High-voltage circuit breaker fault diagnosis method and system based on improved deep forest model

Abstract

The application discloses a high-voltage circuit breaker fault diagnosis method and system based on an improved deep forest model. The method comprises the following steps: acquiring sample data of the movement process of each mechanical component of a first high-voltage circuit breaker under different state types; pre-analyzing and comparing the sample data of different state types, evaluating the feature importance of the sample data of different state types through a Gini index, determining a multi-granularity scanning window parameter according to an initial multi-granularity scanning window size and the relative importance of each feature, performing multi-granularity scanning on the sample data, adding a PCA algorithm to a cascade structure of a traditional deep forest model, performing combined dimension reduction processing on a high-dimensional vector, testing the performance of each forest through cross-validation, and establishing an improved deep forest model; and the improved deep forest model is used for diagnosing an unknown fault state of a high-voltage circuit breaker. The embodiment realizes high-voltage circuit breaker fault diagnosis under a very small sample, and improves the precision of fault diagnosis.

Description

Technical Field

[0001] This invention relates to the field of high-voltage circuit breaker fault diagnosis, and in particular to a high-voltage circuit breaker fault diagnosis method and system based on an improved deep forest model. Background Technology

[0002] As a key piece of equipment in high-voltage power grids, the safe and reliable operation of high-voltage circuit breakers is crucial to the normal operation of the grid. The vast majority of accidents caused by circuit breaker failures are mechanical problems. Therefore, preventative monitoring of switching equipment is essential. This not only enables timely fault detection and isolation but also effectively reduces unnecessary workload for staff, playing a vital role in the safety and stability of the power grid.

[0003] Currently, in diagnosing mechanical faults in the operating mechanism of high-voltage circuit breakers, traditional fault signal extraction methods involve installing accelerometers and linear displacement sensors on a moving component. However, in the field, the installation and removal of these sensors not only reduces the efficiency of circuit breaker fault diagnosis but also introduces additional mass into the moving components, potentially interfering with their normal operation. Some installation methods can even damage the mechanical structure of the circuit breaker. However, advancements in hardware have made machine vision-based signal extraction methods possible. These methods utilize high-speed cameras to capture the entire motion of mechanical components and extract their accurate trajectories.

[0004] In fault signal-based diagnostic methods for high-voltage circuit breakers, the rapid development of artificial intelligence technology has made research on the mechanical characteristics and condition assessment of high-voltage circuit breakers more efficient. Therefore, research on high-voltage circuit breaker condition identification technology based on machine learning and deep learning is of great significance for improving safe power consumption and enhancing the reliability of power systems. Currently, there are many methods for fault diagnosis of high-voltage circuit breakers, such as support vector machines (PCA-SVM) based on principal component analysis, random forests (RF), and deep neural networks (DNN). However, while these methods achieve high diagnostic accuracy, they all require a large amount of training data, especially with small training samples, where their performance is not ideal, increasing experimental workload and detection costs for supervised learning. On the other hand, high-voltage circuit breakers operate at a low frequency, making it difficult to obtain a large number of fault samples to meet the training and testing requirements of ordinary fault classification models.

[0005] Currently, in fault diagnosis models with small training samples, the traditional deep forest (DF)-based fault diagnosis method has demonstrated excellent performance in various mechanical fields. However, its application to fault diagnosis of high-voltage circuit breakers presents numerous adaptation problems, making it unsuitable for direct and effective fault diagnosis of high-voltage circuit breakers. Furthermore, due to the complex hierarchical structure of DF, it requires significantly more computation time compared to traditional machine learning methods, which is detrimental to online monitoring of high-voltage circuit breakers. Summary of the Invention

[0006] This invention provides a high-voltage circuit breaker fault diagnosis method and system based on an improved deep forest model, which enables high-voltage circuit breaker fault diagnosis with extremely small sample sizes, improves the accuracy of fault diagnosis, and shortens the diagnosis time of high-voltage circuit breaker faults.

[0007] To address the aforementioned technical problems, embodiments of the present invention provide a high-voltage circuit breaker fault diagnosis method based on an improved deep forest model, including:

[0008] Obtain sample data of the motion process of each mechanical component under different state types of operation of the first high-voltage circuit breaker;

[0009] Pre-analysis and comparison of sample data of different state types are performed to determine the initial multi-granularity scanning window size. The Gini index is used to evaluate the feature importance of sample data of different state types to obtain the relative importance of several features. Based on the initial multi-granularity scanning window size and the relative importance of each feature, the multi-granularity scanning window parameters are determined.

[0010] Based on the multi-granularity scanning window parameters, the sample data is scanned in multiple granularities to obtain high-dimensional vectors. These high-dimensional vectors are then used for training an improved deep forest. The performance of each level of the forest is tested through cross-validation. Training stops when the preset cross-validation conditions are met, thus establishing an improved deep forest model. The improved deep forest training involves adding the PCA algorithm to the cascaded structure of the traditional deep forest model to merge and reduce the dimensionality of the high-dimensional vectors.

[0011] The test data of the second high-voltage circuit breaker under unknown fault conditions is obtained, and the test data is input into the improved deep forest model to obtain the fault diagnosis results of the second high-voltage circuit breaker.

[0012] By implementing this embodiment of the invention, sample data of the motion process of each mechanical component during operation of the first high-voltage circuit breaker under different state types are obtained. The specific parameters of the multi-granularity scanning window are determined by pre-analysis and comparison of the sample data, combined with the relative feature importance feedback from the Gini index. This effectively avoids the generation of interfering data and ensures the accuracy of the multi-granularity scanning window parameters. Based on the multi-granularity scanning window parameters, the sample data is scanned in a multi-granularity manner to obtain high-dimensional vectors. The PCA algorithm is then added to the cascaded structure of the traditional deep forest model to merge and reduce the dimensionality of the high-dimensional vectors. The reduced vectors are used for training the next level of the forest. Cross-validation is used to test the performance of each level of the forest. Training stops when the preset cross-validation conditions are met, establishing an improved deep forest model. This improved method allows each part of the data in a single sample to play its role effectively, thus avoiding interfering data generated during multi-granularity scanning. Simultaneously, the PCA algorithm further reduces data redundancy in the cascaded structure. This results in excellent fault diagnosis accuracy on small training samples, improves fault diagnosis accuracy, reduces computational load, significantly reduces model training time, and thus shortens the fault diagnosis time of the high-voltage circuit breaker. The test data of the second high-voltage circuit breaker under unknown fault conditions is obtained. The test data is input into the improved deep forest model to obtain the fault diagnosis results of the second high-voltage circuit breaker. The improved deep forest model is used to quickly and accurately diagnose the fault of the tested high-voltage circuit breaker (second high-voltage circuit breaker).

[0013] As a preferred approach, sample data of different state types are pre-analyzed and compared to determine the initial multi-granularity scanning window size, specifically:

[0014] Plot sample data of different state types on the same coordinate axis, find the overlapping part of the sample data curves of different state types on the same coordinate axis, obtain the trajectory overlap part based on the overlapping part of the curves between different state types, and set the width of the initial multi-granularity scanning window size to be greater than the length of the trajectory overlap part.

[0015] In implementing this embodiment of the invention, traditional multi-granularity scanning in DF (Digital Functions) typically employs multiple scanning windows of different sizes to obtain more learning data from different angles when processing high-dimensional data samples. If the original sample dimension is d, the scanning window is generally set to three different sizes: d / 16, d / 8, and d / 4. This results in some slice data obtained from traditional multi-granularity scanning being indistinguishable by the algorithm, generating a large amount of interference data. First, a trajectory comparison analysis is performed on the high-voltage circuit breaker sample data to identify the overlapping parts of the corresponding trajectory curves of different samples. The width of the scanning window (the initial width of the multi-granularity scanning window) is then preliminarily determined. That is, the width of the scanning window must be greater than the length of the overlapping trajectory data. If the width of the scanning window is less than the length of the overlapping trajectory data, it will prevent a portion of the scanned slice data from being entirely overlapping data, which cannot be distinguished by the algorithm, thus reducing the accuracy of the model. Setting the width of the scanning window to be greater than the length of the overlapping trajectory improves accuracy and precision.

[0016] As a preferred approach, the Gini index is used to evaluate the feature importance of sample data for different state types, obtaining the relative importance of several features, specifically:

[0017] Sample data of different state types are combined to obtain combined sample data. The combined sample data is then input into a random forest. The Gini index in the random forest is used to calculate the importance of all features in the combined sample data, and the importance of each part of the adjacent features is output to obtain the relative importance of each part of the features.

[0018] This invention further improves upon the shortcomings of traditional multi-granularity scanning by using the Gini index to determine the relative importance of different features in the high-voltage circuit breaker data sample. Features with higher relative importance are more beneficial for the algorithm to classify them. This ensures that each scanned slice of data contains relatively important features, guaranteeing that each slice contains sufficient information for the algorithm's classification and effectively avoiding the generation of interfering data.

[0019] As a preferred approach, the multi-granularity scanning window parameters are determined based on the initial multi-granularity scanning window size and the relative importance of each feature, specifically as follows:

[0020] Find the maximum value among the relative importance of each feature, and determine the features corresponding to the maximum relative importance;

[0021] Based on the width of the preset initial multi-granularity scanning window size, the parameters of the multi-granularity scanning window are determined to ensure that the slice data scanned by each window contains the part of the feature corresponding to the highest relative importance.

[0022] By implementing the embodiments of the present invention, the width of the multi-granularity scanning window is greater than the length of the overlapping part of the trajectories between samples of different fault types, and the slice data scanned by each window must contain relatively important features. This ensures that each slice data contains enough information that is beneficial to the algorithm classification and avoids the generation of confused or interfering data.

[0023] As a preferred approach, the PCA algorithm is incorporated into the cascaded structure of the traditional deep forest model to merge and reduce the dimensionality of high-dimensional vectors. Specifically:

[0024] The high-dimensional vector is input into each level of the forest in the cascaded structure of the traditional deep forest model, and the forest class vectors are output.

[0025] The forest class vectors are merged with the high-dimensional vector to obtain the merged vectors of each level of cascaded forest. The dimensionality of the merged vectors of each level of cascaded forest is reduced by the PCA algorithm. The retained features of each level of cascaded forest are determined by the variance contribution. The retained features of the previous level of cascaded forest are used to train the next level of cascaded forest.

[0026] This invention incorporates PCA (Programmatical Convergence Aspect) algorithm for feature extraction into the cascaded structure of the traditional Deep Forest (DF) model. Dimensionality reduction is performed on the merged data at each level, and then a select few features are used to train the next level of the cascaded forest. This improved method allows each part of a single sample to effectively utilize its capabilities. Since the original travel data samples of high-voltage circuit breakers are high-dimensional, the dimensionality of the vectors obtained after multi-granularity scanning will be even higher. This means that a large amount of redundant data exists. To remove unnecessary redundancy, PCA feature extraction technology is introduced into the cascaded structure of the DF model to reduce the data dimensionality.

[0027] As a preferred approach, the performance of each level of the forest is tested through cross-validation. Training stops when the preset cross-validation conditions are met, specifically:

[0028] The sample data of different state types are divided into several subsets on an equal basis. Each subset of data is used as a validation set, and the remaining subsets are used as training sets. The training sets are input into the current level of the connected forest to obtain several current deep forest models. The average classification accuracy of the validation set of each current deep forest model is used as the performance index of each level of the connected forest.

[0029] If the performance metrics of cross-validation for three consecutive levels of cascaded forests do not improve, or if the performance metrics of cross-validation for the next level of cascaded forest are not greater than those for the previous level of cascaded forest, then training should be stopped.

[0030] As a preferred option, the different state types include: normal state, low voltage control fault, electromagnet defect, trip buffer fault, and jamming fault.

[0031] To address the same technical problem, embodiments of the present invention also provide a high-voltage circuit breaker fault diagnosis device, comprising: a sample acquisition module, a scanning parameter module, a training module, and a testing module;

[0032] Among them, the sample acquisition module is used to acquire sample data of the motion process of each mechanical component when the first high-voltage circuit breaker is running under different state types;

[0033] The scanning parameter module is used to pre-analyze and compare sample data of different state types, determine the initial multi-granularity scanning window size, evaluate the feature importance of sample data of different state types through the Gini index, obtain the relative importance of several features, and determine the multi-granularity scanning window parameters based on the initial multi-granularity scanning window size and the relative importance of each feature.

[0034] The training module is used to perform multi-granularity scanning on the sample data according to the multi-granularity scanning window parameters to obtain high-dimensional vectors, and then train the improved deep forest using the high-dimensional vectors. The performance of each level of the forest is tested by cross-validation. Training stops when the preset cross-validation conditions are met, and the improved deep forest model is established. The improved deep forest training adds the PCA algorithm to the cascaded structure of the traditional deep forest model to merge and reduce the dimensionality of the high-dimensional vectors.

[0035] The test module is used to acquire test data under unknown fault conditions of the second high-voltage circuit breaker, input the test data into the improved deep forest model, and obtain the fault diagnosis results of the high-voltage circuit breaker under test.

[0036] To address the same technical problem, embodiments of the present invention also provide a high-voltage circuit breaker fault diagnosis system based on an improved deep forest model, a high-voltage circuit breaker fault diagnosis device, an imaging device, and an image acquisition card; wherein, the high-voltage circuit breaker fault diagnosis device executes a high-voltage circuit breaker fault diagnosis method based on an improved deep forest model;

[0037] The equipment connections are as follows: the high-voltage circuit breaker fault diagnosis equipment is connected to the image acquisition card, and the imaging equipment is connected to the image acquisition card.

[0038] As a preferred option, the camera is used to capture the entire motion process of the corresponding mechanical components of the high-voltage circuit breaker and output high frame rate video data.

[0039] The image acquisition card is used to transmit high frame rate video data to high voltage circuit breaker fault diagnosis equipment. Attached Figure Description

[0040] Figure 1 : A flowchart illustrating an embodiment of the high-voltage circuit breaker fault diagnosis method based on an improved deep forest model provided by the present invention;

[0041] Figure 2 This is a comparison chart of different fault sample curves of a high-voltage circuit breaker, representing an embodiment of the high-voltage circuit breaker fault diagnosis method based on an improved deep forest model provided by the present invention.

[0042] Figure 3 : A comparison diagram of the relative importance of different features in an embodiment of the high-voltage circuit breaker fault diagnosis method based on an improved deep forest model provided by the present invention;

[0043] Figure 4 : A modified multi-granularity scanning flowchart of an embodiment of the high-voltage circuit breaker fault diagnosis method based on an improved deep forest model provided by the present invention;

[0044] Figure 5 : An improved cascaded structure framework diagram of an embodiment of the high-voltage circuit breaker fault diagnosis method based on an improved deep forest model provided by the present invention;

[0045] Figure 6 : A PCA calculation feature variance contribution and cumulative value diagram of an embodiment of the high-voltage circuit breaker fault diagnosis method based on an improved deep forest model provided by the present invention;

[0046] Figure 7 : An accuracy graph of different diagnostic models for an embodiment of the high-voltage circuit breaker fault diagnosis method based on an improved deep forest model provided by the present invention;

[0047] Figure 8 : A time comparison chart of four algorithm models for high-voltage circuit breaker fault diagnosis based on an improved deep forest model provided by the present invention on high-voltage circuit breaker fault diagnosis.

[0048] Figure 9 : A schematic diagram of the structure of an embodiment of the high-voltage circuit breaker fault diagnosis device provided by the present invention;

[0049] Figure 10 : A connection diagram of an embodiment of the high-voltage circuit breaker fault diagnosis system based on an improved deep forest model provided by the present invention. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] Example 1

[0052] Please refer to Figure 1 This is a flowchart illustrating the fault diagnosis method for high-voltage circuit breakers based on an improved deep forest model provided in this embodiment of the invention. The fault diagnosis method of this embodiment is applicable to different faults in high-voltage circuit breakers. This embodiment improves the accuracy of fault diagnosis by improving the deep forest model to diagnose high-voltage circuit breaker faults with extremely small sample sizes. The fault diagnosis method includes steps 101 to 104, each step as follows:

[0053] Step 101: Obtain sample data of the motion process of each mechanical component under different state types of the first high-voltage circuit breaker during operation.

[0054] Optional, different state types include: normal state, low voltage control fault, electromagnet defect, trip buffer fault, and jamming fault.

[0055] In this embodiment, a high-voltage circuit breaker fault acquisition platform can be built. By adjusting the circuit breaker, different fault states of the high-voltage circuit breaker can be simulated. The platform for acquiring the mechanical characteristic signals of the high-voltage circuit breaker is constructed, including but not limited to a high-voltage circuit breaker, a high-speed camera, a supplementary light, an image acquisition card, and a high-voltage circuit breaker fault diagnosis device. The high-speed camera can capture the entire movement process of the corresponding mechanical components of the high-voltage circuit breaker. The high-speed camera can be used to capture the moving mechanism of the high-voltage circuit breaker and output high frame rate video data, i.e., sample data. A supplementary light can be used to illuminate the monitored position. The high frame rate video signal is transmitted to the high-voltage circuit breaker fault diagnosis device in conjunction with the image acquisition card. Sample data of the high-voltage circuit breaker under different state types (normal state and different fault states) can be repeatedly acquired. As an example in this embodiment, mechanical characteristic data of the high-voltage circuit breaker under five different state types, namely normal state, low voltage control fault, electromagnet defect, trip buffer fault, and jamming fault, will be acquired. A total of 200 sets of data samples will be acquired under the five states (40 sets of samples for each state).

[0056] Step 102: Perform pre-analysis and comparison of sample data of different state types to determine the initial multi-granularity scanning window size. Use the Gini index to evaluate the feature importance of sample data of different state types to obtain the relative importance of several features. Based on the initial multi-granularity scanning window size and the relative importance of each feature, determine the multi-granularity scanning window parameters.

[0057] In this embodiment, sample data of different fault types are pre-analyzed in the high-voltage circuit breaker fault diagnosis equipment and visualized and compared to find the overlapping parts of the trajectory data of different fault types, so as to preliminarily determine the size of the multi-granularity scanning window.

[0058] Optionally, step 102 specifically includes steps 1021 to 1023, each of which is as follows:

[0059] Step 1021: Plot the sample data of different state types on the same coordinate axis, find the overlapping part of the sample data curves of different state types on the same coordinate axis, obtain the trajectory overlap part based on the overlapping part of the curves between different state types, and set the width of the initial multi-granularity scanning window to be greater than the length of the trajectory overlap part.

[0060] In this embodiment, the high-voltage circuit breaker fault diagnosis equipment first needs to perform pre-analysis of the data to identify the overlapping portions between different fault sample data. By plotting sample data of different fault types under different state types on the same coordinate axis, it will be found that certain sample curves overlap on the same coordinate axis. The overlapping curve portion is the data overlap part (trajectory overlap part), used to initially determine the window size in multi-granularity scanning. A comparison chart of different fault sample curves of high-voltage circuit breakers is shown below. Figure 2 As shown, a comparative analysis of data curves was performed in the fault diagnosis equipment for high-voltage circuit breakers. A set of samples was taken for each state and plotted on the same graph. The overlapping parts of the trajectories were found. It can be seen from the graph that the overlapping curve data are between 0 and 300 frames. Each frame corresponds to a data feature. Therefore, the size of the multi-granularity scanning window was initially set to be greater than 300 dimensions.

[0061] Step 1022: Combine sample data of different state types to obtain combined sample data. Input the combined sample data into a random forest. Calculate the importance of all features in the combined sample data using the Gini index in the random forest. Output the importance of each part of adjacent features to obtain the relative importance of each feature.

[0062] In this embodiment, the Gini index is used to evaluate the feature importance of the training samples and returns the relative importance of each feature in the sample. The sample data is combined and then input into a random forest. The Gini index in the random forest is used to calculate the importance of all features in the sample, and the importance of each set of adjacent features is output, i.e., the relative importance of each feature. The comparison of the relative importance of different feature sets calculated by the Gini index is shown below. Figure 3 As shown, the Gini index is used to evaluate the feature importance of all features in the sample data, and the relative importance of each adjacent 100 features is fed back. After comparison, it is found that the relative importance of the 300-400 dimension features is the greatest, that is, the feature data in this range is most conducive to the algorithm's classification. Therefore, each scanned slice of data should include this part of the data as much as possible.

[0063] Step 1023: Find the maximum value among the relative importance of each feature, and determine the partial feature corresponding to the maximum relative importance; based on the width of the preset initial multi-granularity scanning window size, ensure that the slice data scanned by each window contains the partial feature corresponding to the maximum relative importance, and determine the multi-granularity scanning window parameters.

[0064] In this embodiment, the width of the initial multi-granularity scanning window is greater than the length of the overlapping portion of the trajectories between samples of different fault types, and the slice data scanned by each window must contain relatively important features. Determining the multi-granularity scanning window parameters ensures that each slice data contains sufficient information beneficial to the algorithm's classification, avoiding the generation of confused or interfering data. The adjusted multi-granularity scanning process is as follows: Figure 4 As shown, the scanning window size is set to 400 dimensions, and the scanning step count is 1. Since the original sample length is 500 dimensions, 101 slice data points of length 400 dimensions will be obtained after scanning. These slice data points are then input into a random forest and a fully random forest, respectively. In this embodiment, five operating states of high-voltage circuit breakers are provided, so each slice data point will be transformed into a 5-dimensional class vector by the random forest and the fully random forest. Finally, the 202 5-dimensional class vectors output by the two forests are merged into a 1010-dimensional vector.

[0065] Step 103: Based on the multi-granularity scanning window parameters, perform multi-granularity scanning on the sample data to obtain high-dimensional vectors, and train the high-dimensional vectors into an improved deep forest. Validate the performance of each level of the forest through cross-validation. Stop training when the preset cross-validation conditions are met, and establish the improved deep forest model. The improved deep forest training involves adding the PCA algorithm to the cascaded structure of the traditional deep forest model to merge and reduce the dimensionality of the high-dimensional vectors.

[0066] In this embodiment, after determining the multi-granularity scanning window parameters, multi-granularity scanning is performed on the training sample data to obtain a high-dimensional vector after feature transformation. This high-dimensional vector is then input into each stage of the deep forest (DF) cascade structure. The improved cascade structure framework is as follows: Figure 5 As shown, the 1010-dimensional vector obtained in step 1023 is input into each level of the cascaded structure. Each level consists of four forests, corresponding to two random forests and two completely random forests. Each forest also outputs a 5-dimensional class vector. All class vectors at each level are merged with the 1010-dimensional vector, and then PCA is used to reduce the dimensionality, further eliminating redundant data.

[0067] Optionally, the PCA algorithm can be added to the cascaded structure of the traditional deep forest model to merge and reduce the dimensionality of high-dimensional vectors. Specifically, the high-dimensional vector is input into each level of the cascaded structure of the traditional deep forest model, and each forest class vector is output. The forest class vectors are merged with the high-dimensional vector to obtain the merged vector of each level of the cascaded forest. The dimensionality of the merged vector of each level of the cascaded forest is reduced by the PCA algorithm, and the retained features of each level of the cascaded forest are determined by the variance contribution. The retained features of the previous level of the cascaded forest are used to train the next level of the cascaded forest.

[0068] In this embodiment, the class vectors output from each level of the forest are merged with the high-dimensional vectors obtained from multi-granularity scanning, and then PCA is used for dimensionality reduction. The variance contribution is used to determine the number of features retained. This merging and dimensionality reduction step is repeated for each level. PCA calculates the feature variance contribution and cumulative variance, such as... Figure 6 As shown, the variance contribution is used to measure the importance of a feature for fault classification. When the cumulative variance reaches 99%, it is assumed that this feature already contains all the information of all the original data. Figure 6 The cumulative variance of the first 6 features has reached 99.29%, so at each level, the first 6 features with the largest variance contribution are extracted to train the next level of forest.

[0069] Optionally, the performance of each level of the cascaded forest can be tested through cross-validation. Training is stopped when the preset cross-validation conditions are met. Specifically, the sample data of different state types are divided into several subsets on an equal basis. Each subset is used as a validation set, and the remaining subsets are used as training sets. The training sets are input into the current level of the cascaded forest to obtain several current deep forest models. The average classification accuracy of the validation set of each current deep forest model is used as the performance index of each level of the cascaded forest. If the cross-validation performance index of three consecutive levels of cascaded forests does not improve, or the cross-validation performance index of the next level of cascaded forest is not greater than that of the previous level of cascaded forest, then training is stopped.

[0070] In this embodiment, cross-validation is used to test the performance of each level of the forest. Training of the improved deep forest model terminates when the performance of three consecutive levels of the cascaded structure fails to improve. K-fold cross-validation is used to verify the performance of each level of the forest. First, the training data is divided into K subsets. Each subset is used as a validation set, and the remaining K-1 subsets are used as the training set. This results in K models. The average classification accuracy of these K models on the final validation set is used as the performance metric for that level of the forest. Training terminates if the cross-validation score of three consecutive levels of the forest fails to improve, or if the cross-validation score of the next level is less than or equal to that of the previous level. The model is trained using different numbers of training samples, and then the performance is evaluated using test samples. Each state provides 20 samples, and the other 20 samples are used for model testing; that is, the total training set consists of 100 samples, and the test set consists of another 100 samples.

[0071] In this embodiment, the accuracy of commonly used fault diagnosis methods for high-voltage circuit breakers is compared on training sets of different sizes. The performance of four diagnostic models, including the improved PCA-DF (Progressive Processing Tree-Based Algorithm), on different numbers of training samples can be observed. The accuracy rates of different diagnostic models, such as... Figure 7 As shown, PCA-DF is an improved deep forest model, DF is the traditional deep forest model, RF is the random forest model, and PCA_SVM is a support vector machine model based on principal component analysis. The improved deep forest model PCA-DF outperforms the traditional DF model, and also outperforms the other two commonly used models on a small training set. A comparison of the training and testing times of these four models, and a comparison of the running time of the four algorithm models in high-voltage circuit breaker fault diagnosis, are presented. Figure 8 As shown, the traditional DF algorithm takes a total of 42.11 seconds for training and testing. The improved DF model (improved deep forest model, PCA-cascade forest) takes only 7.95 seconds, significantly reducing the computation time. Although the RF model has the shortest time, its performance with a small number of training samples is not ideal. Considering all factors, the improved DF model (improved deep forest model) has the best overall performance.

[0072] Step 104: Obtain test data under unknown fault conditions of the second high-voltage circuit breaker, input the test data into the improved deep forest model, and obtain the fault diagnosis results of the second high-voltage circuit breaker.

[0073] In this embodiment, the trained improved deep forest model is retained in the high-voltage circuit breaker fault diagnosis equipment for subsequent fault analysis of unknown sample data. Sample data of an unknown fault state of a certain high-voltage circuit breaker (the second high-voltage circuit breaker) is obtained, and the corresponding sample data is input into the model in the high-voltage circuit breaker fault diagnosis equipment. First, the sample data undergoes feature transformation through multi-granularity scanning, and then the transformed data is passed into the cascaded structure of the model. Multiple different forests in the cascaded structure perform diagnostic evaluations on the sample data and output diagnostic results respectively. Finally, the average of all results is taken, and the result with the highest average value is taken as the final diagnostic result. This result represents the current fault state of the high-voltage circuit breaker, and the diagnosis is complete.

[0074] In implementing this invention, traditional multi-granularity scanning in high-dimensional circuit breaker (DF) typically employs multiple scanning windows of varying sizes to obtain more learning data from different perspectives when processing high-dimensional data samples. If the original sample dimension is 'd', the scanning window is generally set to one of three different sizes: d / 16, d / 8, or d / 4. However, after comparative analysis of data from high-voltage circuit breaker samples with different fault types, it was found that some trajectory data corresponding to different faults overlapped. This means that some slice data obtained through traditional multi-granularity scanning cannot be distinguished by the algorithm, resulting in a large amount of interference data. To solve this interference problem, a trajectory comparison analysis is first performed on the high-voltage circuit breaker sample data to identify the overlapping parts of the corresponding trajectory curves of different samples. The width of the scanning window is then preliminarily determined; that is, the width of the scanning window must be greater than the length of the overlapping trajectory data. If the width of the scanning window is less than the length of the overlapping trajectory data, a portion of the scanned slice data will be entirely overlapping data, which cannot be distinguished by the algorithm, thus reducing the model's accuracy.

[0075] To further improve upon the shortcomings of traditional multi-granularity scanning, the Gini index is used to determine the relative importance of different features in the high-voltage circuit breaker data sample. Features with higher relative importance are more beneficial for the algorithm to classify them. Ensuring that each window's scanned slice data contains relatively important features guarantees that each slice contains sufficient information for the algorithm's classification, effectively avoiding the generation of interfering data.

[0076] Since the original travel data samples of high-voltage circuit breakers are high-dimensional, the dimensionality of the vectors obtained after multi-granularity scanning will be even higher, meaning there is still a large amount of redundant data. To remove unnecessary redundant data, PCA feature extraction technology will be introduced into the cascaded structure of the DF (Farmer-Derived) model to reduce the dimensionality of the data. The class vectors output by each level of the forest are merged with the high-dimensional vectors obtained from multi-granularity scanning, and then PCA is used to reduce the dimensionality. The variance contribution rate is used to determine the number of features retained, and then these features are input into the next level of the forest, and this step is repeated for each level. When the performance of the cascaded structure does not improve for three consecutive levels, training terminates. Then, the model is tested using a test set, and the performance of the model is evaluated using ROC curves, accuracy, recall, and F1 score.

[0077] Example 2

[0078] Accordingly, see Figure 9 , Figure 9 This is a schematic diagram of a second embodiment of the high-voltage circuit breaker fault diagnosis device provided by the present invention. Figure 9 As shown, the high-voltage circuit breaker fault diagnosis equipment includes a sample acquisition module 901, a scanning parameter module 902, a training module 903, and a testing module 904.

[0079] Among them, the sample acquisition module 901 is used to acquire sample data of the motion process of each mechanical component when the first high voltage circuit breaker is running under different state types;

[0080] The scanning parameter module 902 is used to pre-analyze and compare sample data of different state types, determine the initial multi-granularity scanning window size, evaluate the feature importance of sample data of different state types through the Gini index, obtain the relative importance of several features, and determine the multi-granularity scanning window parameters based on the initial multi-granularity scanning window size and the relative importance of each feature.

[0081] The training module 903 is used to perform multi-granularity scanning on the sample data according to the multi-granularity scanning window parameters to obtain high-dimensional vectors, and then train the high-dimensional vectors into an improved deep forest. The performance of each level of the forest is tested by cross-validation. Training stops when the preset cross-validation conditions are met, and an improved deep forest model is established. The improved deep forest training adds the PCA algorithm to the cascaded structure of the traditional deep forest model to merge and reduce the dimensionality of the high-dimensional vectors.

[0082] Test module 904 is used to acquire test data of the second high-voltage circuit breaker under unknown fault conditions, input the test data into the improved deep forest model, and obtain the fault diagnosis results of the high-voltage circuit breaker under test.

[0083] The aforementioned high-voltage circuit breaker fault diagnosis equipment can implement the high-voltage circuit breaker fault diagnosis method based on the improved deep forest model described in the above method embodiments. The options in the above method embodiments are also applicable to this embodiment, and will not be detailed here. The remaining content of this application's embodiments can be referred to the content of the above method embodiments, and will not be repeated in this embodiment.

[0084] By implementing this embodiment of the invention, sample data of the motion process of each mechanical component during operation of the first high-voltage circuit breaker under different state types are obtained. The specific parameters of the multi-granularity scanning window are determined by pre-analysis and comparison of the sample data, combined with the relative feature importance feedback from the Gini index. This effectively avoids the generation of interfering data and ensures the accuracy of the multi-granularity scanning window parameters. Based on the multi-granularity scanning window parameters, the sample data is scanned in a multi-granularity manner to obtain high-dimensional vectors. The PCA algorithm is then added to the cascaded structure of the traditional deep forest model to merge and reduce the dimensionality of the high-dimensional vectors. The reduced vectors are used for training the next level of the forest. Cross-validation is used to test the performance of each level of the forest. Training stops when the preset cross-validation conditions are met, establishing an improved deep forest model. This improved method allows each part of the data in a single sample to play its role effectively, thus avoiding interfering data generated during multi-granularity scanning. Simultaneously, the PCA algorithm further reduces data redundancy in the cascaded structure. This results in excellent fault diagnosis accuracy on small training samples, improves fault diagnosis accuracy, reduces computational load, significantly reduces model training time, and thus shortens the fault diagnosis time of the high-voltage circuit breaker. The test data of the second high-voltage circuit breaker under unknown fault conditions is obtained. The test data is input into the improved deep forest model to obtain the fault diagnosis results of the second high-voltage circuit breaker. The improved deep forest model is used to quickly and accurately diagnose the fault of the tested high-voltage circuit breaker (second high-voltage circuit breaker).

[0085] Example 3

[0086] Accordingly, see Figure 10 , Figure 10 This is a connection diagram of Embodiment 2 of the high-voltage circuit breaker fault diagnosis system based on the improved deep forest model provided by the present invention. Figure 10 As shown, the high-voltage circuit breaker fault diagnosis system based on the improved deep forest model includes: a high-voltage circuit breaker fault diagnosis device, an imaging device, and an image acquisition card; wherein, the high-voltage circuit breaker fault diagnosis device executes the high-voltage circuit breaker fault diagnosis method based on the improved deep forest model.

[0087] The equipment connections are as follows: the high-voltage circuit breaker fault diagnosis equipment is connected to the image acquisition card, and the imaging equipment is connected to the image acquisition card.

[0088] The camera equipment is used to capture the entire movement process of the corresponding mechanical components of the high-voltage circuit breaker and output high frame rate video data.

[0089] The image acquisition card is used to transmit high frame rate video data to high voltage circuit breaker fault diagnosis equipment.

[0090] In this embodiment, the high-voltage circuit breaker fault diagnosis equipment includes, but is not limited to, a host computer, and the imaging equipment includes, but is not limited to, a high-speed camera. The high-voltage circuit breaker fault diagnosis system may also include supplementary lighting to illuminate the monitored location, thereby obtaining clearer data and improving the accuracy of fault diagnosis.

[0091] The high-voltage circuit breaker fault diagnosis system based on the improved deep forest model described above can implement the high-voltage circuit breaker fault diagnosis method based on the improved deep forest model in the above method embodiments. The options in the above method embodiments are also applicable to this embodiment, and will not be detailed here. The remaining content of this application's embodiments can be referred to the content of the above method embodiments, and will not be repeated in this embodiment.

[0092] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A fault diagnosis method for high-voltage circuit breakers based on an improved deep forest model, characterized in that, include: Sample data of the motion process of each mechanical component of the first high-voltage circuit breaker under different state types during operation are obtained; wherein, the sample data is obtained by shooting the entire motion process of the corresponding mechanical component of the high-voltage circuit breaker with a shooting device, outputting high frame rate video data to an image acquisition card, and then outputting the data from the image acquisition card. The sample data of the different state types are pre-analyzed and compared to determine the initial multi-granularity scanning window size. The Gini index is used to evaluate the feature importance of the sample data of the different state types to obtain the relative importance of several features. Based on the initial multi-granularity scanning window size and the relative importance of each feature, the multi-granularity scanning window parameters are determined. According to the multi-granularity scanning window parameters, the sample data is scanned in a multi-granularity manner to obtain a high-dimensional vector. The high-dimensional vector is then used for training an improved deep forest. The performance of each level of the forest is tested by cross-validation. Training stops when the preset cross-validation conditions are met, and an improved deep forest model is established. The improved deep forest training involves adding the PCA algorithm to the cascaded structure of the traditional deep forest model to perform merging and dimensionality reduction processing on the high-dimensional vector. The test data of the second high-voltage circuit breaker under unknown fault conditions is obtained, and the test data is input into the improved deep forest model to obtain the fault diagnosis results of the second high-voltage circuit breaker.

2. The high-voltage circuit breaker fault diagnosis method based on an improved deep forest model as described in claim 1, characterized in that, The step of pre-analyzing and comparing sample data of different state types to determine the initial multi-granularity scanning window size specifically involves: Plot the sample data of the different state types on the same coordinate axis, find the overlapping part of the curves of the sample data of the different state types on the same coordinate axis, obtain the trajectory overlap part based on the overlapping part of the curves between the different state types, and set the width of the initial multi-granularity scanning window size to be greater than the length of the trajectory overlap part.

3. The high-voltage circuit breaker fault diagnosis method based on an improved deep forest model as described in claim 1, characterized in that, The method of evaluating the feature importance of sample data of different state types using the Gini index to obtain the relative importance of several features is as follows: The sample data of different state types are combined to obtain combined sample data. The combined sample data is then input into a random forest. The Gini index in the random forest is used to calculate the importance of all features in the combined sample data, and the importance of each part of adjacent features is output to obtain the relative importance of each feature.

4. The high-voltage circuit breaker fault diagnosis method based on the improved deep forest model as described in claim 2, characterized in that, The determination of multi-granularity scanning window parameters based on the initial multi-granularity scanning window size and the relative importance of each feature is specifically as follows: Find the maximum value among the relative importance of each feature, and determine the partial features corresponding to the maximum relative importance; Based on the width of the preset initial multi-granularity scanning window size, the parameters of the multi-granularity scanning window are determined to ensure that the slice data scanned by each window contains the partial features corresponding to the maximum relative importance.

5. The high-voltage circuit breaker fault diagnosis method based on the improved deep forest model as described in claim 1, characterized in that, The process of adding the PCA algorithm to the cascaded structure of the traditional deep forest model to perform merging and dimensionality reduction processing on the high-dimensional vectors is as follows: The high-dimensional vector is input into each level of the cascaded structure of the traditional deep forest model, and the forest class vectors are output. The forest class vectors are merged with the high-dimensional vector to obtain the merged vectors of each level of cascaded forest. The dimensionality of the merged vectors of each level of cascaded forest is reduced by the PCA algorithm. The retained features of each level of cascaded forest are determined by the variance contribution. The retained features of the previous level of cascaded forest are used to train the next level of cascaded forest.

6. The high-voltage circuit breaker fault diagnosis method based on the improved deep forest model as described in claim 5, characterized in that, The performance of each level of the forest is tested through cross-validation, and training stops when the preset cross-validation conditions are met. Specifically: The sample data of different state types are divided into several subsets on an equal basis. Each subset of data is used as a validation set, and the remaining subsets are used as training sets. The training sets are input into the current level of the connected forest to obtain several current deep forest models. The average classification accuracy of the validation set of each current deep forest model is used as the performance index of each level of the connected forest. If the performance metrics of cross-validation of three consecutive cascaded forests do not improve, or if the performance metrics of cross-validation of the next cascaded forest are not greater than those of the previous cascaded forest, then training should be stopped.

7. The high-voltage circuit breaker fault diagnosis method based on the improved deep forest model as described in claim 1, characterized in that, The different state types include: normal state, low voltage control fault, electromagnet defect, trip buffer fault, and jamming fault.

8. A fault diagnosis device for high-voltage circuit breakers, characterized in that, include: The module includes a sample acquisition module, a scanning parameter module, a training module, and a testing module. The sample acquisition module is used to acquire sample data of the motion process of each mechanical component of the first high-voltage circuit breaker under different state types during operation; wherein, the sample data is obtained by shooting the entire motion process of the corresponding mechanical component of the high-voltage circuit breaker through a shooting device, outputting high frame rate video data to an image acquisition card, and then outputting the data from the image acquisition card. The scanning parameter module is used to pre-analyze and compare sample data of different state types, determine the initial multi-granularity scanning window size, evaluate the feature importance of sample data of different state types using the Gini index, obtain the relative importance of several features, and determine the multi-granularity scanning window parameters based on the initial multi-granularity scanning window size and the relative importance of each feature. The training module is used to perform multi-granularity scanning on the sample data according to the multi-granularity scanning window parameters to obtain high-dimensional vectors, and to train the high-dimensional vectors into an improved deep forest. The performance of each level of the forest is tested by cross-validation. Training stops when the preset cross-validation conditions are met, and an improved deep forest model is established. The improved deep forest training adds the PCA algorithm to the cascaded structure of the traditional deep forest model to merge and reduce the dimensionality of the high-dimensional vectors. The test module is used to acquire test data under unknown fault conditions of the second high-voltage circuit breaker, and input the test data into the improved deep forest model to obtain the fault diagnosis results of the high-voltage circuit breaker under test.

9. A fault diagnosis system for high-voltage circuit breakers based on an improved deep forest model, characterized in that, include: A high-voltage circuit breaker fault diagnosis device, an imaging device, and an image acquisition card; wherein the high-voltage circuit breaker fault diagnosis device performs the high-voltage circuit breaker fault diagnosis method based on an improved deep forest model as described in any one of claims 1 to 7; The equipment is connected as follows: the high-voltage circuit breaker fault diagnosis equipment is connected to the image acquisition card, and the imaging equipment is connected to the image acquisition card.

10. The high-voltage circuit breaker fault diagnosis system based on the improved deep forest model as described in claim 9, characterized in that, The camera is used to capture the entire movement process of the corresponding mechanical components of the high-voltage circuit breaker and output high frame rate video data. The image acquisition card is used to transmit the high frame rate video data to the high voltage circuit breaker fault diagnosis device.

Images