State perception and fault diagnosis method of switchgear for combined electrical appliances based on multiple characteristic parameters

Abstract

The state perception and fault diagnosis method of combined electrical switchgear based on multiple characteristic parameters relates to the technical field of fault diagnosis of combined electrical switchgear, and accurately diagnoses mechanical faults based on the joint characteristic values ​​of vibration, current and sound signals of GIS equipment. By calculating the abnormal rate of the characteristic value of the vibration signal sub-stage, the vibration sub-stage where the mechanical defect is located is determined, thereby narrowing the fault search range, locating the mechanical fault, and achieving a preliminary judgment on the mechanical fault. Through the SSA-LightGBM clustering learning algorithm, the characteristic vector of the multi-dimensional characteristic signal of the fusion current signal, vibration signal and sound signal is imported into the SSA-LightGBM fault diagnosis model. Compared with the traditional machine learning algorithm, it has high accuracy and good classification effect, and can achieve effective identification of different opening and closing mechanical states of GIS.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis of combined electrical switch equipment, and in particular to a state perception and fault diagnosis method of combined electrical switch equipment based on multiple characteristic parameters. Background Art

[0002] Since GIS equipment entered the power market, scholars have conducted a large number of studies on GIS discharge fault diagnosis. Only in recent years have they begun to study the mechanical failure of GIS equipment. The main research method is to simulate common mechanical defects of GIS on GIS equipment in the laboratory, and then analyze the vibration or sound signals of GIS to determine the type of mechanical failure of GIS. Common detection methods include vibration signal detection method and sound signal detection method. (1) Vibration signal detection method: The vibration signal is an important representation of the mechanical state of the equipment. When the GIS is operating normally, due to the action of electric force, its shell will generate a vibration signal with a base frequency of 100 Hz. When the GIS equipment has mechanical defects such as improper installation and component wear, the GIS will generate long-term abnormal mechanical vibration. The existence of abnormal vibration can be detected by collecting and analyzing the shell vibration signal. (2) Sound signal detection method: When there is a mechanical failure inside the GIS, not only abnormal vibration will be generated, but also abnormal noise will be caused by the equipment. Experienced on-site staff can often judge the mechanical failure of the equipment by listening to the sound of the equipment. The sound signal detection method is to use sound sensors instead of human ears to collect the sound of the equipment, and identify the mechanical state of the equipment through signal feature extraction and fault diagnosis model. Compared with the vibration signal detection method, the sound signal detection is not easily affected by the sensor installation position, and it is non-contact installation, which is more convenient in practical applications. Although vibration signals and sound signals have many applications in the field of mechanical fault diagnosis of GIS equipment, the vibration signal detection method has the disadvantages of inconvenient installation and the detection results are greatly affected by the sensor installation position, while the sound signal detection method has the disadvantage of being easily disturbed by environmental noise. Therefore, better diagnostic effects can be achieved by combining vibration and sound signals for fault diagnosis.

[0003] At present, there are many diagnostic methods for GIS insulation faults, and they are relatively mature. Mechanical faults are the main faults of GIS equipment, but the current research on them is not sufficient. Mechanical faults are often not easy to detect in the early stage, but as the faults evolve, they may cause insulation discharge in the equipment, thereby causing greater equipment failure. Therefore, GIS mechanical defects cannot be ignored. Although mechanical defects will not cause any faults in the early stage, they will evolve into serious electrical faults in the long run, endangering the safe and stable operation of the power grid. Therefore, it is crucial to conduct a systematic study of various mechanical faults of GIS equipment and propose appropriate fault diagnosis methods based on the characteristics of mechanical faults.

[0004] In the field of fault diagnosis, the use of multi-sensor information fusion technology is more accurate than the fault diagnosis results of a single signal. In recent years, it has been widely used in the condition monitoring of electrical equipment. For example, the current defects in diagnosing mechanical faults of GIS circuit breakers are:

[0005] (1) The current mechanical fault diagnosis methods for circuit breakers are all for conventional circuit breakers, and there are few studies on the mechanical fault diagnosis of circuit breakers inside GIS.

[0006] (2) There is still a lack of in-depth research on the location and identification of early mechanical defects in circuit breaker operating mechanisms. Most methods for analyzing circuit breaker vibration signals are traditional linear time series analysis methods, while nonlinear time series analysis methods (such as phase space reconstruction technology based on chaos theory) are more applicable to circuit breaker vibration signals. The current method of extracting fault features by phase space reconstruction of circuit breaker vibration signals is not accurate enough in locating circuit breaker defects.

[0007] (3) Currently, the monitoring of circuit breakers is mostly single signal monitoring and analysis. The monitoring of circuit breaker mechanical failures is not comprehensive enough and is easily affected by the on-site environment, resulting in inaccurate diagnostic results and low reliability. Summary of the invention

[0008] The purpose of the present invention is to provide a state perception and fault diagnosis method for combined electrical switchgear based on multiple characteristic parameters. First, the vibration sub-stage where the mechanical defect is located is located and preliminarily identified by comparing the abnormal rate of the characteristic value of the vibration signal with the reference value, and then the GIS mechanical fault is accurately diagnosed based on the joint characteristic value of the vibration, current and sound signals of the GIS equipment.

[0009] The technical solution adopted by the present invention to solve the technical problem is: a method for state perception and fault diagnosis of a combination electrical switch device based on multiple characteristic parameters, comprising the following steps:

[0010] S1. Multi-dimensional signal feature acquisition

[0011] S1.1. Collect vibration signals, current signals and sound signals of the opening and closing states of GIS equipment;

[0012] S1.2. Division of GIS equipment vibration sub-stages

[0013] According to the action sequence of the components of the operating mechanism of the GIS equipment, the vibration signal is processed in time division to obtain four vibration sub-stages, namely: stage 1-detent tripping stage, stage 2-spring release stage, stage 3-mechanism transmission stage and stage 4-mechanism braking stage;

[0014] S1.3. Constructing a fault dataset

[0015] Extract characteristic parameters of vibration signal, current signal and sound signal and combine them to obtain multi-dimensional signal characteristic vector E;

[0016] S2. GIS equipment defect location

[0017] S2.1, extracting the vibration signals of each sub-stage in step S1.2;

[0018] S2.2, calculate the chaotic feature LLE and phase trajectory features ACD and VDO of the vibration signal of the vibration sub-stage according to the phase space reconstruction method;

[0019] S2.3, calculating the characteristic value abnormality rate k, that is, the percentage of the deviation between the characteristic value of the unknown signal vibration sub-stage and the characteristic value of the normal signal corresponding vibration sub-stage;

[0020] (1)

[0021] in, represents the characteristic value of the signal to be measured, Represents the characteristic value of a normal signal.

[0022] Compare the size of the characteristic value abnormal rate k and the abnormal rate limit value. When the characteristic value abnormal rate k of a certain vibration sub-stage is greater than the abnormal rate limit value, it indicates that the fault is located in the vibration sub-stage;

[0023] S3. Calculate the F1 score based on the LightGBM model to optimize the LightGBM model

[0024] S3.1. Build LightGBM model

[0025] The LightGBM model is obtained by optimizing the gradient boosting tree algorithm through the unilateral gradient sampling algorithm, the mutually exclusive feature bundling algorithm, the histogram algorithm and the leaf-by-leaf growth strategy with depth limit.

[0026] S3.2. Select F1 score as the indicator for LightGBM model parameter optimization

[0027] S3.3. Optimize LightGBM model parameters

[0028] Use the sparrow search algorithm to optimize the hyperparameters of LightGBM: learning rate learning_rate, tree depth max_depth, number of leaf nodes num_leaves and subsample. Subsample is used to train weak learners to obtain the optimal value of each parameter.

[0029] S4. Diagnosis of GIS mechanical faults based on SSA-LightGBM fault diagnosis model

[0030] The vibration signal, current signal and sound signal are input into the SSA-LightGBM fault diagnosis model to obtain the GIS circuit breaker status.

[0031] Furthermore, the specific steps of step S1.3 are:

[0032] S1.3.1, taking the seven time domain features of the GIS equipment opening and closing coil current signal extracted as the feature vector E1;

[0033] S1.3.2, the chaotic features and phase trajectory features of the four sub-stages of the GIS equipment vibration signal, a total of 12 dimensions, are used as the feature vector E2;

[0034] S1.3.3, taking the 9-dimensional F-GFCC coefficients of the sound signal after PCA dimensionality reduction processing as the feature vector E3;

[0035] S1.3.4. Finally, the feature vectors of the current signal, vibration signal and sound signal are combined to obtain the multi-dimensional signal feature vector E.

[0036] Furthermore, the steps of the gradient boosting tree algorithm include:

[0037] Input a given training set , the maximum number of iterations is M, and the loss function is ;

[0038] (1) Initialize the model, c is a constant

[0039] (1)

[0040] (2) For the number of iterations hour:

[0041] a. Calculate the negative gradient of the current iteration:

[0042] (2)

[0043] b. Based on the training set Construct a CART decision tree. For each leaf node of the t-th decision tree, the minimum output value of the loss function is calculated according to the following formula:

[0044] (3)

[0045] in, represents the leaf node region of the mth CART decision tree, and J is the number of leaf nodes;

[0046] c. The calculation formula for the updated model is as follows:

[0047] (4)

[0048] I is the indicator function. When the data belongs to R m j When I=1, otherwise I=0;

[0049] (3) Calculate the final reinforcement learning machine

[0050] (5)

[0051] Furthermore, according to the true category of the sample and the prediction result category of the fault diagnosis model, it is divided into four cases: true positive (TP), false positive (FP), true negative (TN) and false negative (FN). The F1 score indicator F_score is the harmonic mean of precision and recall, and the calculation formula is as follows:

[0052] (6)

[0053] Among them, P is the precision rate and R is the recall rate;

[0054] Precision is the ratio of the number of correctly classified results in a certain category to all classification results of that category. The calculation formula is as follows:

[0055] (7)

[0056] The recall rate is the ratio of the number of correctly classified samples in a certain actual category to all classification results of that category. The calculation formula is as follows:

[0057] (8)

[0058] Furthermore, the sparrow search algorithm optimizes the parameters in the following way:

[0059] The finder's location is updated as follows:

[0060] (9)

[0061] Where t is the current number of iterations; T is the maximum number of iterations; α is a uniform random number between (0,1); Q is a random number that obeys the standard normal distribution; L is a 1×d matrix with all elements being 1; R2∈[0,1] represents the warning value; ST∈[0.5,1] represents the safety value;

[0062] The joiner position is updated as follows:

[0063] (10)

[0064] In the formula, represents the global worst position at the tth iteration; represents the best position of the discoverer at the t+1th iteration; N represents the population size; A represents a 1×d matrix with each element randomly assigned to 1 or -1, and satisfies ;

[0065] The Scout positions have been updated as follows:

[0066] (11)

[0067] In the formula, represents the global optimal position at the tth iteration; β represents the step size control parameter, which is a normally distributed random number with a mean of 0 and a variance of 1; K is a random number between [-1,1]; f i represents the current sparrow's fitness; f g represents the optimal fitness of the current sparrow population; f w Represents the worst fitness of the current sparrow population.

[0068] The beneficial effects of the present invention are: by calculating the abnormal rate of the characteristic value of the vibration signal sub-stage, the vibration sub-stage where the mechanical defect is located is determined, thereby narrowing the fault search range, locating the mechanical fault, and achieving a preliminary judgment of the mechanical fault. Through the SSA-LightGBM clustering learning algorithm, the characteristic vector of the multi-dimensional characteristic signal of the current signal, vibration signal and sound signal is integrated into the SSA-LightGBM fault diagnosis model. Compared with the traditional machine learning algorithm, it has high accuracy and good classification effect, and can effectively identify the different opening and closing mechanical states of GIS. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 Calculate the characteristic value abnormal rate diagram for each stage of the electromagnet loose fault during the closing phase;

[0070] Figure 2 It is the abnormal rate diagram of characteristic value of each stage of the mechanical defect signal of the opening brake;

[0071] Figure 3 It is the histogram algorithm diagram;

[0072] Figure 4 is the layer-by-layer growth strategy diagram;

[0073] Figure 5 This is a leaf growth strategy diagram;

[0074] Figure 6 In order to use the SSA algorithm to find the optimal parameters of LightGBM, a GIS mechanical fault diagnosis model based on SSA-LightGBM was constructed, and a flowchart for fault type identification was performed for different opening and closing mechanical states of GIS equipment;

[0075] Figure 7This is the classification result diagram of the test set;

[0076] Figure 8 Classification confusion matrix diagram for the SSA-LightGBM fault diagnosis model;

[0077] Fig. 9 This is the main hyperparameters and optimal value diagram of the LightGBM model;

[0078] Fig.10 Label diagrams for different types of samples;

[0079] Fig.11 Diagnosis result diagram of test set samples;

[0080] Fig.12 Diagnostic accuracy graphs for different models. DETAILED DESCRIPTION

[0081] In conjunction with the accompanying drawings, taking a GIS circuit breaker as an example, the state perception and fault diagnosis method of a combined electrical switch device based on multiple characteristic parameters of the present invention is described in detail:

[0082] S1. Multi-dimensional signal feature acquisition

[0083] S1.1. Collect vibration signals, current signals and sound signals of the opening and closing states of GIS circuit breaker.

[0084] S1.2, GIS circuit breaker vibration sub-stage division

[0085] According to the action sequence of the components of the GIS operating mechanism, the vibration signal is processed in time division to obtain four vibration sub-stages.

[0086] Stage 1: latch tripping stage (0-10ms)

[0087] The events included in the latch tripping stage are that the circuit breaker receives the trip command, the trip electromagnet moving iron core is attracted, and the trip guide rod is driven to hit the trip latch;

[0088] Phase 2: Spring release phase (10-50ms)

[0089] After the tripping latch is tripped during the spring release phase, the tripping spring energy is released and the mechanism is decoupled;

[0090] Stage 3: Mechanism transmission stage (50-60ms)

[0091] During the mechanism transmission stage, the energy released by the spring causes the main shaft transmission shaft to rotate through the crank arm, driving the circuit breaker contacts to separate;

[0092] Stage 4: Mechanism braking stage (60-200ms)

[0093] During the braking stage of the mechanism, after the moving contact moves to the lowest position, it continues to oscillate under the joint action of the buffer and the opening spring, and the moving contact gradually returns to stability, and the opening operation is completed.

[0094] S1.3. Constructing a fault dataset

[0095] Extract characteristic parameters of vibration signals, current signals and sound signals.

[0096] S1.3.1, taking the 7 time domain features of the circuit breaker opening and closing coil current signal extracted as the feature vector E1;

[0097] S1.3.2, the chaotic features and phase trajectory features of the four sub-stages of the GIS circuit breaker vibration signal, a total of 12 dimensions, are used as the feature vector E2;

[0098] S1.3.3. The 9-dimensional F-GFCC (Fisher weighted Gammatone Filter Cepstral Coefficient) coefficients of the sound signal after the PCA (Principle Component Analysis) dimensionality reduction process are used as the feature vector E3;

[0099] S1.3.4. Finally, the feature vectors of the current signal, vibration signal and sound signal are combined to obtain the multi-dimensional signal feature vector E.

[0100] S2. GIS circuit breaker defect location

[0101] S2.1, extracting the vibration signals of each sub-stage in step S1.2;

[0102] S2.2. The chaotic feature LLE and phase trajectory features ACD and VDO of the vibration signal of the vibration sub-stage are calculated according to the phase space reconstruction method.

[0103] S2.3, calculating the characteristic value abnormality rate k, that is, the percentage of the deviation between the characteristic value of the vibration sub-stage of the unknown signal and the characteristic value of the corresponding vibration sub-stage of the normal signal;

[0104] (1)

[0105] in, represents the characteristic value of the signal to be measured, Represents the characteristic value of a normal signal.

[0106] Compare the size of the characteristic value abnormal rate k and the abnormal rate limit value. When the characteristic value abnormal rate k of a certain vibration sub-stage is greater than the abnormal rate limit value, it indicates that the fault is located in the vibration sub-stage.

[0107] For example, the characteristic value abnormality rate is calculated for each stage of the electromagnet loosening fault during the closing phase, such as Figure 1 As shown. Figure 1 It can be seen that the abnormality rates of the three eigenvalues ​​in stage 1 are all above 20%, while the abnormality rates of the eigenvalues ​​in the other stages are all below 10%. The abnormality rate limit is set to 15%. If the eigenvalue abnormality rate exceeds 15%, the signal of the sub-stage is considered abnormal, thereby locating the mechanical defect. In practical engineering applications, the vibration signal can be processed in time division, and the abnormality rate of the eigenvalues ​​of each vibration sub-stage can be calculated to determine the sub-stage where the mechanical defect occurs, and then the mechanical parts that act in the sub-stage can be checked, so that the mechanical fault parts can be quickly and accurately located.

[0108] For another example, after extracting the characteristic values ​​from the opening signal in stages, it is found that the characteristic value of the vibration signal under the jamming defect of the operating mechanism in the transmission stage is significantly different from that in the normal state, while the characteristic value of the vibration signal under the loosening defect of the opening buffer in the braking stage is significantly different from that in the normal state. The characteristic value abnormality rate k of each vibration sub-stage under the two operating states is calculated, as follows: Figure 2 As shown. It can be seen from the characteristic value abnormality rate change diagram that for the vibration signal under the operating mechanism jam defect, the characteristic value abnormality rate in the mechanism transmission stage is the largest and greater than 15%, while the other stages are all below the limit value of 15%. Therefore, it can be located that the mechanical defect occurs in the mechanism transmission stage. For the vibration signal under the loose defect of the opening brake buffer, only the characteristic value abnormality rate in the mechanism braking stage is above the limit value, which can be located that the mechanical defect occurs in the mechanism braking stage. The buffer is the main moving component in the mechanism braking stage. Therefore, the on-site staff can quickly locate that the opening brake buffer has an abnormality. The jam defect of the operating mechanism mainly affects the chaotic characteristics and attractor morphology of the opening brake mechanism transmission stage, while the loose defect of the buffer mainly affects the chaotic characteristics and attractor morphology of the opening brake mechanism braking stage, which is consistent with the action characteristics of the mechanical parts of the circuit breaker.

[0109] By extracting the attractor morphological features of each stage of the vibration signal and comparing them with the normal operation, the method of calculating the abnormal rate of the characteristic value of each vibration sub-stage helps to quickly determine the vibration sub-stage where the abnormal signal is located. According to the action process of the circuit breaker components, the faulty components are preliminarily located. However, this method of locating the vibration sub-stage based on the abnormal rate of the characteristic value is only a method for preliminary fault positioning, and the positioning only uses the vibration signal. The GIS circuit breaker mechanical fault types that can be identified and the accuracy of identification are relatively limited. Therefore, in order to improve the accuracy of the fault diagnosis results and obtain more reliable identification results, it is still necessary to establish a multi-feature parameter fusion feature vector, and establish a reliable fault diagnosis model through a machine learning algorithm to achieve a comprehensive and accurate diagnosis of GIS mechanical faults.

[0110] S3. Calculate the F1 (balanced F score) score index based on the LightGBM (Light Gradient Boosting Machine) model to optimize the LightGBM model

[0111] S3.1. Build LightGBM model

[0112] In machine learning algorithms, in order to improve the performance of a single learner, it is proposed to build multiple single learners to complete the learning task, thereby obtaining better performance. This method is ensemble learning. The idea of ​​ensemble learning is that even if the prediction result of one of the weak classifiers is wrong, it can be corrected by other classifiers to obtain the correct classification result. Ensemble learning is mainly divided into two categories, Bagging and Boosting.

[0113] The essence of bagging is to randomly sample a sample space with replacement to obtain some independent training samples, which can enhance the perturbation of the samples. After multiple rounds of sampling training, multiple estimates will be formed, and the multiple estimates will be averaged to reduce the variance of the estimates, thereby enhancing the generalization characteristics of the learner.

[0114] The main idea of ​​Boosting is to assemble weak classifiers into a strong classifier, train a base learner through the initial training set, adjust the distribution of training samples according to the performance of the base learner, and obtain the final reinforcement learner. Boosting is widely used, among which Adaptive Boosting (AdaBoost), Gradient Boosting Decision Tree (GBDT) and Extreme Gradient Boosting (XGBoost) are more widely used. GBDT, as an iterative decision tree algorithm, consists of multiple decision trees, usually CART (classification and regression tree) regression tree, also known as the base learner of GBDT. GBDT makes the final conclusion by summing up the conclusions of all trees. When it was first proposed, it was considered to be an algorithm with strong generalization ability together with SVM (Support Vector Machines).

[0115] The basic idea of ​​GBDT is: in the iterative process, each round of iteration generates a weak classifier, and each classifier is trained based on the residual of the previous round of classifiers. Assume that the strong learner obtained in the previous round of iteration is , the loss function is The goal of this iteration is to find a weak classifier for the CART regression tree model. , so that the loss function of this round Min. h m (x) is the mth decision tree, f m (x) = f m-1 (x)+h m (x).

[0116] For the loss function The GBDT algorithm proposes to fit the approximate value of the loss through the negative gradient value of the loss function, and then fit the CART regression tree. In the iterative process, in the mth round of iteration, the negative gradient of the loss function of the i-th input sample is expressed by the following formula:

[0117] (2)

[0118] g m k is the negative gradient.

[0119] The process of the GBDT algorithm is as follows:

[0120] Input a given training set , the maximum number of iterations is M, and the loss function is .

[0121] (1) Initialize the model, C is the mean of all training sample label values, which is a constant; f o (x) is the initial learner.

[0122] (3)

[0123] (2) For the number of iterations hour:

[0124] a. Calculate the negative gradient of the current iteration:

[0125] (4)

[0126] b. Based on the training set Construct a CART decision tree. For the jth leaf node of the mth decision tree, the leaf node parameter is calculated according to the following formula:

[0127] (5)

[0128] in, represents the leaf node area of ​​the mth CART decision tree, J is the number of leaf nodes, C m j is the parameter of the j-th leaf node of the m-th decision tree.

[0129] c. The calculation formula for the updated model is as follows:

[0130] (6)

[0131] I is the indicator function. When the data belongs to R m j When I=1, otherwise I=0.

[0132] (3) Calculate the final reinforcement learning machine

[0133] (7)

[0134] LightGBM (Light Gradient Boosting Machine) is an improvement on the GBDT algorithm in ensemble learning. LightGBM optimizes the GBDT algorithm through the Gradient-based One-Side Sampling (GOSS) algorithm, the Exclusive Feature Bundling (EFB) algorithm, the histogram algorithm, and the leaf-by-leaf growth strategy with depth restrictions.

[0135] The basic idea of ​​the GOSS algorithm is to retain samples with large gradients and randomly sample samples with small gradients. First, sort the samples according to the absolute value of the gradient of the loss function, take the first a% of large gradient data as a subset A, and then randomly sample b% of the data from the remaining small gradient samples as a subset B; then, multiply the small gradient samples by a constant , so that the algorithm will pay more attention to the samples that are not well trained, and there is no need to change the distribution of the original data too much. High-dimensional data is usually very sparse, and there is mutual exclusivity between features, which has a certain impact on the effectiveness and running speed of the model.

[0136] The basic idea of ​​the EFB algorithm is to reduce the feature dimension by bundling mutually exclusive features into a single feature, thereby solving the sparsity problem of high-dimensional data.

[0137] In order to solve the problem of low efficiency and poor scalability caused by the need to traverse the information gain of all data when the GBDT algorithm is looking for suitable segmentation points, LightGBM uses a leaf-by-leaf growth strategy with a depth limit to search for the optimal segmentation point set using a histogram.

[0138] (1) The GBDT algorithm searches for the optimal segmentation point by first pre-sorting the data, enumerating all feature points on the sorted features, and then searching for the optimal segmentation point based on information gain. This is inefficient. The basic idea of ​​the histogram algorithm is to discretize the continuous feature values ​​into K integers and construct a histogram with a width of K, such as Figure 3 As shown in the figure. In the process of traversing the data, the discretized values ​​are used as indexes to accumulate statistics in the histogram. After traversing the data once, the histogram accumulates the required statistics, and then the optimal segmentation point is found based on the discrete values ​​of the histogram. The number of histograms is much smaller than the number of data, so the time and memory consumed in finding the optimal segmentation point through the histogram algorithm is smaller.

[0139] (2) The GBDT algorithm uses a layer-by-layer growth strategy to split leaves on the same layer at the same time. Although this can avoid the overfitting problem, the layer-by-layer growth strategy does not distinguish between leaves on the same layer, which increases the difficulty of data processing and reduces the computational efficiency of the algorithm. The LightGBM algorithm uses a leaf-like growth strategy, splitting only the leaf nodes with the largest information gain among all leaves at a time, while other leaf nodes are not split. Compared with the traditional layer-by-layer splitting method, this splitting method has a smaller error, a faster learning speed, and improves the classification accuracy of the algorithm. It prevents overfitting, and the LightGBM algorithm limits the maximum depth parameter by adding a maximum depth parameter during the training process. The layer-by-layer growth strategy is as follows: Figure 4 As shown, the leaf growth strategy is as follows Figure 5 shown.

[0140] Assume a training set consisting of N samples ,in Represents the input data, Represents the output result. The purpose of the LightGBM algorithm is to find a mapping relationship To approximate the function , so that the loss function Minimum. The objective function is expressed as:

[0141] (8)

[0142] in, is the loss function, Represents the regularization term. Different from the fast descent method of GBDT, LightGBM uses the Newton method to quickly approximate the objective function. Formula (8) is derived as follows:

[0143] (9)

[0144] in, represents the first-order loss function, Represents the second-order loss function. The formula is as follows:

[0145] (10)

[0146] (11)

[0147] The information gain in LightGBM is as follows:

[0148] -γ(12)

[0149] H-γ is the information gain, γ is the complexity penalty term for splitting nodes to prevent the generation of too many nodes; λ is the regularization parameter used to prevent overfitting, I R is the set of leaf nodes on the left, I L is the set of leaf nodes on the right.

[0150] S3.2. Select indicators for LightGBM model parameter optimization

[0151] Taking the binary classification problem as an example, the evaluation indicators of the LightGBM model are introduced. For the binary classification problem, according to the real category of the sample and the prediction result category of the fault diagnosis model, it is divided into four cases: true positive (TP), false positive (FP), true negative (TN), and false negative (FN). The sum of the samples in the four cases is the total number of samples. The following is an introduction to the three commonly used model performance evaluation indicators of precision, recall and F1 score:

[0152] (1) Precision (P)

[0153] Precision is the ratio of the number of correctly classified results in a certain category to all classification results of that category. The calculation formula is as follows:

[0154] (13)

[0155] (2) Recall (R)

[0156] The recall rate is the ratio of the number of correctly classified samples in a certain actual category to all classification results of that category. The calculation formula is as follows:

[0157] (14)

[0158] (3) F1 score indicator (F_score)

[0159] The F1 score combines the precision and recall indicators, which is the harmonic mean of the precision and recall. The calculation formula is as follows:

[0160] (15)

[0161] Among them, the F1 score indicator comprehensively considers the precision and recall rate, and can more comprehensively evaluate the performance of the model. The higher the F1 score, the better the performance of the model. Therefore, the F1 score indicator is selected as the fitness function for LightGBM model parameter optimization.

[0162] S3.3. Optimize LightGBM model parameters

[0163] The LightGBM model has many hyperparameters. The values ​​of the hyperparameters will affect the diagnostic effect of the model. Therefore, it is necessary to optimize the parameters to obtain the best fault diagnosis model. Select the following LightGBM model hyperparameters for optimization:

[0164] (1) learning_rate represents the learning rate. Setting it too small will result in a very slow gradient descent, while setting it too large will cross the optimal value.

[0165] (2) max_depth represents the depth of the tree. Set a reasonable value to avoid the generated tree being too deep.

[0166] (3) num_leaves represents the number of leaf nodes and is used to adjust the complexity of the tree;

[0167] (4) Subsample is used to train weak learners. If the value is too small, overfitting may occur.

[0168] After optimizing the above four hyperparameters of LightGBM using the sparrow search algorithm, the optimal values ​​of each parameter are obtained as follows Fig. 9 shown.

[0169] The Sparrow Search Algorithm (SSA) is an intelligent swarm optimization algorithm proposed by imitating the biological behaviors of sparrows such as group foraging and anti-predation. The principle of the Sparrow Search Algorithm is as follows: the foraging process of sparrows can be considered as a discoverer-joiner model, and a reconnaissance and early warning mechanism is added to it, that is, some sparrows are selected as guards. The position of each sparrow in the population corresponds to a solution. There are three behaviors in the foraging process of sparrows:

[0170] ① As a discoverer with higher fitness, it looks for food and guides the population to search and forage;

[0171] ② In order to gain higher fitness, the joiners follow the discoverers to forage;

[0172] ③ For the safety of the population, the sentinels remind the population to take anti-predator actions when danger occurs.

[0173] The finder's location is updated as follows:

[0174] (16)

[0175] Where t is the current number of iterations; T is the maximum number of iterations; α is a uniform random number between (0,1); Q is a random number that follows a standard normal distribution; L is a 1×d matrix with all elements being 1; R2∈[0,1] represents the warning value; and ST∈[0.5,1] represents the safety value.

[0176] The joiner position is updated as follows:

[0177] (17)

[0178] In the formula, represents the global worst position at the tth iteration; represents the best position of the discoverer at the t+1th iteration; N represents the population size; A represents a 1×d matrix with each element randomly assigned to 1 or -1, and satisfies .

[0179] The Scout positions have been updated as follows:

[0180] (18)

[0181] In the formula, represents the global optimal position at the tth iteration; β represents the step size control parameter, which is a normally distributed random number with a mean of 0 and a variance of 1; K is a random number between [-1,1]; f i represents the current sparrow's fitness; f g represents the optimal fitness of the current sparrow population; f w Represents the worst fitness of the current sparrow population. ε is a small constant that does not affect the calculation result and is used to ensure that the denominator is not 0.

[0182] S4. Diagnosis of GIS mechanical faults based on SSA-LightGBM fault diagnosis model

[0183] The vibration signal, current signal and sound signal are input into the SSA-LightGBM (Sparrow Search Algorithm-Light Gradient Boosting Machine) fault diagnosis model to obtain the GIS circuit breaker status. The following specific experiments are used as examples to illustrate the superiority of the SSA-LightGBM fault diagnosis model:

[0184] The experiment collected current signals, vibration signals and sound signals under five GIS working conditions: normal closing, normal opening, closing electromagnet loosening, operating mechanism stuck, and buffer loosening. 60 sets of data were collected under each working condition, for a total of 300 sets of data. The data were numbered, with groups 1-60 being normal closing data, 61-120 being normal opening data, 121-180 being closing electromagnet loosening data, 181-240 being opening mechanism stuck data, and 241-300 being opening buffer loosening data. Each set of data was labeled in the order of 1-5, and the labels of the five sets of samples after labeling were as follows: Fig.10 shown.

[0185] The feature vectors of all data are mixed, 80% are randomly selected as the training set, and the remaining 20% ​​are used as the test set, that is, 240 sets of data are used as the training set and 60 sets of data are used as the test set. Among them, there are 10 sets of test samples in the normal closing state, 18 sets of test samples in the normal opening state, 11 sets of test samples in the closing electromagnet loose state, 12 sets of test samples in the opening mechanism stuck state, and 9 sets of test samples in the opening buffer loose state. The fault diagnosis results of the test set under various working conditions are shown in Fig.11 , the classification results are as follows Figure 7 shown.

[0186] In order to better evaluate the performance of the SSA-LightGBM fault diagnosis model, the confusion matrix of the model classification is shown as follows Figure 8 As shown. The horizontal axis of the confusion matrix is ​​the predicted value of the GIS working state, the vertical axis is the true value of the GIS working state, and the main diagonal is the proportion of samples with the same predicted value and actual value. From the confusion matrix, we can see that one normal opening sample was mistakenly identified as a stuck opening mechanism, one sample with a loose closing electromagnet was identified as a normal closing state, and one sample with a stuck opening mechanism was identified as a normal opening state. The other samples can be accurately identified. The total fault recognition accuracy of the calculated test set samples reached 95%, and good diagnostic results were achieved.

[0187] Finally, the SSA-LightGBM fault diagnosis model was trained and tested through the five-fold cross-validation principle, and the average accuracy of the five tests was calculated to be 95.3%. It can be seen that the SSA-LightGBM fault diagnosis model can effectively identify different working conditions of GIS.

[0188] In order to verify the superiority of the SSA-LightGBM fault diagnosis model, GBDT and LightGBM without optimized parameters were also trained and tested using five-fold cross validation. The accuracy results obtained were compared with the SSA-LightGBM diagnostic accuracy results. Fig.12As shown. From the diagnostic comparison results, it can be seen that the SSA-LightGBM fault diagnosis model has a higher recognition accuracy for the mechanical state of GIS equipment opening and closing under multiple feature parameters. Compared with the GBDT algorithm, the performance of the LightGBM algorithm has been improved, and the recognition accuracy of samples has also been improved. The SSA-LightGBM fault diagnosis model optimizes the parameters of LightGBM through the sparrow algorithm, so that the model's recognition accuracy for experimental samples has been further improved, which can meet the needs of distinguishing different GIS mechanical defects.

[0189] The present invention determines the vibration sub-stage where the mechanical defect is located by calculating the abnormal rate of the characteristic value of the vibration signal sub-stage, thereby narrowing the fault search range, locating the mechanical fault, and achieving a preliminary judgment of the mechanical fault. Through the SSA-LightGBM clustering learning algorithm, the characteristic vector of the multidimensional characteristic signal of the fusion current signal, vibration signal and sound signal is imported into the SSA-LightGBM fault diagnosis model. Compared with the traditional machine learning algorithm, it has high accuracy and good classification effect, and can effectively identify the different opening and closing mechanical states of GIS.

Claims

1. A method for state perception and fault diagnosis of combined electrical switchgear based on multiple characteristic parameters, characterized in that: The following steps are involved: S1. Multi-dimensional signal feature acquisition S1.

1. Collect vibration signals, current signals and sound signals of the opening and closing states of GIS equipment; S1.

2. Division of GIS equipment vibration sub-stages According to the action sequence of the components of the operating mechanism of the GIS equipment, the vibration signal is processed in time division to obtain four vibration sub-stages, namely: stage 1-detent tripping stage, stage 2-spring release stage, stage 3-mechanism transmission stage and stage 4-mechanism braking stage; S1.

3. Constructing a fault dataset Extract characteristic parameters of vibration signal, current signal and sound signal and combine them to obtain multi-dimensional signal characteristic vector E; S2. GIS equipment defect location S2.1, extracting the vibration signals of each sub-stage in step S1.2; S2.2, calculate the chaotic feature LLE and phase trajectory features ACD and VDO of the vibration sub-stage signal according to the phase space reconstruction method; S2.3, calculating the characteristic value abnormality rate k, that is, the percentage of the deviation between the characteristic value of the unknown signal vibration sub-stage and the characteristic value of the normal signal corresponding vibration sub-stage; Among them, T e Represents the characteristic value of the signal to be measured, N e The characteristic value representing the normal signal; Compare the size of the characteristic value abnormal rate k and the abnormal rate limit value. When the characteristic value abnormal rate k of a certain vibration sub-stage is greater than the abnormal rate limit value, it indicates that the fault is located in the vibration sub-stage; S3. Calculate the F1 score based on the LightGBM model to optimize the LightGBM model S3.

1. Build LightGBM model The LightGBM model is obtained by optimizing the gradient boosting tree algorithm through the unilateral gradient sampling algorithm, the mutually exclusive feature bundling algorithm, the histogram algorithm and the leaf-by-leaf growth strategy with depth limit. S3.

2. Select F1 score as the indicator for LightGBM model parameter optimization S3.

3. Optimize LightGBM model parameters Use the sparrow search algorithm to optimize the hyperparameters of LightGBM: learning rate learning_rate, tree depth max_depth, number of leaf nodes num_leaves and subsample. Subsample is used to train weak learners to obtain the optimal value of each parameter. S4. Diagnosis of GIS mechanical faults based on SSA-LightGBM fault diagnosis model The vibration signal, current signal and sound signal are input into the SSA-LightGBM fault diagnosis model to obtain the GIS circuit breaker status.

2. The method for state perception and fault diagnosis of combined electrical switchgear based on multiple characteristic parameters according to claim 1 is characterized in that: The specific steps of step S1.3 are: S1.3.1, taking the seven time domain features of the GIS equipment opening and closing coil current signal extracted as the feature vector E1; S1.3.2, the chaotic features and phase trajectory features of the four sub-stages of the GIS equipment vibration signal, a total of 12 dimensions, are used as the feature vector E2; S1.3.3, taking the 9-dimensional F-GFCC coefficients of the sound signal after PCA dimensionality reduction processing as the feature vector E3; S1.3.

4. Finally, the feature vectors of the current signal, vibration signal and sound signal are combined to obtain the multi-dimensional signal feature vector E.

3. The method for state perception and fault diagnosis of combined electrical switchgear based on multiple characteristic parameters according to claim 1 is characterized in that: The steps of the gradient boosting tree algorithm include: Input a given training set The maximum number of iterations is M, and the loss function is L(y, F(x)); (1) Initialize the model, c is the mean of all training sample label values, which is a constant; (2) For the number of iterations m∈[1,M]: a. Calculate the negative gradient of the current iteration: b. Based on the training set Construct a CART decision tree. For each leaf node of the t-th decision tree, the minimum output value of the loss function is calculated according to the following formula: in, represents the leaf node region of the mth CART decision tree, and J is the number of leaf nodes; c. The calculation formula for the updated model is as follows: I is the indicator function. When the data belongs to R m j When I = 1, otherwise I = 0; (3) Calculate the final reinforcement learning machine 4. The method for state perception and fault diagnosis of combined electrical switchgear based on multiple characteristic parameters according to claim 1 is characterized in that: According to the true category of the sample and the prediction result category of the fault diagnosis model, it is divided into four cases: true positive example TP, false positive example FP, true negative example TN and false negative example FN. The F1 score indicator F_score is the harmonic mean of precision and recall, and the calculation formula is as follows: Among them, P is the precision rate and R is the recall rate; Precision is the ratio of the number of correctly classified results in a certain category to all classification results of that category. The calculation formula is as follows: The recall rate is the ratio of the number of correctly classified samples in a certain actual category to all classification results of that category. The calculation formula is as follows:

5. The method for state perception and fault diagnosis of combined electrical switchgear based on multiple characteristic parameters according to claim 1 is characterized in that: The sparrow search algorithm optimizes parameters in the following way: The finder's location is updated as follows: Where t represents the current number of iterations; T is the maximum number of iterations; α is a uniform random number between (0, 1); Q is a random number that follows a standard normal distribution; L is a 1×d matrix with all elements being 1; R2∈[0,1] represents the warning value; ST∈[0.5,1] represents the safety value; The joiner position is updated as follows: In the formula, represents the global worst position at the tth iteration; represents the best position of the discoverer at the t+1th iteration; N represents the population size; A represents a 1×d matrix with each element randomly assigned to 1 or -1, and satisfies A + =A T (AA T ) -1 ; The Scout positions have been updated as follows: In the formula, represents the global optimal position at the tth iteration; β represents the step size control parameter, which is a normally distributed random number with a mean of 0 and a variance of 1; K is a random number between [-1, 1]; f i represents the current sparrow's fitness; f g represents the optimal fitness of the current sparrow population; f w Represents the worst fitness of the current sparrow population.

Images