A method for detecting scale thickness on equalizing electrodes based on multidimensional features and deep learning
By combining multidimensional features and deep learning, the problem of accuracy and precision in detecting scale buildup on equalizing electrodes in converter valve cooling systems has been solved, achieving efficient and accurate scale thickness identification and ensuring the safe operation of power transmission systems.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HUNAN UNIV
- Filing Date
- 2023-12-20
- Publication Date
- 2026-06-30
AI Technical Summary
Existing technologies for detecting scale buildup on equalizing electrodes in converter valve cooling systems suffer from low accuracy and a high risk of overfitting, making it difficult to achieve efficient and high-precision scale thickness identification, which affects the safety of power transmission systems.
By combining multidimensional features and deep learning, variational mode decomposition and wavelet packet decomposition are used to denoise the ultrasonic detection signal and extract its time and frequency domain features. Convolutional neural networks and recurrent neural networks are then combined to construct a recognition model, thereby improving the accuracy and precision of scale thickness identification.
It enhances the model's generalization ability and robustness, improves the identification accuracy and precision of the scale thickness of the equalizing electrode, reduces redundant feature information, and improves detection efficiency and accuracy.
Smart Images
Figure CN118094306B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of detection technology for the scale thickness of equalizing electrodes in converter valve cooling systems, specifically, to a method for detecting the scale thickness of equalizing electrodes based on multidimensional features and deep learning. Background Technology
[0002] Converter valves are critical equipment in high-voltage direct current (HVDC) transmission systems. During operation, they generate significant heat, which is removed by the valve cooling system using circulating cooling water, ensuring normal valve operation. However, the equalizing electrodes continuously accumulate scale during operation. Scale shedding can cause blockages in cooling water pipes and even converter station shutdowns. Therefore, scale buildup on the equalizing electrodes in the converter valve cooling system is one of the weak points in current HVDC transmission systems. According to relevant data, scale buildup on the equalizing electrodes and radiator corrosion account for 70% of all transmission system failures caused by valve cooling systems, seriously threatening the safe operation of the transmission system. Currently, converter stations address equalizing electrode scale buildup through manual disassembly and screening, which is not only inefficient but also highly unpredictable, easily leading to scale shedding and water circuit blockages, affecting radiator heat dissipation performance. Therefore, achieving efficient and high-precision detection of the equalizing electrode scale thickness is crucial for ensuring the safe operation of the transmission system.
[0003] In the field of nondestructive testing, ultrasonic testing, with its high sensitivity, strong penetration, and ease of use, is a powerful method for detecting scale buildup on equalizing electrodes. Some studies have already verified the feasibility of ultrasonic testing for scale buildup on equalizing electrodes and achieved certain results. However, problems remain, such as incomplete feature extraction coverage, difficulty in fully representing scale thickness information, and the risk of overfitting in the model. Further improvements in the accuracy and precision of scale thickness identification are still needed. Summary of the Invention
[0004] The main objective of this invention is to propose a method for detecting the scale thickness of equalizing electrodes based on multidimensional features and deep learning. By fully extracting scale thickness information, the method enhances the generalization ability and robustness of the model, improves the identification accuracy and precision of scale formation on equalizing electrodes in converter stations, and thus solves the aforementioned technical problems.
[0005] To achieve the above objectives, this invention proposes a method for detecting the scale thickness of a voltage equalizing electrode based on multidimensional features and deep learning, comprising the following steps:
[0006] S1. Obtain ultrasonic detection signals of the equalizing electrode under different scale thicknesses, and divide the data of the ultrasonic detection signals into a training set and a test set;
[0007] S2. After denoising the ultrasonic detection signal using Variational Mode Decomposition (VMD) and Wavelet Packet Decomposition (WPD), an optimized detection signal is obtained.
[0008] S3. Extract 6-dimensional time-domain and 4-dimensional frequency-domain features from the optimized detection signal to form a one-dimensional feature vector. After standardization, perform principal component analysis (PCA) to reduce the dimensionality and obtain the first feature.
[0009] S4. The optimized detection signal is decomposed using the WPD method, and the energy feature distribution of the last wavelet packet node is calculated. After standardization, the signal is mapped and reduced in dimension using the PCA method to obtain the second feature.
[0010] S5. Combine the first feature and the second feature to form a one-dimensional feature vector;
[0011] S6. The optimized detection signal is converted into a two-dimensional feature matrix using a Markov migration field.
[0012] S7. Construct a convolutional neural network (CNN) with multiple convolutional and pooling layers using the two-dimensional feature matrix.
[0013] S8. Construct a recurrent neural network (RNN) with multiple memory layers using the one-dimensional feature vector;
[0014] S9. Combine the fully connected layer of the CNN with the final memory layer of the RNN, and connect them to the output layer;
[0015] S10. The recognition model constructed through continuous training outputs the detection result of the scale thickness after reaching the preset accuracy.
[0016] Preferably, the training set accounts for 70% of the total dataset, and the test set accounts for 30% of the total dataset.
[0017] Preferably, step S2 specifically includes:
[0018] S2.1. The characteristic modes and noise modes of the ultrasonic detection signal are separated using the VMD method, and the upper limit of the screening frequency f is set according to the modal center frequency characteristics. max With lower limit f min The ultrasonic detection signal is initially denoised by adaptively selecting the feature modes.
[0019] Preferably, after step S21, the method further includes:
[0020] S2.2 The ultrasonic detection signal after preliminary noise reduction is decomposed using the WPD method, and the wavelet packet node coefficients are processed by setting a small threshold using the soft threshold noise reduction method to perform detailed noise reduction on the ultrasonic detection signal to obtain the optimized detection signal.
[0021] Preferably, the 6-dimensional time domain in step S3 includes: waveform features x wPulse characteristics x i Peak characteristics x e Margin feature x l Skewness feature x s kurtosis feature x k The 4-dimensional frequency domain includes: the centroid frequency f c Average frequency f m Root mean square frequency f rms Frequency variance f v The first feature is a time-domain and frequency-domain feature vector [x] w ,x i ,x e ,x l ,x s ,x k ,f c ,f m ,f rms ,f v ]
[0022] Preferably, step S4 specifically includes:
[0023] First, calculate the energy value of each wavelet packet node to obtain the energy vector P. The calculation steps are shown in Formula 1.
[0024]
[0025] Among them, P i Let x(n) represent the energy of the i-th wavelet packet node, x(n) represent the coefficient value of the n-th wavelet packet in the current node, and N represent the length of the wavelet packet coefficient data of the current node.
[0026] Next, calculate the energy distribution characteristics among several wavelet packet nodes. The calculation steps are shown in Formula 2.
[0027]
[0028] Where E represents the wavelet packet energy probability density, P represents the wavelet packet energy vector, and l represents the decomposition level;
[0029] Finally, the principal component features are obtained by mapping the wavelet packet energy distribution feature vector using the PCA method. The principal component features are sorted from largest to smallest according to their contribution rate, and several principal component features with larger cumulative contribution rates are selected as the second features.
[0030] Preferably, in step S7, the convolutional neural network (CNN) sequentially includes an input layer, a convolutional layer, a max-pooling layer, and a fully connected layer; the convolutional layer and the max-pooling layer are alternately set, with three layers set respectively; the activation function of the convolutional layer is "ReLU", and both the convolutional layer and the max-pooling layer adopt a zero-padding strategy.
[0031] Preferably, in step S8, the recurrent neural network (RNN) includes an input layer and a memory layer, with a total of 4 memory layers, and the activation function is selected as "tanh".
[0032] Preferably, step S10 specifically includes,
[0033] Determine whether the output result of the output layer meets the preset precision;
[0034] S10.1 If yes, output the detection result;
[0035] S10.2 If not, return the output result and execute steps S7 and S8 respectively.
[0036] Preferably, step S10.2 specifically includes: returning the two-dimensional feature matrix in the output result as the input of the CNN, and executing step S7; returning the one-dimensional feature vector in the output result as the input of the RNN, and executing step S8; the optimizer of the recognition model uses the "adam" optimizer, and the loss function uses "sparse categorical crossentropy".
[0037] This invention provides a method for detecting the scale thickness of equalizing electrodes based on multidimensional features and deep learning. It analyzes the distribution characteristics of ultrasonic detection signals of scale on equalizing electrodes from multiple dimensions, including time domain, frequency domain, time-frequency domain, and spatiotemporal domain, fully extracting the scale thickness information contained in the signals. Furthermore, it employs principal component analysis to reduce the dimensionality of the features, eliminating redundant information and reducing the number of parameters in the recognition model. This invention utilizes CNN to process the two-dimensional feature matrix and RNN to process the one-dimensional feature vector, fully leveraging their respective performance advantages. Furthermore, it integrates the CNN and RNN models, enhancing the model's generalization ability and robustness, and improving the accuracy and precision of identifying the scale thickness of equalizing electrodes. Attached Figure Description
[0038] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.
[0039] Figure 1 This is a flowchart of the method for multidimensional feature extraction and CNN-RNN parallel fusion model of ultrasonic detection signal for scaling of equalizing electrode in this invention;
[0040] Figure 2The present invention provides time-domain waveforms and frequency-domain distributions of ultrasonic detection signals from equalizing electrodes at seven different scale thicknesses, collected in this embodiment of the invention.
[0041] Figure 3 These are the time-domain and frequency-domain diagrams of two different modes obtained after VMD decomposition of the ultrasonic detection signal in this embodiment of the invention.
[0042] Figure 4 This is a comparison of the time and frequency domains of the ultrasonic detection signals before and after noise reduction in an embodiment of the present invention.
[0043] Figure 5 These are the seven types of one-dimensional feature vectors extracted in step S5 of this embodiment of the invention;
[0044] Figure 6 These are the seven types of two-dimensional feature matrices extracted in step S6 of this embodiment of the invention;
[0045] Figure 7 This is the confusion matrix of the training set recognition results in this embodiment of the invention.
[0046] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation
[0047] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.
[0048] Furthermore, the technical solutions of the various embodiments of the present invention can be combined with each other, but only if they are feasible for those skilled in the art. If the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.
[0049] Reference Figure 1 This invention discloses a method for extracting multi-dimensional features in the time domain, frequency domain, time-frequency domain, and spatiotemporal domain of ultrasonic detection signals for scaling on equalizing electrodes, and a recognition model method using parallel CNN-RNN fusion, comprising the following steps:
[0050] S1. Obtain ultrasonic detection signals of the equalizing electrode under different scale thicknesses, and divide the data of the ultrasonic detection signals into a training set and a test set;
[0051] S2. After denoising the ultrasonic detection signal using Variational Mode Decomposition (VMD) and Wavelet Packet Decomposition (WPD), an optimized detection signal is obtained.
[0052] S2.1. The characteristic modes and noise modes of the ultrasonic detection signal are separated using the VMD method, and the upper limit of the screening frequency f is set according to the modal center frequency characteristics. max With lower limit f min The ultrasonic detection signal is initially denoised by adaptively selecting the feature modes.
[0053] S2.2 The ultrasonic detection signal after preliminary noise reduction is decomposed using the WPD method, and the wavelet packet node coefficients are processed by setting a small threshold using the soft threshold noise reduction method to perform detailed noise reduction on the ultrasonic detection signal to obtain the optimized detection signal.
[0054] S3. Extract 6-dimensional time-domain and 4-dimensional frequency-domain features from the optimized detection signal to form a one-dimensional feature vector. After standardization, perform principal component analysis (PCA) to reduce the dimensionality and obtain the first feature.
[0055] S4. The optimized detection signal is decomposed using the WPD method, and the energy feature distribution of the last wavelet packet node is calculated. After standardization, the signal is mapped and reduced in dimension using the PCA method to obtain the second feature.
[0056] S5. Combine the first feature and the second feature to form a one-dimensional feature vector;
[0057] S6. The optimized detection signal is converted into a two-dimensional feature matrix using a Markov migration field.
[0058] S7. Construct a convolutional neural network (CNN) with multiple convolutional and pooling layers using the two-dimensional feature matrix.
[0059] S8. Construct a recurrent neural network (RNN) with multiple memory layers using the one-dimensional feature vector;
[0060] S9. Combine the fully connected layer of the CNN with the final memory layer of the RNN, and connect them to the output layer;
[0061] S10. After continuously training the constructed recognition model, output the detection result of the scale thickness after reaching the preset accuracy.
[0062] Determine whether the output result of the output layer meets the preset precision;
[0063] S10.1 If yes, output the detection result;
[0064] S10.2 If not, return the output result and execute steps S7 and S8 respectively.
[0065] The following details each step.
[0066] Step S1: Obtain the pressure equalization electrode detection signals under different scale thicknesses using an ultrasonic testing device, and divide them into training set and test set;
[0067] In this embodiment, seven different scale thicknesses of the equalizing electrode were tested using an ultrasonic testing device. Data was collected 100 times for each scale thickness, resulting in a total of 700 sets of sample data. The seven scale thicknesses were A: 0 mm, B: 0.09 mm, C: 0.175 mm, D: 0.195 mm, E: 0.29 mm, F: 0.36 mm, and G: 0.515 mm. The time-domain and frequency-domain graphs of the equalizing electrode signals for the seven different scale thicknesses are shown below. Figure 2 As shown. By Figure 2 It is evident that the seven types of ultrasound detection signals exhibit differences in both the time-domain waveform and the frequency-domain spectrum, providing favorable conditions for feature extraction and intelligent recognition. The 700 sets of sample data were divided in a 7:3 ratio, with 490 sets in the training set and 210 sets in the test set.
[0068] Step S2: The ultrasonic detection signal is denoised using Variational Mode Decomposition (VMD) and Wavelet Packet Decomposition (WPD). In this embodiment, the VMD penalty factor is set to a commonly used value of 2000. When the decomposition level is 4, the characteristic modes and noise modes of the ultrasonic detection signal can be separated, resulting in two different forms of decomposition results, such as... Figure 3 As shown in the VMD decomposition mode diagram, the characteristic mode amplitude of the ultrasonic detection signal is relatively large, and the center frequency is mainly located between 1MHz and 2MHz.
[0069] Based on the center frequency of the mode, a certain margin can be left, allowing for an upper limit f of the screening frequency. max It is 0.5MHz, with a lower limit f. min The frequency is 2.5MHz, thereby enabling adaptive selection of feature modes.
[0070] For the signal after feature mode reconstruction, wavelet packet decomposition is used for noise reduction. The decomposition level is set to 3, the wavelet basis function is "sym10", the thresholding method is soft thresholding, and the threshold is set to 0.008.
[0071] Comparison of waveforms and frequency domains before and after noise reduction Figure 4 As shown in the time-domain comparison diagram, the ultrasonic detection signal waveform becomes smoother, and high-frequency jitter noise is significantly eliminated; the frequency-domain comparison diagram shows that low-frequency and high-frequency noise in the ultrasonic detection signal is effectively suppressed. The average signal-to-noise ratio of 700 sets of ultrasonic detection signals reaches 29.63dB, with a minimum of 22.34dB; the average root mean square error reaches 0.3128, with a maximum of 0.4758, demonstrating good noise reduction effect.
[0072] Step S3: Extract the 6-dimensional time domain and 4-dimensional frequency domain features of the detection signal processed in step S2, form a one-dimensional feature vector, and then perform mapping and dimensionality reduction using principal component analysis (PCA) after standardization.
[0073] In this embodiment, for the denoised signal, the 6-dimensional time-domain features are calculated according to the formula in Table 1;
[0074]
[0075] Table 1
[0076] In the formulas in Table 1, x p =max|x|, N is the length of the wavelet packet coefficient data of the current node.
[0077] Then calculate the 4-dimensional frequency domain features according to the formulas in Table 2;
[0078]
[0079] Table 2
[0080] In the formulas in Table 2, y(i) is the amplitude corresponding to the frequency after Fourier transform, and N is the data length.
[0081] Based on the above calculations, the time-domain and frequency-domain eigenvectors [x] can be obtained. w ,x i ,x e ,x l ,x s ,x k ,f c ,f m ,f rms ,f v ].
[0082] Furthermore, through the minimax standardization formula Process the sample set to eliminate differences in dimensions and magnitudes between features.
[0083] Furthermore, the principal component features are obtained by mapping the time-domain and frequency-domain feature vectors using the PCA method. The principal component features are sorted from largest to smallest according to their contribution rate. When the cumulative contribution rate is greater than 90%, there are a total of 3 principal component features, which successfully reduces the 10-dimensional features to 3-dimensional features, greatly reducing the number of model parameters.
[0084] Step S4: The detected signal processed in step S2 is decomposed by WPD, and the energy feature distribution of the last wavelet packet node is calculated. After standardization, it is mapped and reduced in dimension by PCA.
[0085] In this embodiment, the WPD decomposition level is set to 5 levels, and the wavelet basis function is selected as "sym10". Where the sampling rate f s With a frequency of 12.5MHz, the time-frequency resolution f can be obtained. d At 390.625kHz, it already has a relatively high resolution.
[0086] According to formula (1) Calculate the energy value of each of the 32 wavelet packet nodes to obtain the energy vector P; where P i Let x(n) represent the energy of the i-th wavelet packet node, x(n) represent the coefficient value of the n-th wavelet packet in the current node, and N represent the length of the wavelet packet coefficient data of the current node.
[0087] According to formula (2) Calculate the energy distribution characteristics among 32 wavelet packet nodes; where E represents the wavelet packet energy probability density, P represents the wavelet packet energy vector, and l represents the decomposition level.
[0088] Furthermore, through the minimax standardization formula Processing the sample set eliminates the differences in magnitude between features, providing favorable conditions for principal component analysis.
[0089] Furthermore, the principal component features are obtained by mapping the wavelet packet energy distribution feature vector through the PCA method. The principal component features are sorted from largest to smallest according to their contribution rate. When the cumulative contribution rate is greater than 90%, there are a total of 7 principal component features, which successfully reduces the 32-dimensional features to 7-dimensional features, greatly reducing the number of model parameters.
[0090] Step S5: Combine the features obtained in steps S3 and S4 to form a one-dimensional feature vector;
[0091] In this embodiment, the 3D principal component features obtained in step S3 and the 7D principal component features obtained in step S4 are combined into a 10D feature vector. The one-dimensional features of the 7 types of ultrasound detection signals are as follows: Figure 5 As shown, the distribution of one-dimensional features of the seven types of ultrasound detection signals exhibits strong differences in amplitude and dimension.
[0092] Step S6: For the detection signal processed in step S2, a Markov migration field is used to convert the detection signal into a two-dimensional feature matrix;
[0093] In this embodiment, the image size after Markov migration field transformation is set to 300, the Markov migration field quantile is set to 10, and the quantile width strategy is set to uniform mode, resulting in a two-dimensional feature matrix of 7 types of ultrasound detection signals, as shown below. Figure 6 As shown, the two-dimensional matrix feature distributions of the seven types of ultrasound detection signals all exhibit strong differences.
[0094] Step S7: Construct a convolutional neural network (CNN) with multiple convolutional and pooling layers;
[0095] In this embodiment, the input features of the CNN input layer are a two-dimensional feature matrix with an input size of (300, 300, 1); Convolutional layer 1 has 6 convolutional kernels with a kernel size of 4×4 and a stride of 4, using a zero-padding strategy, and the activation function is "ReLU"; Max pooling layer 1 has a pooling size of 2×2 and a stride of 2, using a zero-padding strategy; Convolutional layer 2 has 9 convolutional kernels with a kernel size of 4×4 and a stride of 4, using a zero-padding strategy, and the activation function is "ReLU"; Max pooling layer 2 has a pooling size of 2×2 and a stride of 2, using a zero-padding strategy; Convolutional layer 3 has 12 convolutional kernels with a kernel size of 2×2 and a stride of 2, using a zero-padding strategy, and the activation function is "ReLU"; Max pooling layer 1 has a pooling size of 2×2 and a stride of 2, using a zero-padding strategy; and a fully connected layer is added at the end.
[0096] Step S8: Construct a recurrent neural network (RNN) with multiple memory layers;
[0097] In this embodiment, the input features of the RNN input layer are one-dimensional feature vectors, and the input size is set to 10; memory layer 1 has 20 neurons; memory layer 2 has 28 neurons; memory layer 3 has 34 neurons; and memory layer 4 has 48 neurons; all four memory layers use the "tanh" activation function.
[0098] Step S9: Combine the fully connected layer of the CNN with the final memory layer of the RNN, and connect them to the output layer;
[0099] In this embodiment, the number of neurons in the fully connected layer of the CNN is 48, which is the same as the number of neurons in the final memory layer of the RNN. Therefore, the joint layer has a total of 96 neurons, the number of neurons in the output layer is set to 7, and the activation function is the "softmax" function.
[0100] Step S10: Train the constructed recognition model and output the recognition result after the accuracy requirement is met.
[0101] Furthermore, step S10 specifically includes,
[0102] Determine whether the output result of the output layer meets the preset precision;
[0103] S10.1 If yes, output the detection result;
[0104] S10.2 If not, return the output result and execute steps S7 and S8 respectively;
[0105] Specifically, step S10.2 includes: returning the two-dimensional feature matrix in the output result as the input of the CNN, and executing step S7; returning the one-dimensional feature vector in the output result as the input of the RNN, and executing step S8; the optimizer of the recognition model uses the "adam" optimizer, and the loss function uses "sparse categorical crossentropy".
[0106] In this embodiment, the model optimizer is set to "adam", the loss function is selected as "sparsecategorical cross entropy", and the training epochs are set to 15.
[0107] The training set accuracy was 98.57% with a loss of 0.1059; the test set accuracy was 98.10% with a loss of 0.1155. The confusion matrix between the training and test sets is as follows: Figure 7 As shown in the confusion matrix diagram, the overall accuracy of identifying the scale thickness of the equalizing electrode is relatively high.
[0108] In summary, in conjunction with the embodiments of the present invention, the feature method of the present invention can fully extract the scale thickness information from the ultrasonic detection signal of the equalizing electrode, providing a good recognition basis for the recognition model. At the same time, the PCA method can effectively eliminate redundant information in the features and reduce the number of parameters in the recognition model. Moreover, the fusion of CNN-RNN models enhances the generalization ability and robustness of the model, and improves the recognition accuracy and recognition rate of the scale thickness of the equalizing electrode.
[0109] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. All equivalent transformations made based on the inventive concept of the present invention and the contents of the specification and drawings of the present invention, or direct / indirect applications in other related technical fields, are included within the patent protection scope of the present invention.
Claims
1. A method for detecting the thickness of scale layer on a voltage equalizing electrode based on multidimensional features and deep learning, characterized in that, Includes the following steps: S1. Obtain ultrasonic detection signals of the equalizing electrode under different scale thicknesses, and divide the data of the ultrasonic detection signals into a training set and a test set; S2. After denoising the ultrasonic detection signal using Variational Mode Decomposition (VMD) and Wavelet Packet Decomposition (WPD), an optimized detection signal is obtained; wherein, step S2 specifically includes: S2.
1. The characteristic modes and noise modes of the ultrasonic detection signal are separated using the VMD method, and an upper limit for the screening frequency is set according to the modal center frequency characteristics. f max and lower limit f min The ultrasonic detection signal is initially denoised by adaptively selecting the feature modalities. S2.2 The ultrasonic detection signal after preliminary noise reduction is decomposed using the WPD method, and the threshold processing wavelet packet node coefficients are set by the soft threshold noise reduction method to perform detailed noise reduction on the ultrasonic detection signal to obtain the optimized detection signal. S3. Extract 6-dimensional time-domain and 4-dimensional frequency-domain features from the optimized detection signal to form a one-dimensional feature vector. After standardization, perform principal component analysis (PCA) to reduce the dimensionality and obtain the first feature. The 6-dimensional time domain in step S3 includes: waveform features Pulse characteristics Peak characteristics Margin characteristics Skewness characteristics kurtosis characteristics The 4-dimensional frequency domain includes: centroid frequency. Average frequency Root mean square frequency Frequency variance The first feature is a time-domain and frequency-domain feature vector. S4. The optimized detection signal is decomposed using the WPD method, and the energy feature distribution of the last wavelet packet node is calculated. After standardization, the signal is mapped and reduced in dimension using the PCA method to obtain the second feature. S5. Combine the first feature and the second feature to form a one-dimensional feature vector; S6. The optimized detection signal is converted into a two-dimensional feature matrix using a Markov migration field. S7. Construct a convolutional neural network (CNN) with multiple convolutional and pooling layers using the two-dimensional feature matrix. S8. Construct a recurrent neural network (RNN) with multiple memory layers using the one-dimensional feature vector; S9. Combine the fully connected layer of the CNN with the final memory layer of the RNN, and connect them to the output layer; S10. The recognition model constructed through continuous training outputs the detection result of the scale thickness after reaching the preset accuracy.
2. The method for detecting the scale thickness of a voltage equalizing electrode based on multidimensional features and deep learning as described in claim 1, characterized in that, The training set accounts for 70% of the total dataset, and the test set accounts for 30% of the total dataset.
3. The method for detecting the scale thickness of a voltage equalizing electrode based on multidimensional features and deep learning as described in claim 1, characterized in that, Step S4 specifically includes: First, calculate the energy value of each wavelet packet node to obtain the energy vector P. The calculation steps are shown in Formula 1. （1） in, This represents the energy of the i-th wavelet packet node. This represents the coefficient value of the nth wavelet packet in the current node, where N represents the length of the wavelet packet coefficient data in the current node. Next, calculate the energy distribution characteristics among several wavelet packet nodes. The calculation steps are shown in Formula 2. （2） Where E represents the wavelet packet energy probability density, P represents the wavelet packet energy vector, and l represents the decomposition level; Finally, the principal component features are obtained by mapping the wavelet packet energy distribution feature vector using the PCA method. The principal component features are sorted from largest to smallest according to their contribution rate, and several principal component features with a cumulative contribution rate greater than 90% are selected as the second features.
4. The method for detecting the scale thickness of a voltage equalizing electrode based on multidimensional features and deep learning as described in claim 1, characterized in that: In step S7, the convolutional neural network (CNN) sequentially includes an input layer, a convolutional layer, a max-pooling layer, and a fully connected layer; the convolutional layer and the max-pooling layer are alternately set, with three layers set respectively; the activation function of the convolutional layer is "ReLU", and both the convolutional layer and the max-pooling layer adopt a zero-padding strategy.
5. The method for detecting the scale thickness of a voltage equalizing electrode based on multidimensional features and deep learning as described in claim 1, characterized in that: In step S8, the recurrent neural network (RNN) includes an input layer and a memory layer, with a total of 4 memory layers and the activation function selected as "tanh".
6. The method for detecting the scale thickness of a voltage equalizing electrode based on multidimensional features and deep learning as described in claim 1, characterized in that: Step S10 specifically includes, Determine whether the output result of the output layer meets the preset precision; S10.1 If yes, output the detection result; S10.2 If not, return the output result and execute steps S7 and S8 respectively.
7. The method for detecting the scale thickness of a voltage equalizing electrode based on multidimensional features and deep learning as described in claim 6, characterized in that: Step S10.2 specifically includes: returning the two-dimensional feature matrix in the output result as the input of the CNN, and executing step S7; returning the one-dimensional feature vector in the output result as the input of the RNN, and executing step S8; the optimizer of the recognition model uses the "adam" optimizer, and the loss function uses "sparse categorical crossentropy".