Power equipment fault diagnosis method and power acquisition terminal

Abstract

The invention belongs to the technical field of fault identification, and particularly relates to a power equipment fault diagnosis method and a power acquisition terminal. The method comprises the following steps: acquiring signals of different sensors when the power equipment operates; the method comprises the following steps: preprocessing each sensor signal, and extracting a feature vector of each signal; inputting the extracted feature vector of each signal into a fault diagnosis model to obtain a fault diagnosis result; the fault diagnosis model is obtained by training a machine learning model based on each signal feature vector data set extracted by the power equipment in different operation states; before data mapping, feature vectors of all signals are subjected to similarity matching, matching positions require first-bit correspondence, middle monotonically increasing and non-simultaneous repetition, dynamic matching is carried out on time sequence features through feature extension or compression, and the problems of signal asynchronism and low signal acquisition hardware precision caused by unstable control voltage and low signal acquisition hardware precision are solved. Or the diagnosis accuracy is low when data missing is caused by sensor faults.

Description

Technical Field

[0001] The present invention belongs to the technical field of fault identification, and in particular relates to a method for diagnosing faults of electric power equipment and a power collection terminal. Background Art

[0002] At present, state monitoring of power equipment is an important means for power companies to ensure power quality and formulate maintenance strategies. Power equipment will produce changes in various physical quantities during state switching, such as current, voltage, vibration, and sound. Therefore, fault diagnosis and analysis of power equipment is usually performed based on multiple signals collected by sensors and combined with machine learning algorithms to monitor the state of power equipment. For example, a patent application document with publication number CN117347913A discloses a three-level inverter open circuit fault diagnosis method and system, which uses an extreme learning machine with a Gaussian kernel function to analyze the fault diagnosis results based on the characteristics of the output three-phase phase voltage and terminal voltage. However, for sensors that trigger acquisition when the control voltage reaches a certain value or that acquire data according to the hardware frequency, when the control voltage is unstable or the signal acquisition hardware is not accurate, the sensor signals at different installation locations are prone to asynchronous acquisition during the acquisition process. The Gaussian kernel function calculates the similarity between the Euclidean distances between each two points (i.e., feature vectors) in different signals, and maps the data to a high-dimensional space based on the Euclidean distances of each pair of input points. When signal acquisition is not synchronized, the data mapped to the high-dimensional space will not be synchronized, and the diagnostic results output by the machine learning model for fault diagnosis based on the data mapped to the high-dimensional space will also have errors. In addition, when a sensor failure causes some data to be missing, the signal of the missing part of the data is inconsistent with the data length of other signals. The Gaussian kernel function cannot perform high-dimensional mapping of the missing data, which will also cause errors in the diagnostic results output by the machine learning model, and the diagnostic accuracy is low. Summary of the invention

[0003] The object of the present invention is to provide a method for diagnosing faults of electric power equipment and a power collection terminal, so as to solve the problem of low fault diagnosis accuracy due to signal asynchrony and loss.

[0004] In order to solve the above technical problems, the present invention provides a method for diagnosing faults of electric equipment, comprising the following steps:

[0005] Obtain signals from different sensors when power equipment is running;

[0006] After preprocessing each sensor signal, extract the feature vector of each signal;

[0007] The extracted feature vectors of each signal are input into the fault diagnosis model to obtain the fault diagnosis result; the fault diagnosis model is obtained by training the machine learning model based on the feature vector data set of each signal extracted from the power equipment under different operating conditions; the machine learning model performs data mapping on the feature vector by introducing a calibration kernel function, and before data mapping, the feature vectors of each signal are matched for similarity, wherein the matching position requires the first position to correspond, the middle to increase monotonically and not to repeat at the same time.

[0008] Furthermore, the expression of the calibration kernel function g is:

[0009]

[0010] Among them, A is the similarity matching combination of different signal feature vectors, B is all possible matching combinations of different signal feature vectors, n and m are the number of different signal feature vectors, l is the similarity matching length, A 1 (i) A 2 (i) are the i-th matching positions of different signal feature vectors, are the feature vectors after re-matching of different signals, and T and γ are the kernel parameters.

[0011] Furthermore, the fault diagnosis model is a two-level classification model, the first level includes N base learners, and the second level includes a meta learner. The input feature vector is processed by the two-level classification model and the fault diagnosis result is output; wherein the base learner and the meta learner are machine learning models, the calibration kernel function g is introduced into the base learner, the base learner is used to establish the relationship between the feature vector and the prediction result of the base learner, and the meta learner is used to establish the relationship between the prediction result of the base learner and the final fault diagnosis result; N is an integer greater than 1.

[0012] Furthermore, the method for training the two-level classification model includes: dividing each signal feature vector data set extracted from the power equipment under different operating conditions into a training set and a test set, dividing the training set into equal subsets using a cross-cycle validation method, taking one portion as validation data and the rest as training data, using the training data to train each base learner in turn, inputting the validation data into the trained base learner for cross-validation, recording the output prediction results as the first data set, and calculating the proportion of the prediction accuracy of each base learner to the total accuracy, and allocating the weight of each base learner according to the proportion; inputting the test set into the trained base learner, multiplying the prediction results output by each base learner by the weight of each base learner, and recording the calculated results as the second data set, using the first data set as new training data to train the meta-learner to establish a fault diagnosis model, and the second data set is used to test the performance of the fault diagnosis model.

[0013] Furthermore, the N is 2, the base learners are support vector machines and extreme learning machines, and the meta learner is a multivariate linear regression model.

[0014] Furthermore, the preprocessing includes denoising and normalization processing.

[0015] Furthermore, the denoising method is to decompose and reconstruct each signal using a variational mode decomposition method, and then sort the component signals obtained after the decomposition and reconstruction according to their importance, remove the component signal with the lowest importance, and then merge the remaining component signals.

[0016] Furthermore, the signals of each sensor acquired when the electric power equipment is running are time-synchronized signals.

[0017] The beneficial effects of the above technical solution are as follows: the present invention is an improved invention, which targets the power equipment fault diagnosis scenario under multiple signals. First, the feature vectors under different signals are similarly matched before the data mapping of the machine learning model. The matching position requires the first position to correspond, the middle to increase monotonically and not to repeat at the same time. The method allows repeated matching and dynamically matches the timing features through feature extension or compression, thereby dealing with the problem that the diagnostic results output by the machine learning model have errors and low diagnostic accuracy when the control voltage is unstable, the signal acquisition hardware is not accurate enough, resulting in signal asynchrony, or the data is missing due to sensor failure.

[0018] In order to solve the above technical problems, the present invention further provides a power collection terminal, including a processor, and the processor is used to execute computer program instructions to implement the above-mentioned power equipment fault diagnosis method.

[0019] Furthermore, it also includes a wireless sensor module, which is used to obtain signals from different sensors when the power equipment is running after time synchronization.

[0020] The beneficial effects of the above technical solution are as follows: the present invention is an improved invention. For the power equipment fault diagnosis scenario under multiple signals, the power acquisition terminal first performs similarity matching on the feature vectors under different signals before data mapping of the machine learning model. The matching position requires the first position to correspond, the middle to increase monotonically and not to repeat at the same time. This method allows repeated matching and dynamically matches the timing features through feature extension or compression, thereby dealing with the problem of errors in the diagnosis results output by the machine learning model and low diagnostic accuracy when the control voltage is unstable, the signal acquisition hardware is not accurate enough, resulting in signal asynchrony, or data is missing due to sensor failure. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 2 is a schematic diagram of the structure of a power collection terminal according to an embodiment of the power collection terminal of the present invention;

[0022] Figure 2 is a flow chart of fault diagnosis of electric power equipment in an embodiment of the electric power collection terminal of the present invention;

[0023] Figure 3 is a flow chart of establishing a fault diagnosis model of an embodiment of a power collection terminal of the present invention;

[0024] Figure 4 is a transformer vibration signal diagram of an embodiment of the method of the present invention;

[0025] Figure 5 is a comparison diagram of diagnostic effects of different fault diagnosis methods obtained based on transformer vibration signals in a method embodiment;

[0026] Figure 6 is a circuit breaker vibration signal diagram of a method embodiment;

[0027] Figure 7 3 is a comparison diagram of diagnostic effects of different fault diagnosis methods obtained based on circuit breaker vibration signals in a method embodiment. DETAILED DESCRIPTION

[0028] In order to make the purpose, technical solutions and advantages of the present invention more clear, the specific implementation methods of the present invention are further described below in conjunction with the accompanying drawings.

[0029] The present invention aims at the power equipment fault diagnosis scenario under multi-source signals. First, the feature vectors under different signals are similarly matched before the data mapping of the machine learning model. The matching position requires the first position to correspond, the middle to increase monotonically and not to repeat at the same time. The method allows repeated matching and dynamically matches the timing features through feature extension or compression, thereby dealing with the problems of errors and low diagnostic accuracy in the diagnosis results output by the machine learning model when the control voltage is unstable, the signal acquisition hardware is not accurate enough, or the data is missing due to sensor failure. This greatly improves the accuracy of power equipment fault diagnosis and improves the accuracy and stability of power equipment status detection.

[0030] Power collection terminal implementation example

[0031] A power collection terminal of the present invention includes a processor, the processor is used to execute stored computer program instructions to implement a method for diagnosing power equipment faults. The power collection terminal also includes a wireless sensor module, the wireless sensor module is used to perform time calibration according to a trigger instruction issued by the processor to collect signals of multiple sensors when the power equipment is running after time synchronization, thereby reducing the deviation of signal asynchrony when collecting multiple signals.

[0032] Specifically, Figure 1As shown, the power collection terminal includes a control module, a communication module, a wireless sensor module, a time synchronization module, a power module, a heat dissipation module, a data storage module, a display module and a signal processing module (i.e., a processor, not shown in the figure). The control module adopts an FPGA architecture, which is used to send control instructions (including trigger instructions and collection instructions); the communication module is used for data interaction between the control module, the wireless sensor module and the signal processing module; the wireless sensor module includes a current sensor, a voltage sensor, a vibration sensor and a temperature sensor, etc.; the time synchronization module is used for time synchronization during signal collection, and the time synchronization module can be a Lora gateway based on Beidou timing, or other modules that can obtain time; the power module is used to provide power for the power collection terminal; the heat dissipation module is used to dissipate heat for the power collection terminal; the data storage module is used to store the data collected by the wireless sensor module and store fault diagnosis results; the display module is used to display the data collected by the wireless sensor module and the fault diagnosis results; the signal processing module is used to analyze and diagnose faults of the collected multi-source signals. The method for diagnosing faults of power equipment is as follows: Figure 2 As shown, the following steps are included:

[0033] 1. Obtain signals from different sensors when power equipment is running.

[0034] The control module sends a trigger instruction to the wireless sensor module. After receiving the trigger instruction, the wireless sensor module sends a message to the time synchronization module to obtain the time in the time synchronization module, and calibrates the time of each sensor module according to the time. After each sensor module completes the time calibration, the control module sends a collection instruction to the wireless sensor module, and each sensor module in the wireless sensor module starts to collect signals during the operation of the power equipment.

[0035] 2. Preprocess the signals of each sensor.

[0036] The preprocessing in the present invention includes denoising and normalization. As an implementation mode, the present invention uses the variational mode decomposition (VMD) method to decompose and reconstruct each signal, and then sorts the component signals obtained after decomposition and reconstruction according to the importance, removes the component signal with the lowest importance, and merges the remaining component signals. Each signal is a time series signal containing time. Specifically, after each signal is decomposed and reconstructed, the information entropy of each component signal is calculated, and the importance of each component signal is sorted using random forest, the component with the lowest importance is removed, and the remaining components are merged.

[0037] 3. Extract the feature vector of each signal.

[0038] After the signal is preprocessed, the preprocessed signals are framed using a sliding window, and feature vectors that can reflect the timing characteristics of the signals, such as the short-time energy and zero-crossing rate of each signal, are calculated.

[0039] 4. Input the extracted characteristic vectors of each signal into the fault diagnosis model to obtain the fault diagnosis results.

[0040] Among them, the method of establishing the fault diagnosis model is as follows Figure 3 As shown, the following steps are included:

[0041] 1) Acquire the sensor signals of the power equipment under different operating states (including normal state and fault state), perform preprocessing according to steps 2 and 3, and extract the feature vectors of each signal under different operating states to form a feature vector data set;

[0042] 2) Calculate the calibration kernel function of the signal based on the feature data set.

[0043] The calibration kernel essentially maps the feature vector to a high-dimensional space to compare the similarity of samples. Signal asynchrony or data missing problems can cause a large deviation between the kernel function calculation result and the actual result. To this end, the present invention matches the feature vectors of each signal for similarity before data mapping, where the matching position requires the first position to correspond, the middle position to increase monotonically, and the position to not repeat at the same time. Take the feature vectors x(x) extracted from the two samples as 1 ,x 2 ,…,x n ) and y(y 1 ,y 2 ,…,y m ) as an example to illustrate the calibration kernel calculation process;

[0044] First, the two feature vectors are matched for similarity. The matching positions are required to correspond to the first position, increase monotonically in the middle, and not repeat at the same time, that is:

[0045]

[0046] Among them, A is the similarity matching combination of different signal feature vectors, n and m are the number of different signal feature vectors, l is the similarity matching length, and A 1 (i) A 2 (i) are the i-th matching positions of different signal feature vectors. Different signal feature vectors refer to x and y. The present invention matches according to the matching position requirements of first correspondence, monotonous growth in the middle and no simultaneous repetition, but here it only says no simultaneous repetition, and there is no absolute prohibition of repetition (for example, an element in one of the feature vectors is repeated multiple times), so there are multiple similarity matching possibilities, which are finally combined into A.

[0047] Then, the calibration kernel function g is used to comprehensively calculate the similarity matching combination A of the obtained different signal feature vectors. The calibration kernel function is the feature calibration kernel function of the signal, which is used to introduce it into the machine learning model to map the feature vector to high-dimensional space data. The expression of the calibration kernel function g is:

[0048]

[0049] Among them, B is all possible matching combinations of different signal feature vectors, which is used to control the matching range. are the feature vectors after re-matching different signals, T and γ are kernel parameters, which are used to control the matching range and function smoothness. After the model is trained, the kernel parameters of the kernel function remain unchanged.

[0050] 3) The calibration kernel function is introduced into the machine learning model, and the machine learning model is trained based on the signal feature vector data sets extracted from the power equipment under different operating conditions to obtain a fault diagnosis model.

[0051] As a preferred implementation, in order to further improve the diagnostic accuracy, the fault diagnosis model is set as a two-level classification model, the first level includes N base learners, the second level includes a meta learner, the input feature vector is processed by the two-level classification model, and the fault diagnosis result is output; wherein the base learner and the meta learner are machine learning models, the calibration kernel function g is introduced into the base learner, the base learner is used to establish the relationship between the feature vector and the prediction result of the base learner, and the meta learner is used to establish the relationship between the prediction result of the base learner and the final fault diagnosis result; N is an integer greater than 1. The base learner can be a machine learning model such as a support vector machine, an extreme learning machine, or KNN, and the meta learner can be a multivariate linear regression model.

[0052] The method for training a two-level classification model includes: dividing each signal feature vector data set extracted from the power equipment under different operating conditions into a training set and a test set, dividing the training set into equal subsets using a cross-circular validation method, taking one set as validation data and the rest as training data, using the training data to train each base learner in turn, inputting the validation data into the trained base learner for cross-validation, and recording the output prediction result as a first data set; inputting the test set into the trained base learner, recording the output result as a second data set, using the combination of the first data set after cross-validation as new training data to train a meta learner to establish a fault diagnosis model, and using the second data set to test the performance of the fault diagnosis model.

[0053] Furthermore, according to the diagnostic accuracy of each base learner, weights are assigned to different base learners. At this time, the method for training the two-level classification model includes: dividing each signal feature vector data set extracted from the power equipment under different operating conditions into a training set and a test set, using the cross-circular validation method to divide the training set into equal subsets, taking one as the validation data and the rest as the training data, using the training data to train each base learner in turn, inputting the validation data into the trained base learner for cross-validation, recording the output prediction results as the first data set, and calculating the proportion of the prediction accuracy of each base learner to the total accuracy, and assigning the weight of each base learner according to the proportion; inputting the test set into the trained base learner, multiplying the prediction results output by each base learner by the weight of each base learner, and recording the calculation results as the second data set, using the first data set as new training data to train the meta learner to establish a fault diagnosis model, and the second data set is used to test the performance of the fault diagnosis model.

[0054] Among them, the weight calculation process of each base learner is:

[0055] First, calculate the prediction accuracy of each base learner:

[0056]

[0057] Then, to ensure that the weights of each base classifier are on the same scale, the probability is normalized and the weights are as follows:

[0058]

[0059] Among them, n i is the accuracy of the ith base learner, k is the number of folds in the cross-validation, c is the number of categories in the data set, N is the number of samples in category p correctly classified by the i-th base learner in the j-th compromise, p is the total number of samples in category p, w i is the weight of the ith base learner, and s is the number of base learners.

[0060] As a specific implementation, N is 2, the base learners are support vector machines and extreme learning machines, and the meta learner is a multivariate linear regression model. The kernel function is introduced into the extreme learning machine and the support vector machine respectively:

[0061]

[0062] Among them, f(x) is the first-level model of the fault diagnosis model, N is the number of samples, I is the unit matrix, C is the penalty factor, Y is the sample label matrix, y i is the label of the i-th sample, α iis the Lagrange multiplier of the i-th sample, and b is the bias.

[0063] Train the two-level classification model according to the above training method, calculate the weights according to the cross-validation results, and introduce the weights into the trained first-level model:

[0064] f w (x) = [w s w e ] T ⊙f(x)

[0065] Among them, f w (x) is the first-level model after weight assignment, w s and w e They represent the weights of the two classifiers respectively.

[0066] The test set is input into the first-level model with weights, and the calculation result ( Figure 3 The average prediction result in ( ) is recorded as the second data set, the first data set is used to train the multivariate linear regression model, the second data set is used for verification, and the multivariate linear regression model outputs the final prediction result.

[0067] f r (x) = [β 0 β 1 β 2 ]·[1f w (x)] T

[0068] Among them, f r (x) is the fault diagnosis model, β 0 is the constant term of multiple linear regression, β 1 and β 2 is the regression coefficient.

[0069] Method Embodiment

[0070] The specific implementation process of the method for diagnosing faults of electric power equipment of the present invention has been described in detail in the embodiment of the electric power collection terminal, and will not be repeated here.

[0071] The power collection terminal designed by the present invention is used to collect different vibration signals during the operation of the dry-type transformer, such as Figure 4As shown. 60 groups of normal state, winding loose fault, core loose fault and connecting rod loose fault are collected respectively. The signal processing program integrated in the power collection terminal first uses VMD to denoise the signal, and then calculates the short-time zero-crossing rate of each group of signals as the feature vector, and divides 70% of the original data into training data and 30% of the data as test data, and the above division is performed randomly. Finally, the training data is used to train the Gaussian kernel function support vector machine, Gaussian kernel function extreme learning machine, random forest and the method proposed in the present invention to establish a fault diagnosis model, wherein the base learner of the method proposed in the present invention adopts the calibration kernel support vector machine and calibration kernel extreme learning machine designed by the present invention. The grid optimization method is used to determine the optimal parameters of the model, and the test data is input into these trained diagnostic models to test the fault diagnosis performance. The above training and testing were carried out 30 times in total, and the accuracy of each diagnostic model for each category of fault is shown as follows. Figure 5 It can be clearly seen that no matter which type of fault is targeted, the method proposed in the present invention is obviously better than other methods.

[0072] The power collection terminal designed by the present invention is used to collect different vibration signals during the closing process of the circuit breaker, such as Figure 6 As shown. Specifically, 30 groups of normal state, closing spring fatigue failure, and buffer spring fatigue failure are collected respectively. The signal processing program integrated in the power collection terminal first uses VMD to denoise the signal, and then calculates the short-time zero-crossing rate of each group of signals as the feature vector, and divides 70% of the original data into training data and 30% of the data as test data, and the above division is performed randomly. Finally, the training data is used to train the Gaussian kernel function support vector machine, Gaussian kernel function extreme learning machine, random forest and the method proposed in the present invention to establish a fault diagnosis model, wherein the base learner of the method proposed in the present invention adopts the calibration kernel support vector machine and calibration kernel extreme learning machine designed by the present invention. The grid optimization method is used to determine the optimal parameters of the model, and the test data is input into these trained diagnostic models to test the fault diagnosis performance. The above training and testing were carried out 30 times in total, and the accuracy of each diagnostic model for each category of fault is shown as follows. Figure 7 It can be clearly seen that no matter which type of fault is targeted, the method proposed in the present invention is obviously better than other methods.

Claims

1. A method for diagnosing faults in electric power equipment, characterized in that: The steps include: Obtain signals from different sensors when power equipment is running; After preprocessing each sensor signal, extract the feature vector of each signal; The extracted feature vectors of each signal are input into the fault diagnosis model to obtain the fault diagnosis result; the fault diagnosis model is obtained by training the machine learning model based on the feature vector data set of each signal extracted from the power equipment under different operating conditions; the machine learning model performs data mapping on the feature vector by introducing a calibration kernel function, and before data mapping, the feature vectors of each signal are matched for similarity, wherein the matching position requires the first position to correspond, the middle to increase monotonically and not to repeat at the same time.

2. The method for diagnosing faults of electric power equipment according to claim 1, characterized in that: The expression of the calibration kernel function g is: Among them, A is the similarity matching combination of different signal feature vectors, B is all possible matching combinations of different signal feature vectors, n and m are the number of different signal feature vectors, l is the similarity matching length, A1(i) and A2(i) are the i-th matching positions of different signal feature vectors, are the feature vectors after re-matching of different signals, and T and γ are the kernel parameters.

3. The method for diagnosing faults of electric power equipment according to claim 1, characterized in that: The fault diagnosis model is a two-level classification model, the first level includes N base learners, the second level includes a meta learner, and the input feature vector is processed by the two-level classification model to output the fault diagnosis result; wherein the base learner and the meta learner are machine learning models, the calibration kernel function g is introduced into the base learner, the base learner is used to establish the relationship between the feature vector and the prediction result of the base learner, and the meta learner is used to establish the relationship between the prediction result of the base learner and the final fault diagnosis result; N is an integer greater than 1.

4. The method for diagnosing faults of electric power equipment according to claim 3, characterized in that: The method for training a two-level classification model includes: dividing each signal feature vector data set extracted from the power equipment under different operating conditions into a training set and a test set, dividing the training set into equal subsets using a cross-circular validation method, taking one set as validation data and the rest as training data, using the training data to train each base learner in turn, inputting the validation data into the trained base learner for cross-validation, recording the output prediction results as the first data set, and calculating the proportion of the prediction accuracy of each base learner to the total accuracy, and allocating the weight of each base learner according to the proportion; inputting the test set into the trained base learner, multiplying the prediction results output by each base learner by the weight of each base learner, and recording the calculation results as the second data set, using the first data set as new training data to train the meta learner to establish a fault diagnosis model, and the second data set is used to test the performance of the fault diagnosis model.

5. The method for diagnosing faults of electric power equipment according to claim 3 or 4, characterized in that: The N is 2, the base learners are support vector machines and extreme learning machines, and the meta learner is a multivariate linear regression model.

6. The method for diagnosing faults of electric power equipment according to claim 1, characterized in that: The preprocessing includes denoising and normalization processing.

7. The method for diagnosing faults of electric power equipment according to claim 6, characterized in that: The denoising method is to decompose and reconstruct each signal using a variational mode decomposition method, and then sort the component signals obtained after the decomposition and reconstruction according to their importance, remove the component signal with the lowest importance, and then merge the remaining component signals.

8. The method for diagnosing faults of electric power equipment according to claim 1, characterized in that: The signals of each sensor obtained when the power equipment is running are time-synchronized signals.

9. A power collection terminal, comprising a processor, characterized in that: The processor is used to execute computer program instructions to implement the power equipment fault diagnosis method as described in any one of claims 1-8.

10. The power collection terminal according to claim 9, characterized in that: It also includes a wireless sensor module, which is used to obtain signals from different sensors when the power equipment is running after time synchronization.

Images