A fault diagnosis method for oil-immersed transformer based on LGEO-gcForest

The LGEO-gcForest model and Golden Eagle optimization algorithm optimize the multi-grained scanning window size and the maximum cascade number of cascading forests, which solves the problem of the large data demand, poor interpretation, high complexity and unbalanced data in transformer fault diagnosis. The problem has achieved higher diagnostic accuracy and stability.

CN117708696BActive Publication Date: 2024-11-12STATE GRID HUBEI EXTRA HIGH VOLTAGE CO +2
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202311632988.9
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-11-29
Publication Date
2024-11-12
Estimated Expiration
2043-11-29

AI Technical Summary

Technical Problem

The existing gcForest model has problems such as large data demand, poor interpretation, high model complexity, difficulty in adjusting hyperparameters and data imbalance in transformer fault diagnosis, resulting in insufficient diagnostic accuracy and stability.

Method used

Using the fault diagnosis method based on LGEO-gcForest, the non-coding ratio of dissolved gas in oil is calculated, and the Golden Eagle optimization algorithm is used to optimize the multi-particle scanning window size and the maximum cascade number of cascades of cascade forests is constructed to improve fault diagnosis. accuracy and stability.

Benefits of technology

Improve the accuracy and stability of transformer fault diagnosis, enhance the generalization performance of the model, overcome the random volatility of the gcForest model, especially in the case of insufficient sample size and unbalanced data, it can more accurately identify transformer faults. .

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117708696B_ABST
    Figure CN117708696B_ABST
Patent Text Reader

Abstract

The present invention belongs to the technical field of transformer fault diagnosis, and relates to an oil-immersed transformer fault diagnosis method based on LGEO-gcForest. According to the data sample of dissolved gas in oil of the transformer, the non-coded ratio of dissolved gas in oil is calculated, the non-coded ratio is standardized, and a training set and a test set are divided; the multi-granularity scanning window size of the gcForest model and the maximum number of cascades allowed by the cascade forest are used as optimization parameters, and the fault diagnosis accuracy of the gcForest model is used as fitness; the gcForest model is optimized using the Golden Eagle optimization algorithm, and the optimal parameters are returned; the LGEO-gcForest fault diagnosis model is constructed according to the returned optimal parameters to obtain the transformer fault diagnosis result. The present invention can improve the accuracy of transformer fault diagnosis by optimizing the key parameters of the gcForest model.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the technical field of transformer fault diagnosis, and in particular relates to an oil-immersed transformer fault diagnosis method based on LGEO-gcForest. Background Art

[0002] Transformer failure will endanger the safe and stable operation of the entire power system. Transformer fault diagnosis methods can analyze equipment status information and are the key to ensuring reliable and efficient operation of equipment. Therefore, quickly and accurately identifying the fault type of the transformer and performing timely maintenance work can provide important guarantees for the normal operation of the power system.

[0003] gcForest is a deep forest model that consists of multiple cascaded random forests. There have been reports on using the gcForest model for oil-immersed transformer fault diagnosis. However, there may be some defects and challenges in using the gcForest model for transformer fault diagnosis, as shown below:

[0004] 1) Requires a large amount of data: gcForest models usually require a large amount of data for training, especially when the model is deep. For some applications, it may be difficult to obtain enough fault samples, which may lead to model overfitting or performance degradation.

[0005] 2) Explanation limitations: Deep forest models lose the interpretability of a single tree to some extent because they consist of multiple random forests, making it difficult to intuitively interpret the model’s decision-making process. This may be challenging for fields that require explanation of model decisions, such as transformer fault diagnosis.

[0006] 3) Model complexity: The gcForest model has a high complexity, including multiple random forest cascades. This may increase the difficulty of model training and tuning, requiring more computing resources and time.

[0007] 4) Hyperparameter Tuning: Configuring the hyperparameters of the gcForest model may require a lot of experimentation and tuning. Incorrect hyperparameter selection may lead to performance degradation.

[0008] 5) Data imbalance: In actual transformer fault diagnosis data, the samples of normal state and fault state may be unbalanced. This may cause the model to favor more common classes and ignore rare fault conditions. Summary of the invention

[0009] In order to overcome the shortcomings of the gcForest model in transformer fault diagnosis, the present invention provides an oil-immersed transformer fault diagnosis method based on LGEO-gcForest.

[0010] The present invention provides a method for diagnosing oil-immersed transformer faults based on LGEO-gcForest, comprising the following steps:

[0011] Step S1: Calculate the non-coded ratio of dissolved gas in oil according to the data sample of dissolved gas in oil of the transformer, and mark the fault type;

[0012] Step S2: standardize the non-coding ratio and divide the standardized data into a training set and a test set;

[0013] Step S3: Initialize two parameters of the gcForest model: the multi-granularity scanning window size and the maximum number of cascades allowed by the cascade forest, and then input the training set into the constructed gcForest model;

[0014] Step S4: taking the multi-granularity scanning window size of the gcForest model and the maximum number of cascades allowed by the cascade forest as optimization parameters, and taking the fault diagnosis accuracy of the gcForest model as fitness; using the Golden Eagle optimization algorithm to optimize the gcForest model, and returning the optimal parameters, including the optimal multi-granularity scanning window size and the optimal maximum number of cascades allowed by the cascade forest;

[0015] Step S5: construct an LGEO-gcForest fault diagnosis model according to the returned optimal parameters, input the test set, obtain the transformer fault diagnosis result, and verify the accuracy; convert the dissolved gas data in oil of unknown fault category into a non-coded ratio, input the LGEO-gcForest fault diagnosis model with qualified verification accuracy, and obtain the transformer fault diagnosis result.

[0016] Further optimization, the process of optimizing the gcForest model using the Golden Eagle optimization algorithm is as follows:

[0017] Step S41: Initialize the parameters of the Golden Eagle optimization algorithm, including the number of Golden Eagle populations and the maximum number of iterations, and use PWLCM mapping to initialize the coordinates of the Golden Eagle population;

[0018] Step S42: randomly selecting an optimal hunting position and calculating an attack vector;

[0019] Step S43: determine whether the attack vector is a zero vector; when the attack vector is a zero vector, return to step S42 and recalculate the attack vector; when the attack vector is not a zero vector, update the cruise vector and use the improved Levy flight mechanism to update the position of the golden eagle;

[0020] Step S44: Determine whether the maximum number of iterations is met. If not, return to step S42 and re-iterate. When the maximum number of iterations is met, output the optimal fitness and the corresponding multi-granularity scanning window size and the maximum number of cascades allowed by the cascade forest.

[0021] Further preferably, the coordinates of the golden eagle population are initialized using PWLCM mapping as follows:

[0022]

[0023] in, is the position of the i-th golden eagle at the t-th iteration, is the position of the i-th golden eagle at the t+1th iteration, and p is the control parameter.

[0024] Further preferably, the formula for updating the position of the golden eagle using the improved Levy flight mechanism is as follows:

[0025]

[0026]

[0027] in, is the position of the i-th golden eagle at the t-th iteration, is the step vector of the i-th golden eagle in the t-th iteration, C is a random target point between (0,1), is the best position reached by the golden eagle individual in t iterations, λ is the proportional coefficient of each optimization, α is the step size scaling factor, Levy (β) is the Levy random path, and T is the maximum number of iterations.

[0028] Further preferably, the attack vector is as follows:

[0029]

[0030] Among them, A i is the attack vector of the i-th golden eagle, is the best position reached by the golden eagle individual in the tth iteration.

[0031] Further preferably, the cruise vector is as follows:

[0032]

[0033] Among them, C i is the cruising vector of the i-th golden eagle, c k is the kth element of the cruise vector, n is the number of elements, rand is a random number, A j is the attack vector of the jth golden eagle, a kis the kth element of the attack vector of the jth golden eagle, and d is the constant of the hyperplane equation where the golden eagle is located.

[0034] Further preferably, the non-coding ratio includes: CH 4 / H 2 , C 2 H 4 / C 2 H 2 , C 2 H 4 / C 2 H 6 , C 2 H 2 / (C 1 +C 2 ), H 2 / (H 2 +C 1 +C 2 ), C 2 H 4 / (C 1 +C 2 ), CH 4 / (C 1 +C 2 ), C 2 H 6 / (C 1 +C 2 ), (CH 4 +C 2 H 4 ) / (C 1 +C 2 ) Among them, C 1 It represents a first-order hydrocarbon.

[0035] More preferably, the non-coding ratio is normalized by adjusting the non-coding ratio to a range of [0, 1] using minimum-maximum value normalization.

[0036] The present invention has the following advantages: (1) The LGEO-gcForest fault diagnosis model can improve the accuracy of transformer fault diagnosis. By adopting the position update strategy of the improved Levy flight mechanism and optimizing the key parameters of the gcForest model using the LGEO algorithm, the random volatility of the gcForest model output is overcome, so that the LGEO-gcForest fault diagnosis model has better generalization performance. (2) As the sample size increases, the diagnostic accuracy of the LGEO-gcForest fault diagnosis model is improved and the diagnostic performance is more stable. The collection of mixed fault cases of non-charging and overheating can also be increased to verify the effectiveness of the proposed method. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a flow chart of the oil-immersed transformer fault diagnosis method based on LGEO-gcForest of the present invention;

[0038] Figure 2 It is a schematic diagram of the multi-granularity scanning process;

[0039] Figure 3 Schematic diagram of the multi-cascade forest process. DETAILED DESCRIPTION

[0040] The present invention is further explained in detail below with reference to the examples.

[0041] like Figure 1 As shown in the figure, a fault diagnosis method for oil-immersed transformer based on LGEO-gcForest is shown in the figure. The steps are as follows:

[0042] Step S1: Calculate the non-coded ratio of dissolved gas in oil according to the data sample of dissolved gas in oil of the transformer, and mark the fault type;

[0043] Step S2: The non-coding ratio is normalized to the range of [0, 1] using the minimum-maximum value, and the normalized data is divided into a training set and a test set;

[0044] Step S3: Initialize two parameters of the gcForest model: the multi-granularity scanning window size and the maximum number of cascades allowed by the cascade forest, and then input the training set into the constructed gcForest model;

[0045] Step S4: taking the multi-granularity scanning window size of the gcForest model and the maximum number of cascades allowed by the cascade forest as optimization parameters, and taking the fault diagnosis accuracy of the gcForest model as fitness; using the improved Golden Eagle optimization algorithm to optimize the gcForest model, and returning the optimal parameters, including the optimal multi-granularity scanning window size and the optimal maximum number of cascades allowed by the cascade forest;

[0046] Step S5: construct an LGEO-gcForest fault diagnosis model according to the returned optimal parameters, input the test set, obtain the transformer fault diagnosis result, and verify the accuracy; convert the dissolved gas data in oil of unknown fault category into a non-coded ratio, input the LGEO-gcForest fault diagnosis model with qualified verification accuracy, and obtain the transformer fault diagnosis result.

[0047] The present invention uses non-coded ratios of nine types of dissolved gases in oil as input, including: CH 4 / H 2 , C 2 H 4 / C 2 H 2 , C2 H 4 / C 2 H 6 , C 2 H 2 / (C 1 +C 2 ), H 2 / (H 2 +C 1 +C 2 ), C 2 H 4 / (C 1 +C 2 ), CH 4 / (C 1 +C 2 ), C 2 H 6 / (C 1 +C 2 ), (CH 4 +C 2 H 4 ) / (C 1 +C 2 ) Among them, C 1 It represents a first-order hydrocarbon, such as CH 4 ; C 2 It represents a second-order hydrocarbon, such as C 2 H 6 The non-coding ratio was adjusted to the range of [0, 1] using minimum-maximum normalization, and the standardized data was finally divided into training and test sets in a ratio of 7:3.

[0048] The gcforest model adopted by the present invention is composed of multi-granularity scanning and cascade forest.

[0049] Among them, the multi-granularity cascade forest uses a multi-granularity scanning process to enhance the cascade forest. For processing the original features, the sliding window method is used. Figure 2 As shown in the figure, assume that there is a 400-dimensional original data and a sliding window of 100-dimensional features. For sequence data, the sliding window takes a 100-dimensional feature vector as a feature; a total of 301 feature vectors will be generated. All positive or negative feature vectors extracted from the training samples will be identified as positive or negative instances. These instances will be used to generate class vectors in the cascade structure, and all instances formed by sliding windows of the same size will be used to train completely random tree forests and random forests, and then these class vectors as converted features will be used to generate and connect the cascade structure.

[0050] The architecture of the entire cascade forest model is a cascade structure. Figure 3As shown in Figure 2, each level is constructed using two types of forests: fully random forest and random forest. In the cascade forest, the first classifier ( Figure 3 The first layer in the diagram) is called the "root classifier", which takes in the raw feature data and transforms it into a new set of features. These new features are passed to the second classifier (the second layer in the diagram), and so on, until the last classifier (the nth layer in the diagram). Each classifier is trained based on the output of the previous classifier, so they form a cascade structure.

[0051] The following is the Cascade Forest Processing Process:

[0052] 1. The first forest (root forest) receives the original data as input and trains a basic classifier.

[0053] 2. For each subsequent forest, the input data is the output of the previous forest. Each forest will train a classifier to classify the data.

[0054] 3. For each forest, the classifier will classify the data into two categories: positive and negative. The positive class refers to the data that the classifier considers to be positive examples, and the negative class refers to the data that the classifier considers to be negative examples.

[0055] 4. If a data point is classified as positive class, then it will be passed to the next forest for classification. If a data point is classified as negative class, then it will be discarded and will not be passed to the next forest.

[0056] 5. For each subsequent forest, the output of the classifier is used to train the classifier of the next forest. This cascade structure can effectively reduce the impact of misclassification and noise and improve the accuracy and stability of the model.

[0057] 6. The final output is a weighted average of the outputs of all cascaded forests. The weighting factor is based on the accuracy of each forest.

[0058] The present invention introduces a chaotic mapping mechanism and a Levy flight mechanism to improve the Golden Eagle Optimization Algorithm (LGEO).

[0059] First, the PWLCM chaotic map is used to replace the original initialization method of the Golden Eagle optimization algorithm to initialize the position of the Golden Eagle and initialize the fitness function, as shown below:

[0060]

[0061] in, is the position of the i-th golden eagle at the t-th iteration, is the position of the ith golden eagle at the t+1th iteration, and p is the control parameter. When p∈(0,0.5), , the system is in a chaotic state.

[0062] Then the Golden Eagle position update begins; the position update of the improved Golden Eagle optimization algorithm can be divided into three parts: (1) attack phase; (2) cruise phase; (3) position update.

[0063] (1) Attack phase

[0064] The first calculation of the attack vector A can randomly select an optimal hunting position. When the prey is close to the golden eagle, the golden eagle attacks. The attack vector A is as follows:

[0065]

[0066] Among them, A i is the attack vector of the i-th golden eagle, is the best position reached by the golden eagle individual in the tth iteration.

[0067] When the attack vector A is equal to 0, return to recalculate the attack vector A; when the attack vector A is not equal to 0, enter the cruising phase.

[0068] (2) Cruise phase

[0069] When the golden eagle is far away from its prey, it will circle around the prey. At this time, the n-dimensional cruise vector is located on the tangent hyperplane when the golden eagle is in circular motion. The formula for the cruise vector is as follows:

[0070]

[0071] Among them, C i is the cruising vector of the i-th golden eagle, c k is the kth element of the cruise vector, n is the number of elements, rand is a random number, A j is the attack vector of the jth golden eagle, a k is the kth element of the attack vector of the jth golden eagle, and d is the constant of the hyperplane equation where the golden eagle is located.

[0072] (3) Location Update

[0073] The Golden Eagle's position update depends on the attack vector and the cruise vector.

[0074] In the tth iteration, the step vector of the i-th golden eagle i can be expressed as follows.

[0075]

[0076] in, is the step vector of the i-th golden eagle in the t-th iteration, r 1 and r 2 is a random number between 0 and 1. and are the attack coefficient and the cruise coefficient respectively, and the calculation is as follows:

[0077]

[0078]

[0079] in, and are the initial values ​​of the attack coefficient and the cruise coefficient, which are 0.5 and 1 respectively, and T is the maximum number of iterations.

[0080] ||A i || is the Euclidean norm of the attack vector; ||C i || is the Euclidean norm of the cruise vector, calculated as follows:

[0081]

[0082] In the formula, C j is the cruising vector of the jth golden eagle;

[0083] Finally, the individual position of the golden eagle is updated by combining the improved Levy flight mechanism. The formula is as follows:

[0084]

[0085]

[0086] Among them, C is a random target point between (0,1), is the best position reached by the golden eagle individual in t iterations, λ is the proportional coefficient of each optimization, α is the step size scaling factor, and Levy (β) is the Levy random path.

[0087] The process of optimizing the gcForest model using the improved Golden Eagle optimization algorithm is as follows:

[0088] Step S41: Initialize the parameters of the improved Golden Eagle optimization algorithm, including the number of Golden Eagle populations and the maximum number of iterations, and use PWLCM mapping to initialize the coordinates of the Golden Eagle population;

[0089] Step S42: randomly selecting an optimal hunting position and calculating an attack vector;

[0090] Step S43: Determine whether the attack vector is a zero vector. When the attack vector is a zero vector, return to step S42 and recalculate the attack vector; when the attack vector is not a zero vector, update the cruise vector and use the improved Levy flight mechanism to update the position of the golden eagle;

[0091] Step S44: Determine whether the maximum number of iterations is met. If not, return to step S42 and re-iterate. When the maximum number of iterations is met, output the optimal fitness and the corresponding multi-granularity scanning window size and the maximum number of cascades allowed by the cascade forest.

[0092] In order to further verify the effect of the oil-immersed transformer fault diagnosis method based on multi-source data fusion proposed in this paper, 650 sets of gas data were collected from relevant literature, and 9 characteristic parameters were constructed using the uncoded ratio method: CH 4 / H 2 , C 2 H 4 / C 2 H 2 , C 2 H 4 / C 2 H 6 , C 2 H 2 / (C 1 +C 2 ), H 2 / (H 2 +C 1 +C 2 ), C 2 H 4 / (C 1 +C 2 ), CH 4 / (C 1 +C 2 ), C 2 H 6 / (C 1 +C 2 ), (CH 4 +C 2 H 4 ) / (C 1 +C 2 ), the non-coding ratio was adjusted to the range of [0,1] by using the minimum-maximum value normalization, and then 455 groups were selected as training sets and 195 groups as test sets in a ratio of 7:3. The sets were input into four models: gcForest, PSO-gcForest, GEO-gcForest and LGEO-gcForest, and the different diagnosis models were compared and analyzed. The diagnosis results of different models are shown in the following table.

[0093]

[0094] It can be seen from the above table that the overall diagnostic accuracy of the gcForest, PSO-gcForest, GEO-gcForest and LGEO-gcForest fault diagnosis models are 82.06%, 85.59% and 93.88% respectively, while the overall diagnostic accuracy of the LGEO-gcForest fault diagnosis model is as high as 96.08%. The results show that no matter which type of fault is diagnosed or the overall diagnosis, the LGEO-gcForest fault diagnosis model has the best diagnostic effect.

[0095] Although the embodiments of the present invention have been shown and described, as detailed above, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A fault diagnosis method for oil-immersed transformer based on LGEO-gcForest, characterized in that: Here are the steps: Step S1: Calculate the non-coded ratio of dissolved gas in oil according to the data sample of dissolved gas in oil of the transformer, and mark the fault type; Step S2: standardize the non-coding ratio and divide the standardized data into a training set and a test set; Step S3: Initialize two parameters of the gcForest model: the multi-granularity scanning window size and the maximum number of cascades allowed by the cascade forest, and then input the training set into the constructed gcForest model; Step S4: taking the multi-granularity scanning window size of the gcForest model and the maximum number of cascades allowed by the cascade forest as optimization parameters, and taking the fault diagnosis accuracy of the gcForest model as fitness; using the Golden Eagle optimization algorithm to optimize the gcForest model, and returning the optimal parameters, including the optimal multi-granularity scanning window size and the optimal maximum number of cascades allowed by the cascade forest; Step S5: construct an LGEO-gcForest fault diagnosis model according to the returned optimal parameters, input the test set, obtain the transformer fault diagnosis result, and verify the accuracy; convert the dissolved gas data in oil of unknown fault category into a non-coded ratio, input the LGEO-gcForest fault diagnosis model with qualified verification accuracy, and obtain the transformer fault diagnosis result; The process of optimizing the gcForest model using the Golden Eagle optimization algorithm is as follows: Step S41: Initialize the parameters of the Golden Eagle optimization algorithm, including the number of Golden Eagle populations and the maximum number of iterations, and use PWLCM mapping to initialize the coordinates of the Golden Eagle population; Step S42: randomly selecting an optimal hunting position and calculating an attack vector; Step S43: determine whether the attack vector is a zero vector; when the attack vector is a zero vector, return to step S42 and recalculate the attack vector; when the attack vector is not a zero vector, update the cruise vector and use the improved Levy flight mechanism to update the position of the golden eagle; Step S44: determine whether the maximum number of iterations is met. If not, return to step S42 and iterate again. When the maximum number of iterations is met, output the best fitness and the corresponding multi-granularity scanning window size and the maximum number of cascades allowed by the cascade forest. The way to initialize the coordinates of the golden eagle population using PWLCM mapping is as follows: in, is the position of the i-th golden eagle at the t-th iteration, is the position of the i-th golden eagle at the t+1th iteration, and p is the control parameter; The formula for updating the position of the golden eagle using the improved Levy flight mechanism is as follows: in, is the position of the i-th golden eagle at the t-th iteration, is the step vector of the i-th golden eagle in the t-th iteration, C is a random target point between (0,1), is the best position reached by the golden eagle individual in t iterations, λ is the proportional coefficient of each optimization, α is the step size scaling factor, Levy (β) is the Levy random path, and T is the maximum number of iterations.

2. The oil-immersed transformer fault diagnosis method based on LGEO-gcForest according to claim 1 is characterized in that: The attack vector is as follows: Among them, A i is the attack vector of the i-th golden eagle, is the best position reached by the golden eagle individual in the tth iteration.

3. The oil-immersed transformer fault diagnosis method based on LGEO-gcForest according to claim 1 is characterized in that: The cruise vector is as follows: Among them, C i is the cruising vector of the i-th golden eagle, c k is the kth element of the cruise vector, n is the number of elements, rand is a random number, A j is the attack vector of the jth golden eagle, a k is the kth element of the attack vector of the jth golden eagle, and d is the constant of the hyperplane equation where the golden eagle is located.

4. The oil-immersed transformer fault diagnosis method based on LGEO-gcForest according to claim 1 is characterized in that: The non-coded ratios include: CH4 / H2, C2H4 / C2H2, C2H4 / C2H6, C2H2 / (C1+C2), H2 / (H2+C1+C2), C2H4 / (C1+C2), CH4 / (C1+C2), C2H6 / (C1+C2), (CH4+C2H4) / (C1+C2) where C1 represents a first-order hydrocarbon.

5. The oil-immersed transformer fault diagnosis method based on LGEO-gcForest according to claim 1 is characterized in that: Normalizing the non-coded ratios is to adjust the non-coded ratios to the range of [0,1] using minimum-maximum normalization.