Transformer fault diagnosis method and system based on pseudo-label-driven sparse deep belief network

Abstract

The invention discloses a sparse deep belief network transformer fault diagnosis method and system based on pseudo label driving. The method comprises the following steps: initializing network structure parameters, and updating a network by using a differential space optimization algorithm; obtaining a new sample label by adopting a pseudo label strategy, and circularly optimizing a network structure; and diagnosing the transformer fault in real time according to the optimal parameter. The system comprises an initial optimization module, a pseudo tag driving module, a loop optimization module and an offline / online diagnosis module. When the network structure is optimized, a sparse regular term based on a generalized Gaussian function is added, which is very beneficial for suppressing an overfitting phenomenon in network training, meanwhile, a pseudo-label is used for guiding a learning strategy, the confidence level of each pseudo-label sample is calculated, and a sample with a relatively high confidence level is selected and added into a training set to circularly optimize the network, so that the network structure is optimized. Therefore, the influence of false label noise is reduced. Therefore, the method can be effectively applied to the fault diagnosis of the key components of the transformer, the diagnosis accuracy is improved, and the possibility of misjudgment is reduced.

Description

Technical Field

[0001] The present invention belongs to the field of transformer fault diagnosis and monitoring, and in particular relates to a transformer fault diagnosis method and system based on a pseudo-label driven sparse deep belief network (DBN). Background Art

[0002] Transformers are a vital component of the power system, used to convert electrical energy between different voltage levels. Their reliability and stability directly affect the operating efficiency and safety of the entire power system. Given that transformers have a sophisticated internal structure and are usually exposed to harsh operating environments, they are inevitably subject to a variety of fault challenges during long-term operation, making fault warning and diagnosis mechanisms the key to preventing major accidents, reducing unplanned downtime, and reducing maintenance costs.

[0003] Transformer faults can be roughly divided into two categories: thermal faults and electrical faults. Thermal faults are mainly caused by overheating, such as partial discharge, winding short circuit, etc. Electrical faults include insulation breakdown, flashover, etc. These faults pose a serious threat to the stable operation of the transformer.

[0004] In recent years, with the development of Internet of Things (IoT), artificial intelligence (AI) and big data technology, transformer fault diagnosis is moving towards a more intelligent, automated and predictive direction. For example, using machine learning algorithms to analyze historical data can build a more accurate fault prediction model; at the same time, the integration of IoT technology realizes remote real-time monitoring and instant data communication, improving fault response speed and maintenance efficiency.

[0005] Among many methods, the deep belief network (DBN) composed of multiple restricted Boltzmann machines (RBMs) is widely used in fault diagnosis due to its effectiveness in processing substation process data. However, most DBN-based fault diagnosis methods assume that there is enough labeled data available, but in practice, labeling fault data is labor-intensive. Especially for some high-reliability equipment, such as winding equipment, most of the data is unlabeled because it is very difficult and costly to manually label data. Therefore, how to effectively mine useful information from monitoring data and make full use of these unlabeled data to improve the fault recognition rate is a challenging problem. In addition, the parameters in DBN have a great influence on the fault diagnosis results, but the process of obtaining the optimal parameter values ​​is extremely complex and difficult. Existing optimization algorithms, such as particle swarm PSO, ， Although it is effective to a certain extent, it is easy to fall into the local optimal solution, thus limiting the search accuracy and efficiency.

[0006] Therefore, developing more efficient and global parameter optimization strategies to overcome the limitations of existing algorithms is of great significance for improving the practical application effect of DBN in transformer fault diagnosis. Summary of the invention

[0007] In view of the above problems such as small sample size and parameter optimization, the present invention proposes a transformer fault diagnosis method and system based on a pseudo-label driven sparse deep belief network. The proposed method can be directly applied to the fault classification of the monitoring signals of key components of the transformer, and the method does not involve any prior knowledge of signal processing. For comparison, several popular fault diagnosis methods are also used to achieve fault classification. The results show that the proposed method can achieve a higher fault recognition rate.

[0008] In the first aspect, a transformer fault diagnosis method based on a pseudo-label driven sparse deep belief network specifically includes:

[0009] (1) First, the six parameters of the generalized Gaussian deep belief network structure are initialized, and the first round of network structure parameters are updated through the differential space optimization algorithm based on the existing transformer fault parameter data samples;

[0010] (2) Obtain the labels of new samples according to the pseudo-label strategy, and cyclically update the network structure parameters based on the new samples through the differential space optimization algorithm;

[0011] (3) Obtain transformer fault diagnosis results in real time based on the optimal network structure parameters.

[0012] Furthermore, the six parameters of the network structure include the number of neurons in the first hidden layer (k1), the number of neurons in the second hidden layer (k2), the number of neurons in the third hidden layer (k3), the learning rate, the regularization coefficient of the sparse term, and the variance of the regularization term.

[0013] Furthermore, the differential space optimization algorithm specifically includes.

[0014] 1) Initialize the six parameter values ​​according to the following formula

[0015]

[0016] in, represents the initial value of the i-th spatial point (i=1,2,…,M) at the j-th dimension (j=1,2,…,d); M is the number of spatial points, and d is the number of differential spaces (d=4); and are the minimum and maximum values ​​allowed in the parameter solution space respectively; r1 is a random number between 0 and 1.

[0017] 2) Calculate the value of the adaptability function, that is, the cross-validation classification error, and then find the global optimal value GB and the average optimal value MG based on the fitness value, and update the spatial point positions of the four differential spaces according to the following formula.

[0018] space1=X i +α i ×(β i ×GB-γ i ×MG i )

[0019] space2=GB+α i ×(β i ×X i -γ i ×MG i )

[0020] space3=MG i +α i ×(β i ×X i -γ i ×GB)

[0021]

[0022] in, is the set of solutions of the ith subspace, k = 1, 2, …, d, r2 is a random number between 0 and 1; β i and γ i is a random integer of 0 or 1. α i Randomly select from:

[0023]

[0024] Among them, r3 is a random number between 0 and 1; δ and ε represent random integers distributed in the interval [0,1].

[0025] 3) After updating the position of the differential space point, recalculate the adaptive function values ​​of all space points to determine whether the maximum number of iterations has been reached. If reached, output the optimal parameter value and the global optimal solution. If not, find the global optimal value GB and the average value MG again according to the new fitness value, and continue iterating.

[0026] Furthermore, the generalized Gaussian deep belief network specifically includes three RBM modules, each of which includes a set of visible layers V = (v1, v2, ..., vn) and hidden layers h = (h1, h2, ..., hm). The bidirectional full connection between layers is realized through weights Wij, and the units in the layers are not connected.

[0027] The improvement of the generalized Gaussian deep belief network lies mainly in its training and optimization parameter process. As follows:

[0028] 1) Define the generalized Gaussian term as:

[0029]

[0030] Where l is the variance of the generalized Gaussian regularization term and Γ() is the Gamma function.

[0031] 2) Define the sparse generalized Gaussian regularization term as:

[0032]

[0033] Among them, refers to the ‖‖0L0 norm, p j is the activation function of the hidden layer of the deep belief network.

[0034] 3) Update and optimize the parameters of the generalized Gaussian deep belief network according to the following formula.

[0035]

[0036] Among them, W ij is the RBM model weight, b j is the bias coefficient of the hidden layer, η is the learning rate, λ is the regularization factor, v and h represent the visible unit and hidden unit of the RBM model respectively, <> d ,<> r The expectation of the training data and the mathematical expectation of the distribution of the reconstructed model parameters.

[0037] Furthermore, the pseudo-label strategy specifically includes:

[0038] 1) The new unlabeled transformer data sample obtains the belonging probability of each category according to the following formula;

[0039]

[0040] in, is the i-th new unlabeled sample, f GG-DBN is the mapping function of the generalized Gaussian deep belief network, k = 1, 2, ..., K, K is the number of transformer fault states, and P() is the attribution probability.

[0041] 2) Calculate the confidence level w of each fault state of each new unlabeled sample according to the following formula: i k ;

[0042]

[0043] in, Hi is the information entropy.

[0044] 3) According to the above formula, the belonging class with high confidence level is selected as the pseudo-labeled sample label. Then, the generalized Gaussian deep belief network is re-optimized using the pseudo-labeled samples together with the existing samples.

[0045] In the second aspect, a transformer fault diagnosis system based on a pseudo-label driven sparse deep belief network specifically includes:

[0046] Initial optimization module, pseudo-label driving module, loop optimization module, offline / online diagnosis module, the system completes the above contents and tasks.

[0047] The initialization module is used to initialize the parameters of the generalized Gaussian deep belief network model;

[0048] The pseudo-label driving module is applied to assign the transformer fault status category label with the highest confidence level to the new sample;

[0049] The loop optimization module is used to loop optimize the parameters of the generalized Gaussian deep belief network according to the difference space algorithm when existing data or new samples are obtained;

[0050] The offline / online diagnosis module is applied to perform real-time diagnosis on the trained and optimized generalized Gaussian deep belief network.

[0051] Compared with the prior art, the present invention has the following beneficial effects:

[0052] 1) By introducing a generalized Gaussian deep belief network and combining it with a differential space optimization algorithm for parameter optimization, the present invention can more accurately capture the complex characteristics of transformer fault data, thereby achieving accurate diagnosis of transformer fault status. Compared with traditional methods, the diagnostic accuracy of the present invention has been significantly improved.

[0053] 2) By using the pseudo-label strategy, the present invention can make full use of the unlabeled transformer fault data and continuously improve the generalization ability of the model through an iterative optimization process. Specifically, by calculating the confidence level of each pseudo-label sample, samples with higher confidence levels can be screened out and added to the training set to cyclically optimize the network. Not only does it make full use of the information of unlabeled data, it also effectively reduces the impact of pseudo-label noise on model training, further improving the accuracy and stability of the model.

[0054] 3) The present invention can be effectively applied to the fault diagnosis of key components of transformers, improving the accuracy of diagnosis and reducing the possibility of misjudgment. The fault diagnosis experiment conducted on a certain type of transformer showed that the average recognition rate of this method reached 96.33%, which not only exceeded some DBN-based modeling methods, but also outperformed other intelligent methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is a flow chart of a transformer fault diagnosis method based on a pseudo-label driven sparse deep belief network (DBN) of the present invention;

[0056] Figure 2 is a schematic diagram of the RBM in the present invention;

[0057] Figure 3 is a schematic diagram of a DBN in the present invention;

[0058] Figure 4 is a flow chart of the differential space optimization algorithm in the present invention;

[0059] Figure 5 It is a training and optimization flow chart of the generalized Gaussian deep belief network in the present invention;

[0060] Figure 6 It is a schematic diagram of a transformer fault diagnosis system based on a pseudo-label driven sparse deep belief network (DBN) of the present invention;

[0061] Figure 7 is the distribution of eligible points of the parameters during the difference space optimization.

[0062] Figure 8 is the confidence level of some pseudo-labeled samples.

[0063] Fig. 9 Confusion matrix for the recognition rate using the proposed method. DETAILED DESCRIPTION

[0064] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments, but the protection scope of the present invention should not be limited thereto.

[0065] This embodiment discloses a transformer fault diagnosis method and system based on a pseudo-label driven sparse deep belief network (DBN), which combines a differential space optimization algorithm, a sparse regularization technique, and a pseudo-label learning strategy to improve the accuracy and efficiency of transformer fault diagnosis.

[0066] like Figure 1 As shown, a transformer fault diagnosis method based on a pseudo-label driven sparse deep belief network includes the following steps:

[0067] S1. Initialization and the first round of parameter update:

[0068] Initialize six parameters of the generalized Gaussian deep belief network structure, including the number of neurons in the first hidden layer, the number of neurons in the second hidden layer, the number of neurons in the third hidden layer, the learning rate, the regularization coefficient of the sparse term, and the variance of the regularization term; wherein the generalized Gaussian deep belief network includes three RBM modules. RBM is a probabilistic modeling method based on energy function, which can be regarded as an undirected graph model, such as Figure 2 As shown, each RBM module consists of a set of visible layers V = (v1, v2, ..., vn) and hidden layers h = (h1, h2, ..., hm). The layers are fully connected in both directions through weights Wij, and the units in the layers are not connected.

[0069] For a given set of layer states (v,h), its energy function is defined as:

[0070]

[0071] Among them, θ = [Wij, ai, bj] represents the parameter set of RBM; a i is the bias of the visible layer; b j is the bias of the hidden layer; n is the number of visible layer nodes; m is the number of hidden layer nodes.

[0072] According to the energy function, the joint probability distribution of (v,h) is:

[0073]

[0074] in,

[0075] The probability distribution of the visible layer v is:

[0076]

[0077] From the perspective of conditional probability, the activation probability of hidden nodes and the reconstruction activation probability of visible nodes can be obtained as:

[0078]

[0079] Where sigmoid(x)=1 / (1+e -x ).

[0080] like Figure 3 As shown in Figure 1, DBN is formed by stacking multiple RBMs, and its training process is mainly divided into two steps: pre-training and fine-tuning. The first step is to pre-train each RBM to ensure that as much useful information as possible can be retained when mapping data to different spaces. To train the RBM, the parameters are adjusted by maximizing the log-likelihood on the training set. Therefore, the parameter W ij 、a i and bj The partial derivative of is calculated as:

[0081]

[0082] Among them, <·>d and <·>m represent the mathematical expectations of the distribution of training data and the distribution of original model parameters, respectively. Due to the high computational complexity of visible and hidden layers, the contrastive divergence (CD) algorithm is used for gradient estimation. Therefore, the parameter update is as follows:

[0083]

[0084] Where, ΔW ij ,Δa i ,Δb j represents the parameter change during the training optimization process; η is the learning rate; <> r Represents the expectation of the updated parameters.

[0085] After updating the parameters, the back propagation algorithm is then used to fine-tune the network parameters. A back propagation network with a softmax activation function is set at the last layer of the DBN, and the output of the last RBM is used as the input feature vector of the back propagation network to train the fault classifier in a supervised manner. In the supervised training phase, the back propagation network propagates error information from top to bottom to each RBM and fine-tunes the entire DBN network.

[0086] The DBN composed of multiple RBMs infers the distribution of data samples and realizes fault diagnosis by establishing a joint distribution between observed data and labeled data.

[0087] The training and optimization process of the generalized Gaussian deep belief network parameters is as follows:

[0088] ①Define the generalized Gaussian term as:

[0089]

[0090] Where l is the variance of the generalized Gaussian regularization term and Γ() is the Gamma function.

[0091] ②Define the sparse generalized Gaussian regularization term as:

[0092]

[0093] Among them, refers to the ‖‖0L0 norm, p j is the activation function of the hidden layer of the deep belief network.

[0094] ③ Update and optimize the parameters of the generalized Gaussian deep belief network according to the following formula:

[0095]

[0096] Among them, W ij is the RBM model weight, b j is the bias coefficient of the hidden layer, η is the learning rate, λ is the regularization factor, v and h represent the visible unit and hidden unit of the RBM model respectively, <> d ,<> r They represent the expectation of the training data and the mathematical expectation of the distribution of the reconstructed model parameters respectively.

[0097] According to the existing transformer fault parameter data samples, the network structure parameters of the first round are updated through the differential space optimization algorithm.

[0098] Specific steps are as follows Figure 4 As shown, including:

[0099] S1.1 initializes six parameter values, the formula is as follows:

[0100]

[0101] In the formula, represents the initial value of the i-th spatial point (i=1,2,…,M) at the j-th dimension (j=1,2,…,d); M is the number of spatial points, and d is the number of differential spaces (d=4); and are the minimum and maximum values ​​allowed in the parameter solution space respectively; r1 is a random number between 0 and 1;

[0102] S2.2 calculates the value of the adaptability function (cross-validation classification error), finds the global optimal value GB and the average optimal value MG according to the fitness value, and updates the spatial point positions of the four differential spaces. The formula is as follows:

[0103] space1=X i +α i ×(β i ×GB-γ i ×MG i )

[0104] space2=GB+α i ×(β i ×X i -γ i ×MG i )

[0105] space3=MG i +α i ×(β i ×X i -γ i ×GB)

[0106]

[0107] In the formula, is the set of solutions of the ith subspace, k = 1, 2, …, d; r2 is a random number between 0 and 1; β i and γ i is a random integer of 0 or 1, α i Randomly select from:

[0108]

[0109] Where r3 is a random number between 0 and 1; δ and ε represent random integers distributed in the interval [0,1];

[0110] S3.3 repeats the process of updating the position of the differential space point and calculating the adaptive function value of the space point until the maximum number of iterations is reached or the stopping condition is met, and outputs the optimal parameter value and the global optimal solution.

[0111] S2. Pseudo labeling strategy and loop parameter update:

[0112] According to the pseudo-label strategy, the label of the new sample is obtained, wherein the pseudo-label strategy refers to calculating the attribution probability and confidence level of the new sample in the generalized Gaussian deep belief network, and selecting the attribution class with a high confidence level as the sample label of the pseudo-label;

[0113] According to the new samples and pseudo labels, the network structure parameters are cyclically updated through the differential space optimization algorithm to optimize the network performance.

[0114] Specific as Figure 5 As shown, the steps are as follows:

[0115] S2.1 Calculate the probability P() and confidence level w of the new sample in the generalized Gaussian deep belief network i k , the formula is as follows:

[0116]

[0117] in, is the i-th new unlabeled sample, f GG-DBN is the mapping function of the generalized Gaussian deep belief network, k = 1, 2, ..., K, K is the number of transformer fault states, H i is the information entropy,

[0118]

[0119] S2.2 selects the belonging class with high confidence level as the pseudo-labeled sample label.

[0120] S2.3 uses pseudo-labeled samples together with existing samples to re-optimize the generalized Gaussian deep belief network.

[0121] S3. Real-time fault diagnosis:

[0122] According to the optimal network structure parameters, transformer fault diagnosis results are obtained in real time.

[0123] See also Figure 6 , Figure 6 The present invention is a schematic diagram of a transformer fault diagnosis system based on a pseudo-label driven sparse deep belief network (DBN), as shown in the figure, comprising: an initial optimization module, a pseudo-label driving module, a cyclic optimization module and an offline / online diagnosis module; wherein,

[0124] Initialization module, used to initialize the parameters of the generalized Gaussian deep belief network;

[0125] A pseudo-label driving module is used to assign a transformer fault state category label with the highest confidence level to a new sample;

[0126] A loop optimization module is used to loop optimize the parameters of the generalized Gaussian deep belief network according to the difference space algorithm when existing data or new samples are obtained;

[0127] The offline / online diagnosis module is used to perform real-time diagnosis on the trained and optimized generalized Gaussian deep belief network.

[0128] In order to better illustrate the beneficial effects of the method of the present invention, we conduct comparative tests on the same type of transformers in a substation to verify the effect. The transformer faults in the selected area include cooling system failure, partial discharge, core failure, winding short circuit, and oil quality deterioration. A total of 11,643 samples were collected, of which 1,800 were labeled and 9,843 were unlabeled. Therefore, this dataset is suitable for developing pseudo-label-guided sparse DBN learning methods for fault diagnosis of key components of transformers. In addition, the labeled samples have 6 health conditions, including one normal condition and five fault conditions. Under each health condition, 250 samples are randomly selected as training sets and 50 samples are selected as test sets.

[0129] The maximum number of iterations of the difference space optimization algorithm is set to 100. For the number of hidden layer neurons, i.e., k1, k2, and k3, the range in the solution space is [1, 100], and for η, σ, and λ, the range in the solution space is [0, 1].

[0130] For visualization, Figure 7The distribution of qualified points for parameters k1 and k2 during the difference space optimization is shown. From this three-dimensional space diagram, it can be seen that the qualified space points are able to obtain information related to the solution during the search process and gradually gather near the optimal solution. To be precise, in the initial stage, the qualified points are randomly scattered in the analysis space. Then, through the update of the four subspaces proposed above, the difference space algorithm makes the group gradually gather towards the optimal solution area. Finally, the optimal solution for configuring the network parameters is found, namely κ1=53,κ2=75,κ3=57,η=0.1326,σ=0.4617,λ=0.0005.

[0131] Using the optimized generalized Gaussian DBN, the next step is to obtain pseudo labels for unlabeled samples, and pseudo-labeled samples with high confidence levels will be retained. The figure below shows the confidence levels of some pseudo-labeled samples. It can be seen that the confidence levels of different samples vary greatly. For example, the confidence level of sample number 14 is 0.79, while the confidence level of sample number 35 is 0.97. In order to determine the threshold of the confidence level, some validation tests have been conducted, and it was found that the classification error is minimal at a confidence level of 0.9. Therefore, pseudo-labeled samples with a confidence level greater than or equal to 0.9 will be added to the training set in order to retrain the network. At the same time, the network configuration parameters need to be re-optimized in the above manner, such as Figure 8 shown.

[0132] exist Fig. 9 In the confusion matrix of , the diagonal elements represent the prediction accuracy, while the off-diagonal elements represent the errors that the classifier did not predict correctly. The higher the diagonal value of the confusion matrix, the better the model prediction performance. The following figure shows the confusion matrix of the recognition rate using the proposed method. Label 1 in this figure represents the normal situation, while the next 5 labels correspond to 5 fault situations. Fig. 9 It can be seen that the proposed method has high recognition rates for tags 1, 2, 3, 4, and 6, which are 98%, 96%, 100%, 100%, and 98%, respectively.

Claims

1. A transformer fault diagnosis method based on pseudo-label driven sparse deep belief network, characterized in that: Includes steps: S1. Initialization and the first round of parameter update: Initialize six parameters of the generalized Gaussian deep belief network, including the number of neurons in the first hidden layer (k1), the number of neurons in the second hidden layer (k2), the number of neurons in the third hidden layer (k3), the learning rate, the regularization coefficient of the sparse term, and the variance of the regularization term; Using the existing transformer fault parameter data samples and the differential space optimization algorithm, the network parameters are updated to complete the first round of optimization; S2. Pseudo label acquisition and loop parameter update: Obtain the label of the new sample according to the pseudo-label strategy, and cyclically update the network structure parameters again through the differential space optimization algorithm based on the new sample until the optimal state is reached; wherein the label of the new sample is obtained in the process of cyclically updating the network structure parameters through the differential space optimization algorithm based on the new sample; S3. Real-time fault diagnosis: According to the optimal network structure parameters, transformer fault diagnosis results are obtained in real time.

2. The transformer fault diagnosis method based on pseudo-label driven sparse deep belief network according to claim 1 is characterized in that: The step S1. specifically comprises the following steps: S1.1 initializes six parameter values, the formula is as follows: In the formula, represents the initial value of the i-th spatial point (i=1,2,…,M) at the j-th dimension (j=1,2,…,d); M is the number of spatial points, and d is the number of differential spaces (d=4); and are the minimum and maximum values ​​allowed in the parameter solution space respectively; r1 is a random number between 0 and 1; S2.2 calculates the value of the adaptability function (cross-validation classification error), finds the global optimal value GB and the average optimal value MG according to the fitness value, and updates the spatial point positions of the four differential spaces. The formula is as follows: space1=X i +a i ×(β i ×GB-g i ×MG i ) space2=GB+a i ×(β i ×X i -c i ×MG i ) space3=MG i +a i ×(β i ×X i -c i ×GB) In the formula, is the set of solutions of the ith subspace, k = 1, 2, …, d; r2 is a random number between 0 and 1; β i and γ i is a random integer of 0 or 1, α i Randomly select from: Where r3 is a random number between 0 and 1; δ and ε represent random integers distributed in the interval [0,1]; S1.3 After updating the position of the differential space point, recalculate the adaptive function values ​​of all space points to determine whether the maximum number of iterations has been reached; If it is reached, the optimal parameter value and the global optimal solution are output; if the maximum number of iterations is not reached, the global optimal value GB and the average value MG are found again according to the new fitness value, and the iteration continues.

3. The transformer fault diagnosis method based on pseudo-label driven sparse deep belief network according to claim 1 is characterized in that: The generalized Gaussian deep belief network includes three RBM modules, each of which includes a set of visible layers and hidden layers. The layers are bidirectionally fully connected through weights, and the units in the layers are not connected.

4. The transformer fault diagnosis method based on pseudo-label driven sparse deep belief network according to claim 3 is characterized in that: The training and optimization process of the generalized Gaussian deep belief network is as follows: 1) Define the generalized Gaussian term as: Where l is the variance of the generalized Gaussian regularization term, Γ() is the Gamma function; 2) Define the sparse generalized Gaussian regularization term as: Among them, refers to the ‖‖0L0 norm, p j is the activation function of the hidden layer of the deep belief network; 3) Update and optimize the parameters of the generalized Gaussian deep belief network. The formula is as follows: Among them, W ij is the RBM model weight, b j is the bias coefficient of the hidden layer, η is the learning rate, λ is the regularization factor, v and h represent the visible unit and hidden unit of the RBM model respectively, < > d ,< > r They represent the expectation of the training data and the mathematical expectation of the distribution of the reconstructed model parameters respectively.

5. The transformer fault diagnosis method based on pseudo-label driven sparse deep belief network according to claim 1 is characterized in that: The pseudo label strategy includes: S2.1 Calculate the probability P() and confidence level of a new sample in the generalized Gaussian deep belief network The formula is as follows: in, is the i-th new unlabeled sample, f GG-DBN is the mapping function of the generalized Gaussian deep belief network, k = 1, 2, ..., K, K is the number of transformer fault states, H i is the information entropy, S2.2 selects the belonging class with high confidence level as the pseudo-labeled sample label. S2.3 uses pseudo-labeled samples together with existing samples to re-optimize the generalized Gaussian deep belief network.

6. A transformer fault diagnosis system based on pseudo-label driven sparse deep belief network, characterized in that: include: Initialization module, used to initialize the parameters of the generalized Gaussian deep belief network; A pseudo-label driving module is used to assign a transformer fault state category label with the highest confidence level to a new sample; A loop optimization module is used to loop optimize the parameters of the generalized Gaussian deep belief network according to the difference space algorithm when existing data or new samples are obtained; The offline / online diagnosis module is used to perform real-time diagnosis on the trained and optimized generalized Gaussian deep belief network.

Images