A disconnector defect analysis method based on a Bayesian classifier

Abstract

The application discloses a kind of based on the defect analysis method of isolating switch of bayesian classifier, comprising the following steps: S1, simulating setting several groups of defect categories of isolating switch and defect degree, obtain corresponding motor current under each group of defect category and defect degree state;S2, motor current data is sampled according to time interval, constitutes discrete time sequence, and carries out normalization processing, calculates corresponding current characteristic quantity;S3, each group of defect category, defect degree and current characteristic quantity are input into bayesian network and are trained, obtain bayesian classification model;S4, measure the motor current in the actual work of isolating switch, calculate corresponding current characteristic quantity, input the bayesian classification model, obtain corresponding defect type and defect degree;The application adopts bayesian classifier, and the defect type of isolating switch is intelligently identified, and the quantitative diagnosis of isolating switch defect state is realized.

Description

Technical Field

[0001] This invention relates to the field of disconnector defect detection technology, specifically to a disconnector defect analysis method based on a Bayesian classifier. Background Technology

[0002] The contact status of the moving and stationary contacts of a GIS disconnector switch in the closed state is an important concern for the power system. Poor contact status can lead to overheating of the contacts and eventual meltdown of the equipment. Therefore, monitoring the contact status of GIS disconnector switches is a key focus for the power grid.

[0003] The monitoring of the contact status of GIS disconnect switches is mainly divided into two categories:

[0004] The first category is direct state detection, which mainly includes: (1) Pressure sensor detection. The pressure sensor detection method involves installing pressure sensors at positions such as the contact springs of the disconnector switch, and determining the open / closed state of the disconnector switch by measuring the pressure between the contact fingers. This method is mainly applied to open disconnector switches, and there are no GIS application cases or related patents yet; (2) Attitude sensor. The attitude sensor detection method involves installing attitude sensors at the same position on each phase of the disconnector switch, and determining the open / closed state of the disconnector switch by measuring the rotation angle of the ground potential transmission component; (3) Micro-motion limit switch online monitoring system. This is a method of monitoring the open / closed position of the disconnector switch by installing micro-motion limit switches on a mechanism and determining the position of the moving contact of the disconnector switch by the movement position of the mechanism.

[0005] The drawback of the direct detection method is that, due to the enclosed structure, limited internal space, compact transmission mechanism, and operation in a high-voltage, high-current environment of the GIS disconnect switch, direct detection is unsuitable for detecting this object. Introducing a sensor at a high potential makes it difficult to lead out the signal line, the signal is susceptible to strong electromagnetic interference, it can damage the original structural seal, and it can easily lead to insulation problems.

[0006] The second category is state perception methods, which mainly include: (1) Motor current detection, which monitors the state of the disconnect switch by measuring the current of the motor driven by the GIS disconnect switch. Specifically, the state of the GIS disconnect switch contact is calculated based on the following principle: after the moving and stationary contacts touch during the movement, the current will change. This point is taken as the starting point of time, and the current cut-off point is taken as the ending point of time. The running time of the moving contact after the moving and stationary contacts touch is measured. Assuming that the motor speed remains unchanged, the distance of the moving contact is measured by time, and the spring force is reflected by the trend of the change in the amplitude of the current; (2) Vibration signal detection, which is a method of diagnosing GIS faults by vibration detection. It mainly involves using GIS equipment in actual applications in the laboratory to simulate the operating state of various mechanical defects and using appropriate sensors to detect the vibration signal of the GIS equipment shell.

[0007] For motor current detection, existing methods cannot reflect the phase change relationship between voltage and current, and the current of different motor models has a certain degree of dispersion, resulting in unclear current characteristics. Even when the current characteristics are relatively clear, existing diagnostic algorithms are still physical analysis methods. They perform defect analysis by analyzing the current envelope, which is a threshold judgment. This judgment causes two problems: First, it is difficult to directly quantify the threshold setting for different disconnect switches. Expert intervention is required to perform physical analysis of the disconnect switch's movement process before applying the threshold, which is costly. Second, it is also difficult to adapt to the differences in threshold caused by the discreteness of the mechanical state of the switches in industrial settings. Summary of the Invention

[0008] The purpose of this invention is to address the problems existing in the prior art by providing a method for analyzing disconnector defects based on a Bayesian classifier. By employing a Bayesian classifier, the defect types of disconnectors are intelligently identified, and a quantitative diagnosis of the defect status of the disconnectors is achieved.

[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0010] A method for analyzing the defects of disconnect switches based on a Bayesian classifier includes the following steps:

[0011] S1. Simulate and set several groups of defect categories and defect severity of the disconnecting switch, and obtain the corresponding motor current under each group of defect categories and defect severity.

[0012] S2. Sample the motor current data according to time intervals to form a discrete time series, and perform normalization processing to calculate the corresponding current characteristic quantities.

[0013] S3. Input the defect categories, defect severity and current characteristics of each group into the Bayesian network for training to obtain the Bayesian classification model.

[0014] S4. Measure the motor current during the actual operation of the disconnecting switch, calculate the corresponding current characteristic quantity, input it into the Bayesian classification model, and obtain the corresponding defect type and defect degree.

[0015] Step S2 specifically includes, assuming the discrete time series is...

[0016] I[P n ] = {[P1], [P2], [P3]...[P n ]}

[0017] Calculate I[P] n The average value P mean Then let I[P nSubtract P from each element in ] mean , thus obtaining the normalized matrix I[P]

[0018] I[P]={[P1-P mean ], [P2-P mean ], [P3-P mean ...[P n -P mean ]}

[0019] The current characteristic quantity includes the mean value. Mean square value Root amplitude Standard deviation variance Rotation angle θ r =max(θ(n))-min(θ(n)); Peak value C p =max(|P(n)|); peak-to-valley value C p =max(|P(n)|); Peak factor C pi =C p / C rms Waveform factor Skewness Kudo

[0020] Step S3 specifically includes collecting data and establishing a sample set, where each sample includes the defect category, defect severity, and the corresponding current characteristic value.

[0021] The defect category and defect severity are set as the root nodes of the Bayesian network, and the current feature is set as the leaf nodes of the Bayesian network. The sample set is input into the Bayesian network, and the conditional probability of the root node is modeled using a multinomial distribution and the conditional probability of the leaf node is modeled using a Gaussian distribution.

[0022] The defect categories include jamming, three-phase asynchrony, incomplete opening and closing, and spring fatigue; the defect severity includes 10%, 20%, ..., 100%.

[0023] Compared with the prior art, the beneficial effects of the present invention are:

[0024] 1. A Bayesian classifier is used to perform pattern recognition by finding the minimum evaluation function of the Bayesian network, thereby realizing intelligent diagnosis of the mechanical state of the disconnect switch;

[0025] 2. When training the Bayesian pattern recognition model, quantization input is performed, and the model training results can realize the quantitative diagnosis of the disconnect switch status. Attached Figure Description

[0026] Figure 1This is an overall flowchart of the present invention;

[0027] Figure 2 This is a diagram showing the correspondence between the current characteristic quantities and their definitions in this invention;

[0028] Figure 3 This is a three-dimensional matrix diagram illustrating the defect categories, defect severity, and current characteristic quantities of the present invention. Detailed Implementation

[0029] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] In the description of this invention, it should be noted that the terms "upper", "lower", "left", "right", "top", "bottom", "inner", "outer", "horizontal", "vertical", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limiting this invention.

[0031] like Figure 1 As shown, a method for analyzing the defects of disconnect switches based on a Bayesian classifier includes the following steps:

[0032] S1. Perform defect simulation on the disconnector switch test platform. Simulate several groups of defect categories and defect severity for the disconnector switch. Perform opening and closing operations on the disconnector switch under defect conditions to obtain the motor current corresponding to each group of defect categories and defect severity conditions, forming the table shown in Table 1:

[0033]

[0034] Table 1

[0035] S2. Sample the motor current data according to time intervals to form a discrete time series, and perform normalization processing to calculate the corresponding current characteristic quantities.

[0036] S3. Input the defect categories, defect severity and current characteristics of each group into the Bayesian network for training to obtain the Bayesian classification model.

[0037] S4. Measure the motor current during the actual operation of the disconnecting switch, calculate the corresponding current characteristic quantity, input it into the Bayesian classification model, and obtain the corresponding defect type and defect degree.

[0038] Specifically, the defect categories include jamming, three-phase asynchrony, incomplete opening and closing of the circuit breaker, and spring fatigue; the defect severity includes 10%, 20%, ..., 100%.

[0039] Step S2 specifically includes, assuming the discrete time series is...

[0040] I[P n ] = {[P1], [P2], [P3]...[P n ]}

[0041] Calculate I[P] n The average value P mean Then let I[P n Subtract P from each element in ] mean , thus obtaining the normalized matrix I[P]

[0042] I[P]={[P1-P mean ], [P2-P mean ], [P3-P mean ...[P n -P mean ]}

[0043] By adopting a current data normalization method, the problem of threshold evaluation that cannot be uniformly evaluated due to the discreteness of the initial state of switching machinery in industrial practice is avoided.

[0044] Calculate the characteristic data of I[P] as follows Figure 2 As shown, current characteristic quantities include the mean. Mean square value Root amplitude Standard deviation variance Rotation angle θ r =max(θ(n))-min(θ(n)); Peak value C p =max(|P(n)|); peak-to-valley value C p =max(|P(n)|); Peak factor C pi =C p / C rms Waveform factor Skewness Kudo

[0045] Step S3 specifically includes, to Figure 2 The various current characteristic quantities in the table are combined with the data in Table 1 to form the following: Figure 3 The three-dimensional matrix shown is used to establish a sample set, where each sample includes the defect category, defect severity, and the corresponding current characteristic value.

[0046] The defect category and defect severity are set as the root nodes of the Bayesian network, and the current feature is set as the leaf nodes of the Bayesian network. The sample set is input into the Bayesian network, and the conditional probability of the root node is modeled using a multinomial distribution and the conditional probability of the leaf node is modeled using a Gaussian distribution.

[0047] The principle of the Bayesian classifier is as follows:

[0048]

[0049] P(c|x) is the class-conditional probability of sample x with respect to classification label c. P(x) is the normalization factor, and P(c) is the prior probability of the class.

[0050] The basic form is

[0051] The above equation, following the conditional independence assumption adopted in the literature, can be rewritten as follows:

[0052]

[0053] The above formula is the Naive Bayes classifier:

[0054] The above formula can be considered as the following training process: using a sample set D to estimate the class prior probability P(c), and estimating the conditional probability P(x) for each attribute. i |c).

[0055] A Bayesian network is a dependency network where nodes represent entities such as molecules or genes. Interacting nodes are connected by edges that influence the direction of the interaction. Generally, a Bayesian network is a directed acyclic graph, disallowing cycles. Importantly, each node is assigned a probability that defines the chance of finding the node in a given state. If the state of one node depends on the state of another node, conditional probabilities are used. These dependencies propagate through the network and influence the probabilities of other nodes, which are updated as new information about the nodes becomes available. Therefore, Bayesian networks are also known as probabilistic causal models.

[0056] A key characteristic of the successful adoption of Bayesian networks as a decision modeling tool compared to other modeling methods is their relative simplicity. Characterizing the strength of relationships between variables using probability strength means that both quantitative and qualitative information can be used to update the model as new knowledge or data becomes available. Bayesian networks can easily incorporate uncertainty, which is reflected in the model output. Sensitivity analysis tools allow for the characterization of uncertainty to identify key causal factors and knowledge gaps. The model results are testable.

[0057] However, when the data source is relatively limited, the conclusions may have some data "bias". But for the fault diagnosis of disconnecting switches, we need to focus more on the fault type rather than fault quantification. Therefore, the advantages of Bayesian networks are well applicable in this application.

[0058] Specifically, in this application, it is assumed that two current characteristic quantities X1 and X2 can be measured (in this invention, there are not only two characteristic quantities X1 and X2, but also as follows). Figure 1 The 12 features shown are illustrated here, but for simplicity, only two state variables are used as examples. These two features affect the system's operating state. Simultaneously, the system's defect type F and defect severity G are recorded. Defect type F includes two defect types, A and B, and each defect type is further divided into two different defect severity levels, G1 and G2. A Bayesian network is constructed to predict the system's fault type and severity. The specific steps are as follows:

[0059] (1) Variable definition:

[0060] Define the variables to be modeled, including observed variables and latent variables. In this example, the observed variables are X1, X2, F, and G, and the latent variables are the relationships between the nodes. We use the following notation to represent the variables:

[0061] X1: State variable 1

[0062] X2: State variable 2

[0063] F: Fault Type

[0064] G: Fault severity

[0065] (2) Model construction:

[0066] We use a Bayesian network to represent this model. A Bayesian network is a directed acyclic graph where nodes represent variables and edges represent dependencies between variables. In this example, we can construct the following Bayesian network:

[0067] Two random variables: F represents the fault type, and G represents the fault severity. There are also two state variables: X1 and X2, which reflect the state of the mechanical system. The Bayesian network model is as follows:

[0068] F and G are independent of each other; they are the root nodes of the model.

[0069] X1 and X2 are leaf nodes of the model, and they directly depend on nodes F and G.

[0070] The conditional probability table in Table 2 shows how X1 and X2 depend on F and G.

[0071]

[0072]

[0073] Table 2

[0074] In this Bayesian network, the parent nodes of F and G are latent variables, and there are no direct edges between them. The parent nodes of X1 and X2 are F and G, and there are direct edges between them. This structure indicates that the values ​​of X1 and X2 are affected by the fault type and severity, which are independent of each other.

[0075] (3) Data collection:

[0076] To train this Bayesian network, some data needs to be collected. In this example, some data is randomly generated to represent the operating state and fault conditions of the mechanical system. For example, 1000 samples could be generated, each including the values ​​of X1, X2, F, and G. The data can be generated using Python or other programming languages.

[0077] (4) Learning model parameters:

[0078] With the data, we can use maximum likelihood estimation or maximum a posteriori probability estimation to learn the model parameters. In this example, we use maximum likelihood estimation to learn the conditional probability table. Specifically, for each node, we need to compute its conditional probability table, for example:

[0079] P(F): Represents the prior probability of F.

[0080] P(G): Represents the prior probability of G.

[0081] P(X1|F,G): Represents the conditional probability of X1 given F and G.

[0082] P(X2|F,G): Represents the conditional probability of X2 given F and G.

[0083] In this example, we need to compute a conditional probability table for each node. For X1 and X2, we can use a Gaussian distribution to model their conditional probabilities. For F and G, we can use a multinomial distribution to model their conditional probabilities.

[0084] Suppose our dataset contains data as shown in Table 3:

[0085]

[0086]

[0087] Table 3

[0088] Based on this data, a conditional probability table can be calculated:

[0089] P(F=A)=0.5, P(F=B)=0.5

[0090] P(G=G1)=0.5, P(G=G2)=0.5

[0091] For X1 and X2, their conditional probabilities can be modeled using a Gaussian distribution. For example, P(X1|F=A,G=G1) can be obtained by calculating the mean and variance of X1 given F=A and G=G1. Assume the calculated results are:

[0092] P(X1|F=A,G=G1)=N(1.5,0.5^2)

[0093] This means that, given F = A and G = G1, the conditional probability of X1 follows a Gaussian distribution with a mean of 1.5 and a variance of 0.5.

[0094] (5) Inference:

[0095] With a model and parameters, Bayesian inference can be used to predict the values ​​of unknown variables. Suppose we have measured X1 = 2 and X2 = 3, and we want to predict the type and severity of system failures. We can use Bayes' theorem to calculate the posterior probability:

[0096] P(F,G|X1=2,X2=3) is proportional to P(F)P(G)P(X1=2|F,G)P(X2=3|F,G)

[0097] P(X1=2|F,G) and P(X2=3|F,G) can be calculated using conditional probability tables. Assume the calculated results are:

[0098] P(X1=2|F=A,G=G1)=0.6, P(X2=3|F=A,G=G1)=0.7

[0099] P(X1=2|F=A,G=G2)=0.4, P(X2=3|F=A,G=G2)=0.5

[0100] P(X1=2|F=B,G=G1)=0.8, P(X2=3|F=B,G=G1)=0.9

[0101] P(X1=2|F=B,G=G2)=0.6, P(X2=3|F=B,G=G2)=0.8

[0102] Assuming we have no prior knowledge of P(F) and P(G), we can assume they follow a uniform distribution. Then we can obtain:

[0103] P(F=A)=P(F=B)=0.5

[0104] P(G=G1)=P(G=G2)=0.5

[0105] Substituting the above parameters into Bayes' theorem, we get:

[0106] P(F=A, G=G1|X1=2, X2=3)∝0.5·0.5·0.6·0.7=0.105

[0107] P(F=A, G=G2|X1=2, X2=3)∝0.5·0.5·0.4·0.5=0.05

[0108] P(F=B, G=G1|X1=2, X2=3)∝0.5·0.5·0.8·0.9=0.18

[0109] P(F=B, G=G2|X1=2, X2=3)∝0.5·0.5·0.6·0.8=0.12

[0110] These probabilities can then be normalized to obtain the posterior probability distribution:

[0111] P(F=A, G=G1|X1=2, X2=3)=0.244

[0112] P(F=A, G=G2|X1=2, X2=3)=0.117

[0113] P(F=B, G=G1|X1=2, X2=3)=0.421

[0114] P(F=B, G=G2|X1=2, X2=3)=0.218

[0115] This means that, given X1 = 2 and X2 = 3, the possible fault types and degrees of the system are B / G1 and A / G2, respectively. These results can be used to predict the types and degrees of future faults that may occur in the system.

[0116] In summary, this invention employs a Bayesian classifier and performs pattern recognition by finding the minimum evaluation function of the Bayesian network, thereby achieving intelligent diagnosis of the mechanical state of disconnect switches. Quantized input is used during the training of the Bayesian classification model, and the model training results can achieve quantitative diagnosis of the disconnect switch state.

[0117] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for analyzing defects in disconnect switches based on a Bayesian classifier, characterized in that, Includes the following steps: S1. Simulate and set several groups of defect categories and defect severity of the disconnecting switch, and obtain the corresponding motor current under each group of defect categories and defect severity. S2. Sample the motor current data at time intervals to form a discrete-time series, and perform normalization processing to calculate the corresponding current characteristic quantities; the current characteristic quantities include the mean value. Mean square value Root amplitude Standard deviation variance Rotation angle Peak value ; Peak and valley values Peak factor C pi =C p / C rms Waveform factor Skewness Kudo ; S3. Input the defect categories, defect severity and current characteristics of each group into the Bayesian network for training to obtain the Bayesian classification model. The defect category and defect severity are set as the root nodes of the Bayesian network, the current feature is set as the leaf nodes of the Bayesian network, the sample set is input into the Bayesian network, the multinomial distribution is used to model the conditional probability of the root nodes, and the Gaussian distribution is used to model the conditional probability of the leaf nodes. S4. Measure the motor current during the actual operation of the disconnecting switch, calculate the corresponding current characteristic quantity, input it into the Bayesian classification model, and obtain the corresponding defect type and defect degree.

2. The method for analyzing disconnector defects based on a Bayesian classifier according to claim 1, characterized in that, Step S2 specifically includes, assuming the discrete time series is: ； Calculate I[P] n The average value P mean Then let I[P n Subtract P from each element in ] mean Obtain the normalized matrix I[P]: 。 3. The method for analyzing disconnector defects based on a Bayesian classifier according to claim 1, characterized in that, Step S3 specifically includes collecting data and establishing a sample set, where each sample includes the defect category, defect severity, and the corresponding current characteristic value.

4. The method for analyzing the defects of disconnect switches based on a Bayesian classifier according to claim 1, characterized in that, The defect categories include jamming, three-phase asynchrony, incomplete opening and closing, and spring fatigue; the defect severity includes 10%, 20%, ..., 100%.

Images