A method and system for substation equipment fault identification based on deep forest algorithm
By combining the deep forest algorithm with DS evidence theory and cascaded forest, the problems of accuracy and speed in substation equipment fault diagnosis are solved, and efficient fault identification is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- STATE GRID JIBEI ELECTRIC POWER COMPANY LIMITED CHENGDE POWER SUPPLY
- Filing Date
- 2022-12-29
- Publication Date
- 2026-07-17
AI Technical Summary
Existing methods for diagnosing substation equipment faults are insufficient in terms of accuracy and speed. In particular, the parameter tuning of deep neural network models is complex and their generalization ability is insufficient, making it difficult to meet the requirements for accurate identification of substation equipment faults.
A substation equipment fault identification method based on the deep forest algorithm is adopted. Combining DS evidence theory and cascaded forest, the accuracy and speed of fault identification are improved through time-frequency domain feature extraction, multi-granularity scanning, cascaded forest training and DS evidence fusion.
It effectively overcomes the problem of feature redundancy, improves the accuracy and speed of power equipment fault identification, and meets the diagnostic requirements of industrial sites.
Smart Images
Figure CN116304918B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of substation equipment fault identification technology, specifically relating to a substation equipment fault identification method and system based on deep forest algorithm. Technical Background
[0002] With the continuous increase in electricity consumption, higher demands are being placed on the stability and intelligence of the power grid. Substations play a crucial role in transforming voltage levels, collecting current, distributing electrical energy, and regulating voltage. Therefore, they are an extremely important link in the transmission line and a frequent point of failure, requiring close monitoring. A failure in substation equipment can lead to large-scale power outages, causing incalculable losses to the national economy and industrial production. Therefore, ensuring the security of power supply has become a vital component of safeguarding national strategic energy security. Monitoring the operational status of equipment within substations and promptly detecting faults is of paramount importance for the safe operation of substations.
[0003] Substation equipment includes busbar equipment, contact equipment, coil equipment, and other related equipment. Busbar equipment includes busbars of various voltage levels and connecting cables, responsible for providing functions such as three-phase load, neutral line, and grounding line. Contact equipment includes circuit breakers, contactors, disconnect switches, and pressure plates, with a focus on the contact and arc-extinguishing elements in the corresponding switching equipment. Coil equipment is the most widely distributed and diverse equipment in a substation, including transformers, instrument transformers, and grounding arc extinguishers.
[0004] The operation of equipment within substations is complex. A single fault may manifest in multiple ways in the equipment's operational information, and the same fault may indicate multiple fault types. For faults caused by natural disasters, it is necessary to optimize the allocation of power system resources, improve the disaster resistance of electrical equipment, and propose advanced condition monitoring technologies and fault diagnosis methods. For faults caused by equipment issues, it is necessary to strengthen manufacturers' control over product quality and strictly control product defect rates. For faults caused by human factors, it is necessary to improve the technical skills of maintenance personnel and take relevant measures to prevent external damage. For complex equipment with intricate fault relationships, simple judgments based on inspection personnel are insufficient for accurate fault identification. Therefore, machine learning algorithms are widely used in fault identification of substation equipment.
[0005] In recent years, the collection of massive amounts of fault data has provided new opportunities for data-driven fault identification methods, attracting increasing attention from researchers and engineers. Commonly used fault diagnosis methods include Support Vector Machines (SVM), k-nearest neighbors (KNN), and artificial neural networks (ANN). The identification effect mainly depends on whether the fault features can accurately reflect the essential differences in fault types. Fault diagnosis models based on deep neural networks have good diagnostic effects, but their disadvantages include the need for significant manual time to adjust the training parameters, and the need to improve the model's generalization ability. To address the problem of overly complex parameter tuning for deep neural networks, the Deep Forest model has been proposed. This model is based on random forests and uses a deep structure similar to deep learning, making it an effective model with controllable parameters that can adapt to data of different scales. Deep Forest has achieved comparable results to deep neural networks in areas such as cancer diagnosis and recommender systems, but it has not yet been applied to the field of substation equipment fault identification technology. Summary of the Invention
[0006] This invention provides a substation equipment fault identification method and system based on the deep forest algorithm. It combines DS evidence theory with cascaded forests to overcome the feature redundancy problem encountered in the process of expanding cascaded forests, effectively improving the accuracy and speed of power equipment fault identification and meeting the requirements of industrial field for power equipment fault diagnosis.
[0007] This invention provides a substation equipment fault identification method based on the deep forest algorithm, which includes the following steps:
[0008] S1. Perform time-frequency domain index calculation and normalization on the collected monitoring dataset of fault signals from existing substation equipment to obtain a feature set;
[0009] S2. Perform multi-granularity scanning on the feature set, obtain all feature sub-vectors through a sliding window, and output class vectors by passing these feature sub-vectors through ordinary random forest and fully random forest respectively. Concatenate all the output class vectors to obtain a high-dimensional feature vector.
[0010] S3. Construct a multi-layer cascaded forest using multiple fully random forests and ordinary random forests, and input the obtained high-dimensional feature vectors into each layer of the cascaded forest for training.
[0011] S4. Perform DS evidence fusion on the multiple class probability distributions of the training output of the previous layer of the cascaded forest to obtain a fused feature vector, and concatenate it with a randomly selected class vector and the output of multi-granularity scanning as the input of the next layer of the cascaded forest. Finally, take the average of the class distribution probabilities of each random forest, and finally take the class with the highest prediction probability as the substation equipment fault prediction probability identification result.
[0012] S5. Perform k-fold cross-validation on the output value of each layer of the cascaded forest, and calculate the recognition accuracy of each cross-validation. When the accuracy improvement is greater than the preset value, add another layer of cascaded forest and retrain the entire new cascaded forest model. Otherwise, stop adding new layers and complete the training process of the cascaded forest.
[0013] S6. The real-time monitored substation equipment signals are processed and then input into the trained cascaded forest model to obtain online fault identification results.
[0014] Furthermore, the step of obtaining the feature set includes:
[0015] Extract multiple time-domain and frequency-domain features from the existing monitoring dataset of substation equipment fault signals;
[0016] An improved energy function is used to calculate the energy and energy entropy of the sampled dataset.
[0017] Furthermore, the formula for calculating the energy as a function is as follows:
[0018]
[0019] The formula for calculating energy entropy is as follows:
[0020]
[0021] In the formula, N is the length of the signal; p is the amplitude factor; x i It is the amplitude of the signal vibration point; when x i If the value is positive, add p; otherwise, subtract p.
[0022] Furthermore, the step of obtaining the high-dimensional feature vector includes:
[0023] For a set of samples containing K-dimensional time-domain or frequency-domain features, sliding sampling is performed through a sliding window of length L, with a step size of S for each sliding. After the sliding sampling is completed, N L-dimensional feature sub-vectors are obtained, where N = (KL) / S + 1.
[0024] All the obtained feature vectors are used as inputs to ordinary random forest and fully random forest. Each random forest yields N probability vectors of dimension P.
[0025] The N probability vectors of dimension P are concatenated to obtain 2PN-dimensional feature vectors. The concatenation result constitutes the final output of multi-granularity scanning.
[0026] Furthermore, the steps for constructing a multi-layered cascaded forest include:
[0027] The multi-layer cascaded forest contains multiple layers, and each layer of the cascaded forest contains T1 fully random forests and T2 ordinary random forests for processing data features layer by layer.
[0028] Each of the fully random forests contains t1 decision trees. A feature is randomly selected and split at each node of the decision tree until each node contains only the same class.
[0029] Each ordinary random forest contains t2 decision trees, randomly selected. A number of features are selected as candidates, where d is the number of randomly input features in the time domain or frequency domain. The information gain value is calculated and sorted, and the feature with the largest information gain value is selected to grow the tree.
[0030] Furthermore, the functional expression for the information gain is:
[0031] G ain (D FS |x i )=H(D FS )-H(D FS |x i )
[0032] in,
[0033]
[0034]
[0035]
[0036] D FSy = c ={(x,y)∈D FS |y=c}
[0037] In the formula: D FS Represents the feature set in the time or frequency domain; |D FS | represents the feature capacity, y represents the class of the feature set; x i Elements representing characteristics, where i = 1, 2, ..., n; H(D FS ) represents the feature set D FS Uncertainty of H(D) FS|xi) represents the uncertainty in classifying the feature set using DFS given that feature xi is known; Pi(D FS ) is the feature dataset D FS The proportion of feature xi to the total features; a is the proportion of feature x. i The value of D; FSxi =a is the characteristic x i The feature set; The characteristic is x i The characteristic capacity of time; c is the element x i Possible class labels; C is the set of classes of xi; D FSy =c is the set of features belonging to class c; |D FSy =c| represents the feature capacity belonging to class c.
[0038] Furthermore, the DS evidence fusion process includes:
[0039] Within a DS evidence identification framework, the basic probability assignment functions for two independent evidence bodies are m1 and m2, corresponding to propositions {A1,A2,…,As} and {B1,B2,…B}, respectively. r Using the Dempster combination rule, the basic probability allocation function of the new proposition after evidence fusion is obtained. The specific calculation formula is as follows:
[0040]
[0041] In the formula, A is A i and B j The new proposition after integration; A k Indicates the category of situation; M j (A k ) represents the basic probability assignment function for predicting the occurrence of the k-th scenario in the j-th forest; m j (A k ) represents the original output probability of the j-th forest predicting the occurrence of the k-th case; n is the total number of forests contained in the current cascade layer.
[0042] Furthermore, the output splicing includes the following steps:
[0043] The output of multi-granularity scanning is used as the input to the first layer of the cascaded forest. After prediction by two types of random forests, the class vector is output.
[0044] The randomly selected class vector from the first layer output is used as the enhanced feature vector, which is then concatenated with the high-dimensional feature vector from the multi-granularity scan output as the input to the next layer.
[0045] The enhanced feature vector output from each subsequent layer is concatenated with the high-dimensional feature vector output from the multi-granularity scan as the input to the next layer, repeating the above process.
[0046] Furthermore, the cross-validation includes the following steps:
[0047] The training set data is divided into k parts. K-1 parts are used for training, and the remaining part is used for testing. Finally, the average test error is taken as the generalization error.
[0048] Each time a new cascade layer is added, the classification accuracy of the entire cascade is evaluated on the new validation dataset. If the classification accuracy no longer increases, the addition of layers is stopped and the training process of the cascade forest is completed.
[0049] This invention also discloses a substation equipment fault identification system based on a deep forest algorithm, comprising:
[0050] The feature set acquisition module is used to perform time-frequency domain index calculation and normalization processing on the collected monitoring dataset of existing substation equipment fault signals to obtain the feature set;
[0051] The high-dimensional feature vector acquisition module is used to perform multi-granularity scanning on the feature set, obtain all feature sub-vectors through a sliding window, and output class vectors by ordinary random forest and fully random forest respectively. The high-dimensional feature vector is obtained by concatenating all the output class vectors.
[0052] The cascaded forest construction and training module is used to construct a multi-layer cascaded forest from multiple fully random forests and ordinary random forests, and input the obtained high-dimensional feature vectors into each layer of the cascaded forest for training.
[0053] The cascaded forest output module is used to perform DS evidence fusion on the multiple class probability distributions of the training output of the previous layer of the cascaded forest to obtain a fused feature vector, which is then concatenated with a randomly selected class vector and the output of the multi-granularity scan as the input of the next layer of the cascaded forest. Finally, the average value of the class distribution probabilities of each random forest is taken, and the class with the highest prediction probability is taken as the substation equipment fault prediction probability identification result.
[0054] The k-fold cross-validation module is used to perform k-fold cross-validation on the output value of each layer of the cascaded forest and calculate the recognition accuracy of each cross-validation. When the accuracy improvement is greater than the preset value, a new layer of cascaded forest is added and the entire new cascaded forest model is retrained. Otherwise, the addition of new layers is stopped and the training process of the cascaded forest is completed.
[0055] The fault identification result output module is used to process the real-time monitored substation equipment signals and input them into the trained cascaded forest model to obtain online fault identification results.
[0056] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0057] 1. This invention provides a substation equipment fault identification method based on deep forest, which uses time-domain and frequency-domain signal feature extraction methods to effectively avoid noise disturbance;
[0058] 2. This invention designs an improved deep forest algorithm based on DS evidence fusion, treating each forest as an independent evidence body, thus overcoming the feature redundancy problem encountered in the process of expanding cascaded forests;
[0059] 3. This invention uses an improved deep forest algorithm to establish a fault identification model, which effectively improves the accuracy and speed of fault identification of power equipment and meets the requirements of industrial sites for fault diagnosis of power equipment. Attached Figure Description
[0060] Figure 1 A schematic diagram of the substation equipment fault identification process based on the deep forest algorithm provided by the present invention;
[0061] Figure 2 This is a multi-granularity scanning structure diagram provided by the present invention;
[0062] Figure 3 This is a cascaded forest structure diagram after DS evidence fusion provided by the present invention;
[0063] Figure 4 This is a schematic diagram of the confusion matrix of the k-fold cross-validation results provided by the present invention;
[0064] Figure 5 This is a performance comparison chart between the substation equipment fault identification method provided by this invention and the unimproved multi-granularity cascaded forest algorithm. Detailed Implementation
[0065] The following is in conjunction with the appendix Figures 1-5 The present invention will be further described in detail with reference to specific embodiments.
[0066] like Figure 1 As shown in the figure, this embodiment proposes a substation equipment fault identification method based on the deep forest algorithm, including the following steps:
[0067] S1. Perform time-frequency domain index calculation and normalization on the collected monitoring dataset of fault signals from existing substation equipment to obtain a feature set;
[0068] S2. Perform multi-granularity scanning on the feature set, obtain all feature sub-vectors through a sliding window, and output class vectors by passing these feature sub-vectors through ordinary random forest and fully random forest respectively. Concatenate all the output class vectors to obtain a high-dimensional feature vector.
[0069] S3. Construct a multi-layer cascaded forest using multiple fully random forests and ordinary random forests, and input the obtained high-dimensional feature vectors into each layer of the cascaded forest for training.
[0070] S4. Perform DS evidence fusion on the multiple class probability distributions of the training output of the previous layer of the cascaded forest to obtain a fused feature vector, and concatenate it with a randomly selected class vector. The randomly selected class vector is obtained from the completely random forest. At the same time, it is concatenated with the output of the multi-granularity scan as the input of the next layer of the cascaded forest. Finally, take the average of the class distribution probabilities of each random forest, and finally take the class with the highest prediction probability as the substation equipment fault prediction probability identification result.
[0071] S5. Perform k-fold cross-validation on the output value of each layer of the cascaded forest, and calculate the recognition accuracy of each cross-validation. When the accuracy improvement is greater than the preset value (in this embodiment, the value is 0.1%), then add another layer of cascaded forest and retrain the entire new cascaded forest model. Otherwise, stop adding new layers and complete the training process of the cascaded forest.
[0072] S6. The real-time monitored substation equipment signals are processed and then input into the trained cascaded forest model to obtain online fault identification results.
[0073] In this embodiment, the steps for obtaining the feature set include:
[0074] Nine time-domain features and six frequency-domain features are extracted from the monitoring dataset of existing substation equipment fault signals. The time-frequency domain indicators include time-domain features and frequency-domain features. The time-domain features are root square amplitude, absolute mean, root mean square amplitude, kurtosis, skewness, waveform index, peak value index, margin index, and impulse index. The frequency-domain features are mean, standard deviation, root mean square frequency, center frequency, frequency concentration, and frequency kurtosis.
[0075] In addition, to take into account the frequency domain information of the signal, an improved energy function is used to calculate the energy and energy entropy of the sampled data.
[0076] The formula for calculating the energy as a function is as follows:
[0077]
[0078] The formula for calculating energy entropy is as follows:
[0079]
[0080] In the formula, N is the signal length; p is the amplitude factor, which is 0.02 in this embodiment, and can increase the signal peak value; x i It is the amplitude of the signal vibration point; when x iWhen the value is positive, add p; otherwise, subtract p. This embodiment uses time-domain and frequency-domain signal feature extraction methods, which can effectively avoid noise disturbances.
[0081] In this embodiment, the step of obtaining the high-dimensional feature vector includes:
[0082] For a set of samples containing K-dimensional time-domain or frequency-domain features sampled from substation data, sliding sampling is performed through a sliding window of length L, with a step size of S for each sliding. After the sliding sampling is completed, N L-dimensional feature sub-vectors are obtained, where N = (KL) / S + 1.
[0083] All the obtained feature vectors are used as inputs to ordinary random forest and fully random forest. Each random forest yields N probability vectors of dimension P.
[0084] The N probability vectors of dimension P are concatenated to obtain 2PN-dimensional feature vectors. The concatenation result constitutes the final output of multi-granularity scanning.
[0085] In this embodiment, the values of K, L, and S are determined according to the specific circumstances.
[0086] In this embodiment, the steps for constructing a multi-level cascaded forest include:
[0087] The multi-layer cascaded forest contains multiple layers. Each layer of the cascaded forest contains two fully random forests and two ordinary random forests to process data features layer by layer. Here, the number of fully random forests and ordinary random forests is the same, and their values can also be other values other than 2.
[0088] Each of the fully random forests contains t1 decision trees. A feature is randomly selected and split at each node of the decision tree until each node contains only the same class.
[0089] Each ordinary random forest contains t2 decision trees, randomly selected. 15 features are selected as candidates, where d is the number of randomly input features in the time domain or frequency domain. In this embodiment, the value of d is no greater than 15. The information gain value is calculated and sorted, and the feature with the largest information gain value is selected for the growth tree.
[0090] In this embodiment, the functional expression for information gain is:
[0091] G ain (D FS |x i )=H(D FS )-H(D FS |x i (3)
[0092] in,
[0093]
[0094]
[0095]
[0096] D FSy＝c ={(x,y)∈D FS |y=c} (7)
[0097] In the formula: D FS Represents the feature set in the time or frequency domain; |D FS | represents the feature capacity, y represents the class of the feature set; x i Elements representing characteristics, where i = 1, 2, ..., n; H(D FS ) represents the feature set D FS Uncertainty of H(D) FS |x i ) represents feature x i Given the feature set D FS Uncertainty in classification; P i (D FS ) is the feature dataset D FS Middle feature x i The proportion of the total features; a is the feature x i The value; For the feature x i The feature set; The characteristic is x i The characteristic capacity of time; c is the element x i Possible class tags; C is x i A collection of classes; D FSy＝c Let D be the set of features belonging to class c; FSy＝c | indicates the feature capacity belonging to class c.
[0098] When applying the deep forest algorithm to solve fault identification problems, directly concatenating the prediction results with multi-granularity scan results as input to the next level can lead to feature vector redundancy. To address this, an improved method based on Dempster's evidence theory is proposed. This method treats each forest as an independent evidence body and obtains the fused feature vector of all forest outputs in this cascaded layer after evidence fusion according to Dempster's combination rule. This overcomes the feature redundancy problem encountered in expanding cascaded forests.
[0099] The D-S evidence theory uses probability to measure the possibility of each proposition occurring and can handle uncertainty problems. The recognition framework of the D-S evidence theory contains a finite number of basic propositions. Under one recognition framework, the basic probability assignment functions of two independent evidence bodies are m1 and m2 respectively. In this embodiment, the two independent evidence bodies respectively correspond to a complete random forest and an ordinary random forest. The propositions corresponding to m1 and m2 are {A1, A2, …, A s} and {B1, B2, …, B r}, and the basic probability assignment function of the new proposition after evidence fusion is obtained by using the Dempster combination rule. The specific calculation formula is as follows:
[0100]
[0101] In the formula, A is the new proposition after the fusion of A i and B j ; m(A) represents the basic probability assignment function of proposition A; the intersection of the two is empty, indicating that proposition A i and B j have no common points. In formula (8), it is to screen out the completely conflicting situation of A i and B[[ID=The basic probability assignment function for each case in the current cascaded layer is obtained using the above formula. Then, the feature vector after evidence fusion is obtained according to formulas (8-10). To compensate for the information loss that may result from feature vector fusion, a class vector of a randomly selected forest decision result is concatenated with the fused feature vector, and simultaneously concatenated with the multi-granularity scan output as the input feature for the next layer. Finally, the average class probability output of each forest is the final predicted probability, and the class with the highest predicted probability is taken as the identification result.
[0108] In this embodiment, the output splicing includes the following steps:
[0109] The output of multi-granularity scanning is used as the input to the first layer of the cascaded forest. After prediction by two types of random forests, the class vector is output.
[0110] The randomly selected class vector from the first layer output is used as the enhanced feature vector, which is then concatenated with the high-dimensional feature vector from the multi-granularity scan output as the input to the next layer.
[0111] The enhanced feature vector output from each subsequent layer is concatenated with the high-dimensional feature vector output from the multi-granularity scan as the input to the next layer, repeating the above process.
[0112] In this embodiment, cross-validation includes the following steps:
[0113] The training set data is divided into k parts. K-1 parts are used for training, and the remaining part is used for testing. Finally, the average test error is taken as the generalization error.
[0114] Each time a new cascade layer is added, the classification accuracy of the entire cascade is evaluated on the new validation dataset. If the classification accuracy no longer increases, the addition of layers is stopped and the training process of the cascade forest is completed.
[0115] In this embodiment, obtaining the online fault identification result includes: averaging the class distribution probabilities of each ordinary random forest and fully random forest, outputting the final predicted probability, and taking the class with the largest predicted probability value as the substation equipment fault identification result.
[0116] This embodiment also discloses a substation equipment fault identification system based on the deep forest algorithm, including:
[0117] The feature set acquisition module is used to perform time-frequency domain index calculation and normalization processing on the collected monitoring dataset of existing substation equipment fault signals to obtain the feature set;
[0118] The high-dimensional feature vector acquisition module is used to perform multi-granularity scanning on the feature set, obtain all feature sub-vectors through a sliding window, and output class vectors by ordinary random forest and fully random forest respectively. The high-dimensional feature vector is obtained by concatenating all the output class vectors.
[0119] The cascaded forest construction and training module is used to construct a multi-layer cascaded forest from multiple fully random forests and ordinary random forests, and input the obtained high-dimensional feature vectors into each layer of the cascaded forest for training.
[0120] The cascaded forest output module is used to perform DS evidence fusion on the multiple class probability distributions of the training output of the previous layer of the cascaded forest to obtain a fused feature vector, which is then concatenated with a randomly selected class vector and the output of the multi-granularity scan as the input of the next layer of the cascaded forest. Finally, the average value of the class distribution probabilities of each random forest is taken, and the class with the highest prediction probability is taken as the substation equipment fault prediction probability identification result.
[0121] The k-fold cross-validation module is used to perform k-fold cross-validation on the output value of each layer of the cascaded forest and calculate the recognition accuracy of each cross-validation. When the accuracy improvement is greater than the preset value, a new layer of cascaded forest is added and the entire new cascaded forest model is retrained. Otherwise, the addition of new layers is stopped and the training process of the cascaded forest is completed.
[0122] The fault identification result output module is used to process the real-time monitored substation equipment signals and input them into the trained cascaded forest model to obtain online fault identification results.
[0123] Figure 4 The confusion matrix in this embodiment is compiled based on the results of a certain k-fold cross-validation. A total of 80 identifications were performed, and only 3 sets of data were not accurately identified, achieving an accuracy rate of 96.2%.
[0124] Figure 5 To compare the performance of multi-granularity cascaded forest and the method of this invention, Scheme 1 uses the unmodified multi-granularity cascaded forest algorithm, while Scheme 2 is the substation equipment fault identification method of this invention using an improved deep forest algorithm. Scheme 2 demonstrates superior performance in all five metrics: accuracy, precision, recall, F1 score, and runtime. This embodiment uses the improved deep forest algorithm to establish a fault identification model, effectively improving the accuracy and speed of fault identification for power equipment and meeting the requirements of industrial field fault diagnosis for power equipment.
[0125] The above-described invention merely illustrates implementation methods of the present invention and should not be construed as limiting the scope of the invention patent, nor as imposing any form of limitation on the structure of the embodiments of the present invention. It should be noted that those skilled in the art can make various changes and improvements without departing from the concept of the embodiments of the present invention, and these all fall within the protection scope of the embodiments of the present invention.
Claims
1. A method for substation equipment fault identification based on deep forest algorithm, characterized in that, Includes the following steps: S1. Perform time-frequency domain index calculation and normalization on the collected monitoring dataset of fault signals from existing substation equipment to obtain a feature set; S2. Perform multi-granularity scanning on the feature set, obtain all feature sub-vectors through a sliding window, and output class vectors by passing these feature sub-vectors through ordinary random forest and fully random forest respectively. Concatenate all the output class vectors to obtain a high-dimensional feature vector. S3. Construct a multi-layer cascaded forest using multiple fully random forests and ordinary random forests, and input the obtained high-dimensional feature vectors into each layer of the cascaded forest for training. S4. Perform DS evidence fusion on the multiple class probability distributions of the training output of the previous layer of the cascaded forest to obtain a fused feature vector, and concatenate it with a randomly selected class vector and the output of multi-granularity scanning as the input of the next layer of the cascaded forest. Finally, take the average of the class distribution probabilities of each random forest, and finally take the class with the highest prediction probability as the substation equipment fault prediction probability identification result. S5. Perform k-fold cross-validation on the output value of each layer of the cascaded forest, and calculate the recognition accuracy of each cross-validation. When the accuracy improvement is greater than the preset value, add another layer of cascaded forest and retrain the entire new cascaded forest model. Otherwise, stop adding new layers and complete the training process of the cascaded forest. S6. The real-time monitored substation equipment signals are processed and then input into the trained cascaded forest model to obtain online fault identification results. The steps for obtaining the feature set include: Extract multiple time-domain and frequency-domain features from the existing monitoring dataset of substation equipment fault signals; An improved energy function is used to calculate the energy and energy entropy of the sampled dataset; The steps for constructing a multi-level cascaded forest include: The multi-layer cascaded forest contains multiple layers, and each layer of the cascaded forest contains T1 fully random forests and T2 ordinary random forests for processing data features layer by layer. Each of the fully random forests contains t1 decision trees. A feature is randomly selected and split at each node of the decision tree until each node contains only the same class. Each ordinary random forest contains t2 decision trees, randomly selected.
1. A number of features are selected as candidates, where d is the number of randomly input features in the time domain or frequency domain. The information gain value is calculated and sorted, and the feature with the largest information gain value is selected to grow the tree. The DS evidence fusion process includes: Within a DS evidence identification framework, the basic probability assignment functions for two independent evidence bodies are m1 and m2, respectively, corresponding to propositions {A1, A2, ..., A...}. s } and {B1,B2,…B r Using the Dempster combination rule, the basic probability allocation function of the new proposition after evidence fusion is obtained. The specific calculation formula is as follows: In the formula, A is A i and B j The new proposition after integration; A k Indicates the category of situation; M j (A k ) represents the basic probability assignment function for predicting the occurrence of the k-th scenario in the j-th forest; m j (A k ) represents the original output probability of the j-th forest predicting the occurrence of the k-th case; n is the total number of forests contained in the current cascade layer; The formula for calculating the energy as a function is as follows: The formula for calculating energy entropy is as follows: In the formula, N is the length of the signal; p is the amplitude factor; x i It is the amplitude of the signal vibration point; when x i If the value is positive, add p; otherwise, subtract p. The functional expression for the information gain is: in, In the formula: D FS Represents the feature set in the time or frequency domain; |D FS | represents the feature capacity, y represents the class of the feature set; x i Elements representing characteristics, where i = 1, 2, ..., n; H(D FS ) represents the feature set D FS Uncertainty of H(D) FS |x i ) represents feature x i Given the feature set D FS Uncertainty in classification; P i (D FS ) is the feature dataset D FS Middle feature x i The proportion of the total features; a is the feature x i The value; For the feature x i Feature set; | | indicates that the feature is x i The characteristic capacity of time; c is the element x i Possible class tags; C is x i A collection of classes; The set of features belonging to class c; | | indicates the feature capacity belonging to class c.
2. The substation equipment fault identification method based on deep forest algorithm according to claim 1, characterized in that, The steps to obtain the high-dimensional feature vector include: For a set of samples containing K-dimensional time-domain or frequency-domain features, sliding sampling is performed through a sliding window of length L, with a step size of S for each sliding step. After the sliding sampling is completed, N L-dimensional feature sub-vectors are obtained, where N = (KL) / S + 1. All the obtained feature vectors are used as inputs to ordinary random forest and fully random forest. Each random forest yields N probability vectors of dimension P. The N probability vectors of dimension P are concatenated to obtain 2PN-dimensional feature vectors. The concatenation result constitutes the final output of multi-granularity scanning.
3. The substation equipment fault identification method based on deep forest algorithm according to claim 1, characterized in that, The output splicing includes the following steps: The output of multi-granularity scanning is used as the input to the first layer of the cascaded forest. After prediction by two types of random forests, the class vector is output. The randomly selected class vector from the first layer output is used as the enhanced feature vector, which is then concatenated with the high-dimensional feature vector from the multi-granularity scan output as the input to the next layer. The enhanced feature vector output from each subsequent layer is concatenated with the high-dimensional feature vector output from the multi-granularity scan as the input to the next layer, repeating the above process.
4. The substation equipment fault identification method based on deep forest algorithm according to claim 1, characterized in that, The cross-validation includes the following steps: The training set data is divided into k parts. K-1 parts are used for training, and the remaining part is used for testing. Finally, the average test error is taken as the generalization error. Each time a new cascade layer is added, the classification accuracy of the entire cascade is evaluated on the new validation dataset. If the classification accuracy no longer increases, the addition of layers is stopped and the training process of the cascade forest is completed.
5. A system utilizing the substation equipment fault identification method based on the deep forest algorithm as described in any one of claims 1 to 4, characterized in that, include: The feature set acquisition module is used to perform time-frequency domain index calculation and normalization processing on the collected monitoring dataset of existing substation equipment fault signals to obtain the feature set; The high-dimensional feature vector acquisition module is used to perform multi-granularity scanning on the feature set, obtain all feature sub-vectors through a sliding window, and output class vectors by ordinary random forest and fully random forest respectively. The high-dimensional feature vector is obtained by concatenating all the output class vectors. The cascaded forest construction and training module is used to construct a multi-layer cascaded forest from multiple fully random forests and ordinary random forests, and input the obtained high-dimensional feature vectors into each layer of the cascaded forest for training. The cascaded forest output module is used to perform DS evidence fusion on the multiple class probability distributions of the training output of the previous layer of the cascaded forest to obtain a fused feature vector, which is then concatenated with a randomly selected class vector and the output of the multi-granularity scan as the input of the next layer of the cascaded forest. Finally, the average value of the class distribution probabilities of each random forest is taken, and the class with the highest prediction probability is taken as the substation equipment fault prediction probability identification result. The k-fold cross-validation module is used to perform k-fold cross-validation on the output value of each layer of the cascaded forest and calculate the recognition accuracy of each cross-validation. When the accuracy improvement is greater than the preset value, a new layer of cascaded forest is added and the entire new cascaded forest model is retrained. Otherwise, the addition of new layers is stopped and the training process of the cascaded forest is completed. The fault identification result output module is used to process the real-time monitored substation equipment signals and input them into the trained cascaded forest model to obtain online fault identification results.