Magnetic element excitation waveform clustering recognition algorithm based on FT-PCA-RF

Abstract

The invention discloses a magnetic element excitation waveform classification algorithm based on FT-PCA-RF, and the method comprises the steps: determining the time-frequency characteristics of different excitation waveforms, extracting the time-domain characteristics and frequency-domain characteristics of the excitation waveforms according to the types of the excitation waveforms, and carrying out the integration of the time-domain characteristics and frequency-domain characteristics; then, on the basis of label calibration, a PCA principal component analysis method is used for judging the significance of each feature, and three labels are selected through comparison to establish a preliminary classification model; on the basis, a training set is re-divided into a test set and a training set according to the ratio of 3: 7, a random forest algorithm and a logistic regression algorithm are respectively utilized to establish a clustering recognition model, and a magnetic element excitation waveform classification algorithm of FT-PCA-RF is established to finally achieve the purpose of clustering recognition of original input waveform signals. The solution result shows that the recognition precision of the clustering recognition model can reach 98%, the excitation waveform signal can be effectively recognized according to the feature information, and the coupling relation between the excitation waveform and the magnetic core loss can be further researched.

Description

Technical Field

[0001] The invention relates to the field of excitation waveform recognition, and in particular to a magnetic element excitation waveform clustering recognition algorithm based on FT-PCA-RF. Background Art

[0002] The research on excitation waveform classification technology is constantly advancing, and new algorithms and models have been proposed and applied to practical systems. For example, methods based on wavelet transform, Fourier transform and deep learning have been widely used in excitation waveform analysis. With the development of technology, the application scenarios of waveform classification technology in power systems are also expanding, such as in emerging fields such as smart grids, distributed power generation, and wind power generation. In motors and electrical equipment, the magnitude of the excitation current has a great influence on their working stability and efficiency. Through the study of excitation waveform recognition methods, it is possible to monitor and optimize the excitation system, improve the operating efficiency of the motor, and reduce energy consumption. Different forms of excitation waveforms will have a certain impact on the electrical characteristics of the equipment, such as magnetic field distribution, torque fluctuation, efficiency and temperature rise. Accurate classification will help improve equipment performance.

[0003] Abnormal changes in the excitation waveform are usually early signs of motor, power equipment or system failure. By classifying the excitation waveform, it is possible to effectively monitor the equipment operating status, identify abnormal waveforms, and then predict faults, reducing the impact of equipment failures on production. At the same time, through waveform analysis and classification, it is possible to further locate the cause of the fault and improve the efficiency and accuracy of equipment maintenance.

[0004] With the continuous development of artificial intelligence fields such as machine learning and deep learning, the integration of the above methods can effectively improve the accuracy and automation level of excitation signal recognition. For example, using convolutional neural networks to classify excitation signals can achieve real-time and effective recognition of excitation signals and perform fault diagnosis on them. However, with the increasing complexity of power equipment and power systems, the large-scale data generated has put forward higher requirements for waveform classification technology. However, the existing excitation waveform classification technology may have problems such as low efficiency and processing bottlenecks when processing massive data. How to improve the classification performance in a big data environment and optimize the data processing process is one of the bottlenecks of current research. In summary, it is very necessary to find an algorithm that is specifically for magnetic component excitation waveform signals and can concisely and efficiently complete the clustering and recognition of such signals. Summary of the invention

[0005] In view of the above-mentioned deficiencies in the prior art, the present invention provides a magnetic component excitation waveform clustering recognition algorithm based on FT-PCA-RF.

[0006] The technical solution adopted by the present invention is a magnetic component excitation waveform clustering recognition algorithm based on FT-PCA-RF, and the overall process includes the following 6 steps.

[0007] Step 1: Construct a magnetic element excitation waveform function f[t], which is used to represent the time series of different waveforms directly acquired by waveform signal acquisition devices such as waveform signal collectors and oscilloscopes during the operation of the magnetic element.

[0008] Furthermore, the data of the magnetic element excitation waveform function f[t] in the above step 1 is composed of excitation waveform functions f1[t], f2[t], f3[t]... of different waveforms, and there is a discrete relationship between the different excitation waveform functions, and the specific structural relationship is shown in formula (1).

[0009] f(t)=f1(t)+f2(t)+…+f i (t) (1)

[0010] Step 2: Construct the operation operator TF i (·)(i=1,2,..,N) and f j (·)(j=1,2,..,M), respectively complete the purpose of extracting the time-frequency features of the input data “·”.

[0011] Furthermore, f in the above step 2 j Before the (·) operator is calculated, the Fourier transform function F(ω) needs to be used to process the magnetic element excitation waveform function f[t]. The specific structural relationship is shown in formula (2).

[0012]

[0013] In the formula, e -jωt It is a complex signal.

[0014] Furthermore, in step 2, the operation operator TF is used i (·) and f j The specific steps for solving the time-frequency characteristics of the input data “·” are as follows.

[0015] Step 2.1: Input data “·” and extract its N time domain features TF i .

[0016] Step 2.2: Perform Fourier transform on the input data “·” and extract its M frequency domain features f j .

[0017] Step 2.3: Using the operator TF i (·) and f j (·) to achieve the purpose of extracting the time-frequency features of the original magnetic element excitation waveform function f[t] of the difference waveform.

[0018] Step 2.4: According to the waveform differences, N+M time-frequency features of the difference waveforms are sorted to form an excitation waveform feature matrix H.

[0019] Step 3: Time domain characteristics FT of the magnetic component excitation waveform function f[t] i (·) and frequency domain features f j (·) Perform PCA principal component analysis to calculate the contribution C of the N time-frequency characteristics in the magnetic element excitation waveform function f[t] to the waveform difference, and sort the N time-frequency characteristics from high to low according to the calculation results.

[0020] Furthermore, the process of step 3 is as follows.

[0021] Step 3.1: Standardize the original excitation waveform feature matrix H data so that the mean of each data is 0 and the standard deviation is 1. The standardization formula is shown in formula (3).

[0022]

[0023] In the formula, n is the number of samples, m is the number of physical indicators of a single sample, μ j is the average value of the rational index, σ j is the standard deviation of the physical index.

[0024] Step 3.2: Calculate the covariance matrix COV of the excitation waveform characteristic matrix H data. The matrix calculation equation is shown in formula (4).

[0025]

[0026] In the formula, r ij is the covariance coefficient between the ith physical index and the jth physical index, and its calculation equation is shown in formula (5).

[0027]

[0028] Step 3.3: Obtain the eigenvalue λ and eigenvector v through eigenvalue decomposition. The eigenvalue decomposition equation is shown in formula (6).

[0029] COV·v=λv (6)

[0030] The eigenvalue can be solved by the characteristic polynomial, and the calculation method is shown in formula (7).

[0031] det(COV-λI)=0 (7)

[0032] The eigenvalue represents the principal component contribution C of each data. The eigenvectors are arranged according to the contribution to form an eigenvector matrix. N principal components are selected to obtain a new coordinate system.

[0033] Step 3.4: Based on the obtained orthogonal basis, perform regression analysis on the principal components, restore the indicator variables, and obtain the data after dimensionality reduction.

[0034] Step 4: According to the contribution degree C ranking and the actual situation, five time-frequency features are selected to form a feature matrix TFS, which is regarded as an identification label. Based on the feature matrix TFS data of the magnetic component excitation waveform function f[t], the RF algorithm is used to perform data clustering operation training.

[0035] Furthermore, the specific steps of the RF algorithm in step 4 are as follows.

[0036] Step 4.1: Divide the feature matrix TFS of the magnetic component excitation waveform function f[t] into k n-dimensional vectors by row, randomly select 30% of the data to form the accuracy detection matrix TETFS, and the remaining data to form the original training data matrix TFS′, determine the decision tree depth depth, the number of decision trees estimators and the maximum features of random feature selection of the RF algorithm.

[0037] Step 4.2: Use bootstrap sampling to extract samples with replacement from the original training data matrix TFS′ to form the training set matrix RFTFS of each decision tree, so as to achieve the purpose of clustering model training. The samples that are not selected form the out-of-bag sample set OBTFS, so as to achieve the purpose of estimating the general error of the model.

[0038] Step 4.3: Select random features. Select the feature with the maximum number of features as the candidate feature to split the decision tree node. Among the candidate features, select the mean square error to maximize the node purity. The mean square error calculation method is shown in formula (8).

[0039]

[0040] In the formula, MSE is the mean square error, x i is the sample data, x GT is the true label.

[0041] Step 4.4: Each decision tree starts from the root node and recursively splits until the stopping condition is met. The internal nodes that still need to be recursively split are marked, and the samples are split according to the selected features and thresholds at the internal nodes. The leaf nodes that meet the stopping condition are marked, and the cluster recognition results are determined according to the majority class or average value.

[0042] Step 4.5: Repeat steps 4.3 and 4.4 until the decision tree depth or the minimum number of samples of the leaf node is reached, and perform cluster identification on all trees of each sample of the trained model to achieve the purpose of determining the excitation waveform category based on the majority voting method in subsequent cluster identification.

[0043] Step 4.6: Input the accuracy detection matrix TETFS data and draw the confusion matrix M. The confusion matrix is ​​shown in formula (9).

[0044]

[0045] Where, T ii To correctly identify the number of the i-th waveform, F 1i is the number of waveforms that are incorrectly identified as the ith waveform.

[0046] Based on the confusion matrix M, the precision Precision is calculated to achieve the purpose of displaying the authenticity of the sample situation. The calculation method is shown in formula (10).

[0047]

[0048] Step 5: Construct the magnetic component excitation waveform test function f * [t] is used to represent the time series of unknown waveforms collected during the operation of the magnetic element.

[0049] Step 6: Based on the operator TF i (·) and f j (·), based on the feature set TFS, complete the magnetic component excitation waveform test function f * The extraction purpose of the 5 labels in [t] constitutes the recognition test matrix S * , use RF algorithm to test matrix S * Perform excitation waveform clustering and recognition operations to realize the magnetic component excitation waveform test function f * [t]The purpose of cluster identification.

[0050] In step 6 above, the RF algorithm measures the matrix S * The excitation waveform clustering recognition operation uses the majority voting method to determine the excitation waveform. The specific structural relationship of the clustering recognition operation is shown in formula (11).

[0051]

[0052] In the formula, S i is the sample set, C k is the selection decision tree, η(C k ,S i ) are the decision values ​​of the two variables, max i,K It is a voting mechanism function.

[0053] The beneficial effects of the above technical solution are as follows: the present invention provides a magnetic component excitation waveform clustering recognition algorithm based on FT-PCA-RF, constructs the obtained TFi (·) and f based on FFT decomposition j (·) specific operation operators, and the time-frequency feature separation of the original signal is realized for the magnetic component excitation waveform function; the waveform difference contribution of the time-frequency feature is obtained based on the PCA principal component analysis method and sorted based on the contribution; the feature matrix TFS containing 5 significant time domain features is obtained, and it is vectorized and trained by the RF algorithm for data clustering operation; the clustering recognition features of the magnetic component excitation waveform test function are extracted based on the feature matrix TFS, and the trained RF algorithm model is used to realize the magnetic component excitation waveform test function f * [t]The purpose of cluster identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 The present invention is a flowchart of the overall steps of a magnetic component excitation waveform clustering recognition algorithm based on FT-PCA-RF.

[0055] Figure 2 The calculation operator TF of the present invention i (·) and f j (·) The time-frequency feature calculation method involved.

[0056] Figure 3 It is a flow chart of the PCA principal component analysis steps of the time-frequency characteristics of the excitation waveform signal of the magnetic element of the present invention.

[0057] Figure 4 This is a PCA principal component analysis result diagram of the time-frequency characteristics of the excitation waveform signal of the magnetic element according to an embodiment of the present invention.

[0058] Figure 5 This is a flow chart of the steps of clustering the time-frequency characteristics of the excitation waveform TFS of a magnetic element using the RF algorithm of an embodiment of the present invention.

[0059] Figure 6 This is a graph showing the clustering results of the time-frequency characteristics of the magnetic element excitation waveform TFS using the RF algorithm of an embodiment of the present invention.

[0060] Among them, (a) is the confusion matrix; (b) is the eigenvalue importance. DETAILED DESCRIPTION

[0061] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0062] This embodiment takes the original data of three excitation waveforms, namely, sine wave, triangle wave and trapezoidal wave, directly acquired when the magnetic element is working as an example. Figure 1The six overall steps S1 to S6 shown in the figure successively complete the construction of data function, extraction of time-frequency features, feature sorting, model training, and signal recognition, thereby realizing clustering recognition of the excitation waveform signal of the new original component, as described in detail as follows.

[0063] Step 1: When the magnetic element is working, three kinds of excitation waveform signal data, namely sine wave, triangle wave and trapezoidal wave, are directly acquired, and the magnetic element excitation waveform function f[t] is constructed based on the data to represent the time series of original characteristics of different waveforms.

[0064] Furthermore, the data of the magnetic element excitation waveform function f[t] in the above step 1 is composed of excitation waveform functions f1[t], f2[t], f3[t] of different waveforms, and there is a discrete relationship between the different excitation waveform functions, and the specific structural relationship is shown in formula (1).

[0065] f(t)=f1(t)+f2(t)+f3(t) (1)

[0066] Step 2: Construct the operation operator TF i (·) and f j (·), respectively completing the purpose of extracting the time-frequency features of the input three waveform data or the processed data “·”.

[0067] Furthermore, the above-mentioned operation operator TF i (·) and f j (·) There are 5 steps to solve the time-frequency characteristics of the input three waveform data or the processed data “·”, and the specific steps are as follows.

[0068] Step 2.1: Input data “·” and extract its 16 time domain features TF i .

[0069] Step 2.2: Perform Fourier transform on the input data “·” and extract its 13 frequency domain features f j .

[0070] Step 2.3: Using the operator TF i (·) and f j (·) to achieve the purpose of extracting the time-frequency features of the original magnetic element excitation waveform function f[t] of the difference waveform.

[0071] Step 2.4: According to the waveform difference, 29 time-frequency features of the difference waveform are sorted to form the excitation waveform feature matrix H;

[0072] Furthermore, f in the above step 2 j Before the (·) operator is calculated, the Fourier transform function F(ω) needs to be used to process the magnetic element excitation waveform function f[t]. The specific structural relationship is shown in formula (2).

[0073]

[0074] In the formula, e -jωt It is a complex signal.

[0075] Step 3: Time domain characteristics FT of the magnetic component excitation waveform function f[t] i (·)(i=1...16) and frequency domain features f j (·)(j=1...13) PCA principal component analysis is performed to calculate the contribution C of the N time-frequency characteristics in the magnetic element excitation waveform function f[t] in the waveform difference, and the 29 time-frequency characteristics are sorted from high to low according to the calculation results. The steps are as follows Figure 3 As shown, the specific steps are as follows.

[0076] Step 3.1: Standardize the original excitation waveform feature matrix H data so that the mean of each data is 0 and the standard deviation is 1. The standardization formula is shown in formula (3).

[0077]

[0078] In the formula, n is the number of samples, m is the number of physical indicators of a single sample, μ j is the average value of the rational index, σ j is the standard deviation of the physical index.

[0079] Step 3.2: Calculate the covariance matrix COV of the excitation waveform characteristic matrix H data. The matrix calculation equation is shown in formula (4).

[0080]

[0081] Furthermore, ij is the covariance coefficient between the ith physical index and the jth physical index, and its calculation equation is shown in formula (5).

[0082]

[0083] Step 3.3: Obtain the eigenvalue λ and eigenvector v through eigenvalue decomposition. The eigenvalue decomposition equation is shown in formula (6).

[0084] COV·v=λv (6)

[0085] Furthermore, the eigenvalue can be solved by the characteristic polynomial, and the calculation method is shown in formula (7).

[0086] det(COV-λI)=0 (7)

[0087] The eigenvalue represents the contribution C of each data to the principal component. The eigenvectors are arranged according to the contribution to form an eigenvector matrix. Five principal components are selected to obtain a new coordinate system.

[0088] Step 3.4: Based on the obtained orthogonal basis, perform regression analysis on the principal components, restore the indicator variables, and obtain the data after dimensionality reduction.

[0089] Step 4: Select 5 time-frequency features according to the contribution degree C and the actual situation to form a feature matrix TFS, which is regarded as an identification label. Based on the feature matrix TFS data of the magnetic component excitation waveform function f[t], the RF algorithm is used to perform data clustering operation training. The execution flow chart is as follows: Figure 5 As shown, the specific steps are as follows.

[0090] Step 4.1: Divide the feature matrix TFS of the magnetic element excitation waveform function f[t] into 5 1024-dimensional vectors by row, randomly select 30% of the data to form the accuracy detection matrix TETFS, and the remaining data to form the original training data matrix TFS′, and determine the decision tree depth of the RF algorithm depth = 15, the number of decision trees estimators = 325, and the maximum random feature selection features = 5.

[0091] Step 4.2: Use bootstrap sampling to extract samples with replacement from the original training data matrix TFS′ to form the training set matrix RFTFS of each decision tree, so as to achieve the purpose of clustering model training. The samples that are not selected form the out-of-bag sample set OBTFS, so as to achieve the purpose of estimating the general error of the model.

[0092] Step 4.3: Select random features. Select the feature with the maximum number of features as the candidate feature to split the decision tree node. Among the candidate features, select the mean square error to maximize the node purity. The mean square error calculation method is shown in formula (8).

[0093]

[0094] In the formula, MSE is the mean square error, x i is the sample data, x GT is the true label.

[0095] Step 4.4: Each decision tree starts from the root node and recursively splits until the stopping condition is met. The internal nodes that still need to be recursively split are marked, and the samples are split according to the selected features and thresholds at the internal nodes. The leaf nodes that meet the stopping condition are marked, and the recognition results are determined according to the majority class or the average value.

[0096] Step 4.5: Repeat steps 4.3 and 4.4 until the decision tree depth or the minimum number of samples of the leaf node is reached, and cluster identification is performed on all trees of the three waveforms of the trained model to achieve the purpose of determining the excitation waveform category based on the majority voting method in subsequent cluster identification.

[0097] Step 4.6: Input the accuracy detection matrix TETFS data and draw the confusion matrix M. The confusion matrix is ​​shown in formula (9).

[0098]

[0099] Furthermore, the specific structure of the confusion matrix is ​​shown in formula (10).

[0100]

[0101] Where, T ii To correctly identify the number of the i-th waveform, F 1i is the number of waveforms that are incorrectly identified as the ith waveform.

[0102] Based on the confusion matrix M, the accuracy Precision is calculated to be 98.876%, achieving the purpose of displaying the authenticity of the sample situation. The calculation method is shown in formula (11).

[0103]

[0104] Step 5: The three excitation waveform signal data of unknown waveforms directly acquired when the magnetic element is working are used to construct the magnetic element excitation waveform test function f * [t] is used to represent the time series of unknown waveforms collected during the operation of the magnetic element.

[0105] Step 6: Based on the operator TF i (·) and f j (·), based on the feature set TFS, complete the magnetic component excitation waveform test function f * The purpose of extracting the 5 identification labels in [t] is to form the identification test matrix S * , use RF algorithm to test matrix S * Perform excitation waveform clustering and recognition operations to realize the magnetic component excitation waveform test function f * [t] For the purpose of clustering identification, the clustering results of the test set are as follows.

[0106]

[0107] In step 6 above, the RF algorithm measures the matrix S *The excitation waveform clustering recognition operation uses the majority voting method to determine the excitation waveform. The specific structural relationship of the clustering recognition operation is shown in formula (12).

[0108]

[0109] In the formula, S i is the sample set; C k is the selection decision tree; η(C k ,S i ) is the decision value of the two variables; max i,K It is a voting mechanism function.

[0110] Although the present invention has been disclosed as above in terms of preferred embodiments, they are not intended to limit the present invention. Anyone skilled in the art can make various changes or modifications without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection defined by the claims of this application.

Claims

1. A magnetic component excitation waveform clustering recognition algorithm based on FT-PCA-RF, characterized in that: The method includes the following six steps: Step 1: construct a magnetic element excitation waveform function f[t], which is used to represent the time series of different waveforms directly acquired by waveform signal acquisition devices such as waveform signal acquisition devices and oscilloscopes during the working process of the magnetic element; In the above step 1, the magnetic element excitation waveform function f[t] data is composed of different waveform excitation waveform functions f1[t], f2[t], f3[t]..., and there is a discrete relationship between different excitation waveform functions. The specific structural relationship is shown in formula (1): f(t)=f1(t)+f2(t)+...+f i (t) (1) Step 2: Construct the operation operator TF i (·) and f j (·), respectively completing the purpose of extracting the time-frequency features of the input data "·"; f in step 2 above j Before the (·) operator is calculated, the Fourier transform function F(ω) needs to be used to process the magnetic element excitation waveform function f[t]. Its specific structural relationship is shown in formula (2): In the formula, e -jωt is a complex signal; Step 3: Time domain characteristics FT of the magnetic component excitation waveform function f[t] i and frequency domain characteristics f j Perform PCA principal component analysis to calculate the contribution C of N time-frequency characteristics in the magnetic element excitation waveform function f[t] to the waveform difference, and sort the N time-frequency characteristics from high to low according to the calculation results; Step 4: Select 5 time-frequency features to form a feature matrix TFS according to the contribution degree C ranking and the actual situation, which is regarded as the identification label. Based on the feature matrix TFS data of the magnetic component excitation waveform function f[t], the RF algorithm is used to perform data clustering operation training; Step 5: Construct the magnetic component excitation waveform test function f * [t], used to represent the time series of unknown waveforms collected during the operation of the magnetic element; Step 6: Based on the operator TF i (·) and f j (·), based on the feature set TFS, complete the magnetic component excitation waveform test function f * The purpose of extracting the 5 identification labels in [t] is to form the identification test matrix S * , use RF algorithm to test matrix S * Perform excitation waveform clustering and recognition operations to realize the magnetic component excitation waveform test function f * [t] The purpose of cluster identification; In step 6 above, the RF algorithm measures the matrix S * The excitation waveform clustering recognition operation uses the majority voting method to determine the excitation waveform. The specific structural relationship of the clustering recognition operation is shown in formula (3): In the formula, S i is the sample set, C k is the selection decision tree, η(C k ,S i ) are the decision values ​​of the two variables, max i,K It is a voting mechanism function.

2. The magnetic component excitation waveform clustering recognition algorithm based on FT-PCA-RF according to claim 1 is characterized in that: In step 2, the operator TF is calculated i (·) and f j The solution of (·) is as follows: Step 2.1: Input data "·" and extract its N time domain features TF i ; Step 2.2: Perform Fourier transform on the input data "·" and extract its M frequency domain features f j ; Step 2.3: Using the operator TF i (·) and f j (·) to achieve the purpose of extracting the time-frequency features of the original magnetic element excitation waveform function f[t] of the difference waveform; Step 2.4: According to the waveform differences, N+M time-frequency features of the difference waveforms are sorted to form an excitation waveform feature matrix H.

3. The magnetic component excitation waveform clustering recognition algorithm based on FT-PCA-RF according to claim 1 is characterized in that: The process of step 3 is as follows: Step 3.1: Standardize the original excitation waveform feature matrix H data so that the mean of each data is 0 and the standard deviation is 1. The standardization formula is shown in formula (4): In the formula, n is the number of samples, m is the number of physical indicators of a single sample, μ j is the average value of the rational index, σ j is the standard deviation of the physical index; Step 3.2: Calculate the covariance matrix COV of the excitation waveform characteristic matrix H data. The matrix calculation equation is shown in formula (5): In the formula, r ij is the covariance coefficient between the ith physical index and the jth physical index, and its calculation equation is shown in formula (6): Step 3.3: Obtain the eigenvalue λ and eigenvector v through eigenvalue decomposition. The eigenvalue decomposition equation is shown in formula (7): COV·v=λv (7) The eigenvalue can be solved by the characteristic polynomial, and the calculation method is shown in formula (8): det(COV-λI)=0 (8) The eigenvalue represents the principal component contribution C of each data. The eigenvectors are arranged according to the contribution to form an eigenvector matrix. N principal components are selected to obtain a new coordinate system. Step 3.4: Based on the obtained orthogonal basis, perform regression analysis on the principal components, restore the indicator variables, and obtain the data after dimensionality reduction.

4. The magnetic component excitation waveform clustering recognition algorithm based on FT-PCA-RF according to claim 1 is characterized in that: The solution method of the RF algorithm in step 4 is as follows: Step 4.1: Divide the characteristic matrix TFS of the magnetic component excitation waveform function f[t] into k n-dimensional vectors by row, randomly select 30% of the data to form the accuracy detection matrix TETFS, and the remaining data to form the original training data matrix TFS ′ , determine the decision tree depth of the RF algorithm, the number of decision trees estimators and the maximum features of random feature selection; Step 4.2: Use bootstrap sampling to extract the original training data matrix TFS ′ The samples drawn with replacement constitute the training set matrix RFTFS of each decision tree, which realizes the purpose of clustering model training. The samples not drawn constitute the out-of-bag sample set OBTFS, which realizes the purpose of estimating the general error of the model. Step 4.3: Select random features. Select the feature with the maximum number of features as the candidate feature to split the decision tree node. Among the candidate features, select the mean square error to maximize the node purity. The mean square error calculation method is shown in formula (9): In the formula, MSE is the mean square error, x i is the sample data, x GT is the true label; Step 4.4: Each decision tree starts from the root node and recursively splits until the stopping condition is met. The internal nodes that still need to be recursively split are marked, and the samples are split according to the selected features and thresholds at the internal nodes. The leaf nodes that meet the stopping condition are marked, and the prediction results are determined according to the majority class or the average value. Step 4.5: Repeat steps 4.3 and 4.4 until the decision tree depth or the minimum number of samples of the leaf node is reached, and perform clustering identification on all trees of each sample of the trained model to achieve the purpose of determining the excitation waveform category based on the majority voting method in subsequent predictions; Step 4.6: Input the precision detection matrix TETFS data and draw the confusion matrix M. The confusion matrix is ​​shown in formula (10): Where, T ii To correctly identify the number of the i-th waveform, F 1i is the number of waveforms that are mistakenly identified as the ith waveform; Based on the confusion matrix M, the precision Precision is calculated to achieve the purpose of displaying the authenticity of the sample situation. The calculation method is shown in formula (11):

Images