the unlabeled sample set S U1 and S U2 The categories of all samples in the validation set S V1 and S V2 The categories of all samples in S V1 and S V2 The accuracy rate of

[0015] Step 9: According to the semi-supervised quantitative selection strategy, classifiers C1 and C2 are respectively trained from the unlabeled training set S U1 and S U2 Select the sample Q1 with high confidence ( m 0+ k ) and Q2( m 0+ k ) and add it to S after assigning it a label I1 and S I2 In the new labeled training set S L1 ( m ) and SL2 ( m );S L1 ( m )=S I1 +Q2( m 0+ k ), S L2 ( m )=S I2 +Q1( m 0+ k );

[0016] Step 10: Use the new labeled training set S L1 ( m ) and S L2 ( m ) Reconstruct two different classifiers C1 and C2 to re-predict S U1 and S U2 All samples in n discriminant scores and predict the unlabeled sample set S U1 and S U2 The categories of all samples in the validation set S V1 and S V2 The categories of all samples in S V1 and S V2 The accuracy rate on the final S V1 and S V2 The average accuracy of

[0017] Step 11: If satisfied k The iteration stops if the upper limit value is reached; otherwise, return to step 8 to perform the m 0+ k +1 iteration.

[0018] The embodiments of this specification can at least achieve the following beneficial effects:

[0019] Active learning is integrated into semi-supervised learning. A small number of noise samples with large uncertainty that are easy to mislabel are manually labeled to update the model. At the same time, semi-supervised learning is used to enhance the robustness of the initial model, so that the initial classification model is improved. Then, the collaborative training in semi-supervised learning is used to fully explore the implicit information in the remaining unlabeled data, that is, the samples with high confidence that are consistent with the distribution of labeled samples are used for model training. On the basis of reducing the cost of manual labeling, the overall performance of transformer fault diagnosis is further improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0021] Figure 1 It is a flow chart of the two-stage semi-supervised transformer fault diagnosis method incorporating active learning involved in the present invention. DETAILED DESCRIPTION

[0022] In the following, only some exemplary embodiments are briefly described. As those skilled in the art will appreciate, the described embodiments may be modified in various ways without departing from the spirit or scope of the embodiments of the present invention. Therefore, the drawings and descriptions are considered to be exemplary and non-restrictive in nature.

[0023] The disclosure below provides many different embodiments or examples to implement different structures of the embodiments of the present invention. In order to simplify the disclosure of the embodiments of the present invention, the parts and settings of specific examples are described below. Of course, they are only examples, and the purpose is not to limit the embodiments of the present invention. In addition, the embodiments of the present invention can repeat reference numbers and / or reference letters in different examples, and this repetition is for the purpose of simplification and clarity, which itself does not indicate the relationship between the various embodiments and / or settings discussed.

[0024] The embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0025] like Figure 1 As shown, this embodiment provides a two-stage semi-supervised transformer fault diagnosis method incorporating active learning, including:

[0026] initialization:

[0027] Step 1: DGA data feature space selection: select the characteristic gas content, the proportion of characteristic gas components and the effective percentage that have been proven to best reflect the fault category and severity.

[0028] Step 2: DGA data preprocessing, including using synthetic minority class oversampling technology to improve the imbalanced distribution of data, using inverse tangent transformation to deal with the situation where the numerator or denominator of the feature is 0, and normalization processing.

[0029] Step 3: Divide the DGA dataset into training set S T and validation set S V , the training set is divided into the initial labeled sample set S I and the unlabeled sample set SU , the number of labeled sample sets is much smaller than that of unlabeled sample sets.

[0030] Step 4: Use One-against-all (OAA) to identify the initial labeled sample set S I On training n FLDA (Fisher Linear Discrimination Analysis) two classifiers ( n is a category and ≥ 2), obtain the unlabeled sample set S U All samples in n discriminant scores and predict the unlabeled sample set S U and validation set S V The categories of all samples in . Calculated in S V The accuracy rate on .

[0031] Phase 1: Initial Startup Iteration Phase Steps 5 to 7 describe the m The process of iteration ( m =1, 2, ..., m S )

[0032] Step 5: m In the iteration, the qualitative selection strategy is selected based on the active fit semi-supervised learning, from S U Select samples U with large uncertainty m And manually mark the real label and add it to S I In the middle, the sample C with high confidence m Add the predicted labels to S I In the new labeled training set S L ( m ), labeled sample set S I and the unlabeled sample set S U Updated, S L ( m )=S I +U m +C m , S I =S I +U m , S U =S U -U m Note that the predicted samples with high confidence do not change the data distribution of the labeled sample set and the unlabeled sample set.

[0033] Step 6: Use S L ( m ) Re-establish the FLDA model and re-predict the updated S UAll samples in n discriminant scores and predict the updated unlabeled sample set S U and validation set S V The categories of all samples in . Calculated in S U and S V The accuracy rate on .

[0034] Step 7: Observe the m The iteration and m -1 iteration in S V Are the samples marked by the above predictions consistent? If so, the iteration stops. Otherwise, the iteration stops until the m The iteration ends. Otherwise, return to step 5 to perform the m +1 iteration. The number of iteration stops is counted as m 0.

[0035] Phase 2: Post-stable iteration phase, steps 8 to 11 describe the m 0+ k The iterative process ( k =1, 2,..., k S )

[0036] Step 8: m 0+ k The training set S is updated in the last iteration of the initial startup iteration. I A collaborative semi-supervised model is established. By establishing FLDA to maximize the inter-class discreteness and minimize the intra-class discreteness, different discrimination rules are formed to construct two different classifiers. Therefore, it is necessary to initialize two identical initial labeled training sets, unlabeled training set data, and validation set S V . S I1 =S I2 =S I , S U1 =S U2 =S U , S V1 =S V2 =S V . Using S I1 and S I2 Two different classifiers C1 and C2 derived from FLDA are established respectively, and the unlabeled sample set S is obtained. U1 and S U2 All samples in n discriminant scores and predict the unlabeled sample set S U1 and S U2 The categories of all samples in the validation set S V1 and S V2 The categories of all samples in SV1 and S V2 The accuracy rate on .

[0037] Step 9: According to the semi-supervised quantitative selection strategy, classifiers C1 and C2 are respectively trained from the unlabeled training set S U1 and S U2 Select the sample Q1 with high confidence ( m 0+ k ) and Q2( m 0+ k ) and add it to S after assigning it a label I1 and S I2 In the new labeled training set S L1 ( m ) and S L2 ( m ). L1 ( m )=S I1 +Q2( m 0+ k ), S L2 ( m )=S I2 +Q1( m 0+ k ).

[0038] Step 10: Use the new labeled training set S L1 ( m ) and S L2 ( m ) Reconstruct two different classifiers C1 and C2 to re-predict S U1 and S U2 All samples in n discriminant scores and predict the unlabeled sample set S U1 and S U2 The categories of all samples in the validation set S V1 and S V2 The categories of all samples in S V1 and S V2 The accuracy rate on the final S V1 and S V2 The average accuracy rate.

[0039] Step 11: If satisfied k The iteration stops if the upper limit of m 0+ k +1 iteration.

[0040] In a specific embodiment, the present invention uses the IEC TC10 public data set of transformers to establish a two-stage semi-supervised transformer fault diagnosis method (TSSM-AL) that integrates active learning and trains the algorithm to select the optimal feature subset. The IEC TC10 public data set contains a total of 167 samples, including six types of faults: low energy discharge (LED), high energy discharge (HED), low and medium temperature (LMT), high temperature (HT), partial discharge (PD) and normal condition (NC). The types of dissolved gases in the oil of each sample are measured to be H2, CH4, C2H2, C2H4, C2H6, CO and CO2. Due to the small amount of partial discharge data, it is not used in model establishment and algorithm training. The final number and type of samples are shown in Table 1.

[0041] Table 1. Transformer DGA sample quantity and category

[0042]

[0043] Feature space construction. Looking at the research at home and abroad, the related work focuses on the existing DGA ratios. The DGA ratios used in the diagnostic models in different literatures are obviously different, and a unified standard for the gas ratios used in the diagnostic models has not yet been formed. The present invention analyzes the relationship between all relevant features of DGA and faults, and selects characteristic gas content, the proportion of characteristic gas components and the effective percentage that have been proven to reflect the fault category and severity, totaling 12 dimensions. (C2H2 / CO, C2H4 / CO2,C2H2 / CH4, CH4 / THH, CH4 / TCH, C2H2 / THD, C2H2 / THH, C2H4 / THH, C2H6 / TH, 10 / (C2H4 / C2H2),C2H2 / THD / 0.21, C2H6 / THD / 0.23, Among them, THH=H2+C2H2+C2H4, TCH=CH4+C2H2+C2H4+C2H, THD=CH4+C2H2+C2H4, TH=H2+CH4+C2H2+C2H4+C2H6).

[0044] The 12-dimensional DGA dataset was preprocessed, including using the Synthetic Minority Oversampling Technique (SMOTE) to change the imbalance of the data, using the Arctangent Transformation (AT) to process the denominator or numerator of the feature ratio corresponding to the balanced data to be 0, and normalization. The balanced data includes 247 samples.

[0045] ;

[0046] ;

[0047] x represents a characteristic quantity of the sample, where x ij and x' ij Respectively represent i The samples before and after normalization j eigenvalues, x j max and x j min Represents all samples j The maximum and minimum values ​​of the features.

[0048] Dataset division: The preprocessed dataset is divided into a training set S including 197 samples according to 20*5-fold cross validation. T and a validation set S of 50 samples V The training set is divided into an initial labeled sample set S consisting of 75 samples I and an unlabeled sample set S of 122 samples U (20*5-fold cross validation divides the sample set into 5 equal parts, 4 of which are used as training sets and 1 as validation set. The order of the sample set is randomly shuffled 20 times to form 100 different combinations of training sets and validation sets).

[0049] Using OAA in the initial labeled sample set S I On training n FLDA (Fisher Linear Discrimination Analysis, FLDA) two classifiers ( n For categories, n ≥2), obtain the unlabeled sample set S U All samples in n discriminant scores and predict the unlabeled sample set S U and validation set S VThe categories of all samples in .

[0050] The goal of FLDA is to find a transformation matrix W , maximize the distance between the two classes and minimize the distance within the class, using the transformation matrix W And the threshold weight b calculates the discriminant score f ( x ).

[0051] ;

[0052] ;

[0053] ;

[0054] ;

[0055] ;

[0056] ;

[0057] ;

[0058] Represents a sample n dimensional feature vector, is called the weight vector. w 0 represents a constant, called the threshold weight, S w represents the intra-class dispersion, S b represents the inter-class dispersion, N 1 and N 2 represents the number of samples of category 1 and category 2 respectively, N represents the total number of samples, m d Indicates d The sample mean vector of the class, m Represents the sample total mean vector.

[0059] FLDA is used to solve two-classification problems, while transformer fault classification is a multi-classification problem. Therefore, FLDA needs to be expanded to multi-classification FLDA. The present invention uses OAA to achieve multi-classification. OAA is used for training n ( n After training, the discrimination scores of all unlabeled samples and validation set samples are calculated based on the binary classifier. f j ( x )( j = 1, 2, …, n). When classifying unlabeled samples and validation set samples, f j ( x ) value is the classification result of the unlabeled sample.

[0060] Conduct the m Iterations ( m =1, 2, ..., m S ), according to the active fit semi-supervised learning qualitative sample selection strategy, from S U Select samples U with large uncertainty m And manually mark the real label and add it to S I In the middle, the sample C with high confidence m Add the predicted labels to S I In the new labeled training set S L ( m ), labeled sample set S I and the unlabeled sample set S U Updated, S L ( m )=S I +U m +C m , S I =S I +U m , S U =S U -U m Note that the predicted samples with high confidence do not change the data distribution of the labeled sample set and the unlabeled sample set.

[0061] Actively adapting the qualitative sample selection strategy for semi-supervised learning: using FLDA to select S L ( m ) is modeled, and in S U Get each sample on n There are discriminant scores scores(x,n), if n At least n -1 discriminant score is greater than 0, then the sample is judged to be highly uncertain, recorded as Select1, and marked as Um. n If at most one of the discriminant scores is greater than 0 (excluding the case where all the discriminant scores are negative), the confidence of this sample is judged to be large, which is recorded as Select2 and marked as Cm.

[0062] ;

[0063] ;

[0064] ;

[0065] ;

[0066] x represents a sample, n represents the number of categories, S L ( m ) is the updated sample set. L ( m ) Re-establish the FLDA model and re-predict the updated S U All samples in n discriminant scores and predict the updated unlabeled sample set S U and validation set S V The categories of all samples in . Calculated in S U and S V The accuracy rate on .

[0067] Observe the m The iteration and m -1 iteration in S V Are the samples marked by the above predictions consistent? If so, the iteration stops. Otherwise, the iteration stops until the m The iteration ends. Otherwise, return to step 5 to perform the m +1 iteration. The number of iteration stops is counted as m 0.

[0068] when m 0+ k The training set S is updated in the last iteration of the initial startup iteration. I A collaborative semi-supervised model is established. By establishing FLDA to maximize the inter-class discreteness and minimize the intra-class discreteness, different discrimination rules are formed to construct two different classifiers. Therefore, it is necessary to initialize two identical initial labeled training sets, unlabeled training set data, and validation set S V . S I1 =S I2 =S I , S U1 =S U2 =S U , S V1 =S V2 =S V . Using S I1 and S I2 Two different classifiers C1 and C2 derived from FLDA are established respectively, and the unlabeled sample set S is obtained. U1 and S U2 All samples in n discriminant scores and predict the unlabeled sample set S U1 and SU2 The categories of all samples in the validation set S V1 and S V2 The categories of all samples in S V1 and S V2 The accuracy rate on .

[0069] The two significantly different classifiers derived from FLDA are denoted as C1 and C2. The goals of C1 and C2 are to maximize the inter-class divergence and minimize the intra-class divergence, respectively. C1 and C2 need to be continuously learned in semi-supervised learning until the two classifiers are similar.

[0070] Defining population dispersion S t :

[0071] ;

[0072] The criteria for judging C1 and C2 are:

[0073] ;

[0074] ;

[0075] in W 1 and W 2 are the projection directions of FLDA1 and FLDA2 classifiers respectively. b1 and b2 are the threshold weights of the two classifiers respectively. According to the semi-supervised quantitative selection strategy, classifiers C1 and C2 are the unlabeled training set S U1 and S U2 Select the sample Q1 with high confidence ( m 0+ k ) and Q2( m 0+ k ) and add it to S after assigning it a label I1 and S I2 In the new labeled training set S L1 ( m ) and S L2 ( m ). L1 ( m )=S I1 +Q2( m 0+ k ), S L2 ( m )=S I2 +Q1( m 0+ k ).

[0076] Semi-supervised quantitative sample selection strategy: Using FLDA to update the two identical training sets S I1and S I2 Modeling is performed separately, and in the unlabeled training set S U1 and S U2 On each sample of n The discriminant scores are calculated, and the highest discriminant score is predicted as the corresponding category. Therefore, all samples have a label corresponding to the corresponding discriminant score mean1 or mean2. The difference between the discriminant score and the average discriminant score of all samples is calculated, and sorted from large to small, recorded as seq1 or seq2. The samples with the top 80% difference are selected as Q1 ( m 0+ k ) or Q2( m 0+ k ).

[0077] ;

[0078] ;

[0079] Using the new labeled training set S L1 ( m ) and S L2 ( m ) Reconstruct two different classifiers C1 and C2 to re-predict S U1 and S U2 All samples in n discriminant scores and predict the unlabeled sample set S U1 and S U2 The categories of all samples in the validation set S V1 and S V2 The categories of all samples in S V1 and S V2 The accuracy rate on .

[0080] If satisfied k The iteration stops if the upper limit of m 0+ k +1 iteration.

[0081] According to the DGA data set of IEC TC 10, the present invention compares the average accuracy of transformer fault diagnosis on the validation set of the two-stage semi-supervised method with active learning for transformer fault diagnosis (TSSM-AL), the two-stage semi-supervised method with random sampling for transformer fault diagnosis (TSSM-RS) and the standard learning method for transformer fault diagnosis (SLM) based on 20*5 fold cross validation, and the results are shown in Table 2. It is found that the proposed TSSM-AL algorithm has the highest accuracy, and SLM has a low accuracy because it does not use unlabeled samples to expand the labeled sample set. Although TSSM-RS uses semi-supervised learning in two stages, the manually labeled samples used to update the model in the initial startup stage are randomly selected, and the amount of information is small, which cannot maximize the robustness of the model. Note: Standard learning means that the labeled sample set is not expanded with unlabeled samples, and only the initial labeled samples are used to build the model. Passive learning means randomly selecting some samples from the unlabeled sample set and having them manually labeled by experts, and then adding the selected samples to the initial training set to update the classifier.

[0082] Table 2. Transformer fault diagnosis accuracy under different algorithms

[0083]

[0084] The present invention also integrates a two-stage semi-supervised method with active learning for transformer fault diagnosis (TSSM-AL), an initial start method with active semi-supervised learning for transformer fault diagnosis (ISM-ASSL), an initial start method with active learning for transformer fault diagnosis (ISM-AL), an initial start method with semi-supervised learning for transformer fault diagnosis (ISM-SSL), and a post-stabilization method with semi-supervised learning for transformer fault diagnosis (Post Stabilization Method with Semi-supervised Learning for Transformer Fault Diagnosis). The average accuracy of transformer fault diagnosis on the validation set of ISM-ASSL, PSM-SSL, and ISM-AL were compared after 20*5 fold cross validation, showing the average fault diagnosis rate of different algorithms under the conditions of active semi-supervised learning qualitative sample selection strategy in the initial startup phase, semi-supervised quantitative sample selection strategy in the post-stabilization phase, and two-stage sample selection strategy. The results are shown in Table 3. TSSM-AL includes two stages of learning, while other algorithms only include one stage of learning (ISM-ASSL, ISM-AL, and ISM-SSL only learn in the initial startup phase, and PSM-SSL only learns in the post-startup phase). The active semi-supervised learning qualitative sample selection strategy in the initial startup phase of TSSM-AL refers to the semi-supervised learning used in the post-stabilization phase, which randomly selects samples from the unlabeled sample set for model training, and its corresponding post-stabilization phase semi-supervised quantitative sample selection strategy refers to the active semi-supervised learning used in the initial startup phase, which randomly selects samples from the unlabeled sample set for manual labeling and prediction to train the model. The only active qualitative strategy of ISM-AL in the initial startup phase means that no semi-supervised learning is performed in this phase, and the only semi-supervised qualitative strategy of ISM-SSL in the initial startup phase means that no active learning is performed in this phase.

[0085] Table 3. Average accuracy of transformer fault diagnosis obtained by different algorithms under different sample selection strategies based on IEC TC 10 DGA dataset

[0086]

[0087] The results in Table 3 show that TSSM-AL uses strategies to select appropriate samples in both the initial startup and post-stabilization stages, which helps improve the accuracy. If only the active semi-supervised learning qualitative sample selection strategy is used in the initial startup stage or the semi-supervised quantitative sample selection strategy is used only in the post-stabilization stage, it means that the samples in the unlabeled sample set are randomly selected in the post-stabilization stage or the samples are randomly selected in the initial startup stage for model training, which weakens the robustness of the model to a certain extent. Obviously, the strategy in the initial startup stage is more effective than the strategy in the post-stabilization stage, and the strategy in the post-stabilization stage is an effective supplement to the training method in the initial startup stage. For the ISM method, learning is only carried out in the initial startup stage. If only the active qualitative strategy or semi-supervised qualitative strategy in the initial startup stage is used, the results obtained are not as good as their joint use (i.e., the active semi-supervised learning qualitative sample selection strategy). For the PSM-SSL method, learning was performed only in the post-stabilization stage, and the accuracy was higher than that of the ISM method using only the semi-supervised qualitative strategy, which shows that the collaborative training in the post-stabilization stage is better than the self-training in the initial startup stage, but the accuracy is lower than that of the ISM method using only the active qualitative strategy, which shows that active learning to correctly label samples can improve the sample distribution and improve the accuracy to a greater extent.

[0088] The above-described embodiments are used to illustrate the present invention, not to limit the present invention, so changes in the exemplified values ​​or replacement of equivalent elements should still fall within the scope of the present invention.

[0089] From the above detailed description, it can be understood by those skilled in the art that the present invention can indeed achieve the aforementioned objectives and is in compliance with the provisions of the Patent Law.

[0090] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they are aware of the basic creative concepts. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention. The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. It should be pointed out that any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

[0091] It should be noted that the above description of the relevant process is only for example and explanation, and does not limit the scope of application of this specification. For those skilled in the art, various modifications and changes can be made to the process under the guidance of this specification. However, these modifications and changes are still within the scope of this specification.

[0092] The basic concepts have been described above. Obviously, for those of ordinary skill in the art who have read this application, the above invention disclosure is only for example and does not constitute a limitation of this application. Although not explicitly stated here, those of ordinary skill in the art may make various modifications, improvements and amendments to this application. Such modifications, improvements and amendments are suggested in this application, so such modifications, improvements and amendments still belong to the spirit and scope of the exemplary embodiments of this application.

[0093] At the same time, the present application uses specific words to describe the embodiments of the present application. For example, "one embodiment", "an embodiment", and / or "some embodiments" refer to a certain feature, structure or characteristic related to at least one embodiment of the present application. Therefore, it should be emphasized and noted that "one embodiment" or "an embodiment" or "an alternative embodiment" mentioned twice or more in different positions in this specification does not necessarily refer to the same embodiment. In addition, some features, structures or characteristics in one or more embodiments of the present application can be appropriately combined.

[0094] In addition, it will be understood by those skilled in the art that various aspects of the present application may be illustrated and described by a number of patentable categories or situations, including any new and useful process, machine, product or combination of substances, or any new and useful improvement thereof. Therefore, various aspects of the present application may be implemented entirely by hardware, entirely by software (including firmware, resident software, microcode, etc.), or by a combination of hardware and software. The above hardware or software may be referred to as a "unit", "module" or "system". In addition, various aspects of the present application may take the form of a computer program product embodied in one or more computer-readable media, wherein computer-readable program code is contained therein.

[0095] The computer program code required for the operation of each part of the present application can be written in any one or more programming languages, including object-oriented programming languages ​​such as Java, Scala, Smalltalk, Eiffel, JADE, Emerald, C++, C#, VB.NET, Python, etc., conventional procedural programming languages ​​such as C programming language, VisualBasic, Fortran2103, Perl, COBOL2102, PHP, ABAP, dynamic programming languages ​​such as Python, Ruby and Groovy, or other programming languages. The program code can be run entirely on the user's computer, or run on the user's computer as an independent software package, or run partially on the user's computer and partially on the remote computer, or run entirely on the remote computer or server. In the latter case, the remote computer can be connected to the user's computer through any network form, such as a local area network (LAN) or a wide area network (WAN), or connected to an external computer (e.g., via the Internet), or in a cloud computing environment, or used as a service such as software as a service (SaaS).

[0096] In addition, unless explicitly stated in the claims, the order of the processing elements and sequences described in this application, the use of alphanumeric characters, or the use of other names are not intended to limit the order of the processes and methods of this application. Although the above disclosure discusses some invention embodiments that are currently considered useful through various examples, it should be understood that such details are only for illustrative purposes, and the attached claims are not limited to the disclosed embodiments. On the contrary, the claims are intended to cover all modifications and equivalent combinations that are consistent with the essence and scope of the embodiments of this application. For example, although the implementation of the various components described above can be embodied in a hardware device, it can also be implemented as a pure software solution, for example, installation on an existing server or mobile device.

[0097] Similarly, it should be noted that in order to simplify the description disclosed in this application and thus help understand one or more embodiments of the invention, in the above description of the embodiments of the application, multiple features are sometimes combined into one embodiment, figure or description thereof. However, this method of the application should not be interpreted as reflecting the intention that the claimed object requires more features than those explicitly stated in each claim. On the contrary, the subject of the invention should have fewer features than the above single embodiment.

Claims

1. A two-stage semi-supervised transformer fault diagnosis method incorporating active learning, characterized in that: include: Step 1: Get DGA data; Step 2: Preprocess the DGA data; Step 3: Divide the preprocessed DGA data into training set S T and validation set S V , the training set is divided into the initial labeled sample set S I and the unlabeled sample set S U ; Step 4: Use one-to-many to initialize the sample set S I On training n FLDA two classifiers, obtain the unlabeled sample set S U All samples in n discriminant scores and predict the unlabeled sample set S U and validation set S V The categories of all samples in S V The accuracy rate of Step 5: m In the iteration, from S U Select samples U with large uncertainty m And manually mark the real label and add it to S I In the middle, the sample C with high confidence m Add the predicted labels to S I In the new labeled training set S L ( m ); Step 6: Use S L ( m ) Re-establish the FLDA model and re-predict the updated S U All samples in n discriminant scores and predict the updated unlabeled sample set S U and validation set S V The categories of all samples in S U and S V The accuracy rate of Step 7: Observe the m The iteration and m -1 iteration in S V Are the samples marked by the above predictions consistent? If so, the iteration stops. Otherwise, the iteration stops until m The upper limit of the value, the iteration ends; Otherwise, return to step 5 to proceed m +1 iteration; The number of iteration stops is counted as m 0; Step 8: m 0+ k The training set S is updated in the last iteration of the initial startup iteration. I A collaborative semi-supervised model is established. By establishing the FLDA model, the inter-class discreteness is maximized and the intra-class discreteness is minimized respectively, and different discrimination rules are formed to construct two different classifiers; that is, two identical initial labeled training sets, unlabeled training set data and validation set S are initialized. V ; S I1 =S I2 =S I , S U1 =S U2 =S U , S V1 =S V2 =S V ; Using S I1 and S I2 Two different classifiers C1 and C2 derived from FLDA are established respectively to obtain the unlabeled sample sets S U1 and S U2 All samples in n discriminant scores and predict the unlabeled sample set S U1 and S U2 The categories of all samples in the validation set S V1 and S V2 The categories of all samples in S V1 and S V2 The accuracy rate of Step 9: According to the semi-supervised quantitative selection strategy, classifiers C1 and C2 are respectively trained from the unlabeled training set S U1 and S U2 Select the sample Q1 with high confidence ( m 0+ k ) and Q2( m 0+ k ) and add it to S after assigning it a label I1 and S I2 In the new labeled training set S L1 ( m ) and S L2 ( m );S L1 ( m )=S I1 +Q2( m 0+ k ), S L2 ( m )=S I2 +Q1( m 0+ k ); Step 10: Use the new labeled training set S L1 ( m ) and S L2 ( m ) Reconstruct two different classifiers C1 and C2 to re-predict S U1 and S U2 All samples in n discriminant scores and predict the unlabeled sample set S U1 and S U2 The categories of all samples in the validation set S V1 and S V2 The categories of all samples in S V1 and S V2 The accuracy rate on the final S V1 and S V2 The average accuracy of Step 11: If satisfied k The iteration stops when the upper limit value of Otherwise, return to step 8 to proceed m 0+ k +1 iteration.

Images