A high-voltage equipment insulation performance testing method based on deep learning

Abstract

The application provides a high-voltage equipment insulation performance test method based on deep learning, and belongs to the technical field of cross between high-voltage electrical equipment detection and deep learning, and the method comprises the following steps: S1, collecting test data; S2, data preprocessing; S3, feature extraction; S4, feature standardization and label construction; S5, deep learning model training and optimization; S6, actual insulation performance test and result output. The deep learning model is used for automatically outputting equipment insulation states (normal, slight aging, moderate aging and severe aging) and related evaluation results, so as to provide an intelligent and accurate solution for high-voltage equipment insulation performance test.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of high-voltage electrical equipment detection and deep learning, and in particular to a high-voltage equipment insulation performance testing method based on deep learning. BACKGROUND

[0002] The insulation performance of high-voltage equipment is the core of ensuring the safe and stable operation of the power system, and the deterioration or defects of the insulation state will directly cause equipment failure and even large-scale power outages. Current high-voltage equipment insulation performance testing mainly relies on traditional detection methods, including single indicator testing such as partial discharge detection, dielectric loss measurement, leakage current monitoring, and data processing methods such as manual experience judgment and simple linear analysis.

[0003] These traditional methods have obvious limitations: on the one hand, the change process of the insulation state from normal to severe aging presents complex nonlinear characteristics, and there are complex local correlations and time series dynamic dependencies between characteristics such as partial discharge, dielectric loss, and leakage current, making it difficult for traditional linear models to accurately depict these deep features; on the other hand, existing testing methods lack the ability to synergistically integrate multi-source features, often analyzing single indicators in isolation, resulting in the loss of feature information and the inability to fully exploit the effective value of the data, which in turn leads to misjudgment and missed judgment of the insulation state, especially the inability to identify early minor aging defects. In addition, traditional methods are low in intelligence and rely on manual intervention, and the testing efficiency and reliability are greatly affected by human factors, making it difficult to meet the needs of large-scale and refined monitoring of the power system. SUMMARY

[0004] The present application provides a high-voltage equipment insulation performance testing method based on deep learning, which first collects core insulation parameters such as partial discharge, dielectric loss, and leakage current of high-voltage equipment through multi-source sensors, and completes data preprocessing through cleaning, missing value repair, and outlier processing; then extracts multi-dimensional features such as time domain, frequency domain, phase, and time series trend from the preprocessed data, forms a comprehensive feature vector, and performs standardization processing; the core part adopts a four-algorithm fusion architecture of separation before combination and bidirectional feedback, first extracts local correlation features by convolutional neural network and captures time series dynamic features by long short-term memory network, preliminarily combines the two to form basic features, then performs nonlinear optimization on the basic features by quantum surface fitting, and finally completes insulation state classification by quantum random forest, and dynamically adjusts the parameters of the previous algorithms through the feedback coefficient to realize the synergistic optimization of the four algorithms; after the model training is completed, the method is applied to newly collected insulation data, and automatically outputs the equipment insulation state (normal, minor aging, moderate aging, and severe aging) and related evaluation results, providing an intelligent and accurate solution for high-voltage equipment insulation performance testing.

[0005] To achieve the above-mentioned purposes, the present application adopts the following technical solutions:

[0006] A high-voltage equipment insulation performance test method based on deep learning, comprising:

[0007] Step 1: Collecting insulation parameter data of high-voltage equipment, including partial discharge parameters, dielectric loss parameters, and leakage current parameters;

[0008] Step 2: Preprocessing the collected insulation parameter data to obtain preprocessed data;

[0009] Step 3: Extracting multi-dimensional features from the preprocessed data to form a feature vector;

[0010] Step 4: Standardizing the feature vector to obtain a standardized feature matrix, and simultaneously constructing a label matrix based on the insulation state of the high-voltage equipment, dividing the standardized feature matrix and the label matrix into a training set, a validation set, and a test set;

[0011] Step 5: Constructing a deep learning model that integrates convolutional neural networks, long short-term memory networks, quantum curve fitting, and quantum random forests, training the model using the training set, dynamically adjusting model parameters using the validation set during training, and evaluating model performance using the test set;

[0012] Step 6: Applying the model that meets the evaluation standards to newly collected high-voltage equipment insulation parameter data to output insulation state classification results;

[0013] In step 5, the core logic of model training is as follows: extracting a local correlation feature vector from the standardized feature matrix using a convolutional neural network, processing the locally correlated feature vector sorted by timestamp using a long short-term memory network to extract a time-series dynamic feature vector, preliminarily fusing the local correlation feature vector and the time-series dynamic feature vector to obtain a preliminary fusion feature vector, performing nonlinear optimization on the preliminary fusion feature vector using quantum curve fitting to obtain an optimized feature vector, and classifying the optimized feature vector using quantum random forests to obtain insulation state classification results, while the mapping coefficient output by quantum curve fitting is used to adjust the parameters of the convolutional neural network, and the feedback coefficient output by quantum random forests is used to adjust the parameters of the long short-term memory network, forming a closed-loop optimization.

[0014] In this specification, the preprocessing in step 2 includes data cleaning, missing value processing, and outlier processing. The data cleaning uses a moving average filter method, the missing value processing uses linear interpolation, K-nearest neighbor interpolation, or re-collection of data according to the missing proportion, and the outlier processing uses the 3σ criterion for detection and replacement with normal range boundary values.

[0015] In the specification, the multi-dimensional features in step 3 include time-domain features, frequency-domain features and phase features of partial discharge, basic features, time-series trend features, working condition response features, correlation features and mutation features of dielectric loss, and basic features, time-series trend features, working condition response features, correlation features, mutation features and harmonic features of leakage current.

[0016] In the specification, the standardized processing in step 4 adopts a Z-score standardization method, and a label matrix is constructed in a one-hot encoding manner, corresponding to four insulation states of normal state, slight aging state, moderate aging state and severe aging state of the high-voltage equipment.

[0017] In the specification, the convolutional neural network adopts a progressive structure of input layer-2 convolutional layers-2 pooling layers-1 fully connected layer, the input layer receives the standardized feature matrix and reshapes it into a four-dimensional tensor, the convolutional layer calculates and captures local correlations between features, the pooling layer retains key local features and reduces dimensions, and the fully connected layer compresses high-dimensional features into a fixed-dimensional local correlation feature vector.

[0018] In the specification, the long short-term memory network adopts a structure of input layer-2 LSTM unit layers-1 fully connected layer, the input layer organizes the local correlation feature vector into a time-series sequence according to time stamps, the LSTM unit layer captures time-series dynamic dependencies through the synergistic effect of forget gate, input gate, cell state and output gate, and the fully connected layer compresses the output of the last time step into a fixed-dimensional time-series dynamic feature vector.

[0019] In the specification, the processing process of quantum surface fitting is as follows: each feature dimension of the preliminary fused feature vector is mapped to a quantum bit ground state, a nonlinear fitting surface is constructed based on the measurement expectation value of the quantum state, the preliminary fused feature vector is optimized to obtain an optimized feature vector, and the average value of the square of the quantum amplitude of all feature dimensions of all samples is calculated as a mapping coefficient.

[0020] In the specification, the quantum random forest is composed of multiple quantum decision trees, each quantum decision tree determines the sample splitting direction through quantum measurement principle, and the quantum random forest determines the final classification result through an integrated voting mechanism. In the voting process, the weight is dynamically adjusted according to the classification accuracy of each quantum decision tree, and the average value of the maximum voting score proportion of all samples is calculated as a feedback coefficient.

[0021] In the specification, the total loss of model training in step 5 is obtained by weighted summation of the cross-entropy loss of the convolutional neural network, the cross-entropy loss of the long short-term memory network, the mean square error loss of the quantum surface fitting, and the cross-entropy loss of the quantum random forest, wherein the loss weight of the quantum surface fitting and the quantum random forest is higher than the loss weight of the convolutional neural network and the long short-term memory network.

[0022] In the specification, the model evaluation indicators in step 5 include accuracy, precision, recall and F1 score. If the recall rate of a certain insulation state is lower than the preset value, the sample data of the state is supplemented to reconstruct the data set; if the overall precision is lower than the preset value, the feature extraction method is optimized; if the loss function converges slowly, the model learning rate is adjusted until the model evaluation meets the standard.

[0023] In summary, the present application has at least the following beneficial effects:

[0024] Deep mining of multi-dimensional features: through the synergistic effect of convolutional neural network and long short-term memory network, the local correlation information and time series dynamic trend of insulation features are comprehensively captured, avoiding information loss caused by single feature analysis, and making the feature representation more consistent with the essential characteristics of the insulation state.

[0025] Improve the discrimination ability of nonlinear features: with the nonlinear mapping advantage of quantum surface fitting, effectively describe the nonlinear change law in the insulation state deterioration process, solve the problem that traditional linear fusion cannot distinguish subtle state differences, and provide high recognition feature basis for accurate classification.

[0026] Enhance the reliability and robustness of classification decision: through the integrated voting mechanism of quantum random forest, combined with dynamically adjusted decision tree weight and split threshold, reduce the overfitting risk of single model, improve the recognition accuracy of different insulation states (especially early slight aging state), and reduce the probability of misjudgment and omission.

[0027] Form a closed-loop collaborative optimization mechanism: through bidirectional feedback interaction adjustment of the parameters of the previous algorithm, realize the dynamic synergy of convolutional neural network, long short-term memory network, quantum surface fitting and quantum random forest, so that the whole test model has self-adaptive optimization capability, and continuously adapts to the change characteristics of insulation data.

[0028] Improve the intelligentization and automation level of the test: without human intervention throughout the process, form a complete automatic process from data preprocessing, feature extraction to state classification, reduce the dependence on human experience, improve the test efficiency, and provide reliable and efficient technical support for preventive maintenance of high-voltage equipment. BRIEF DESCRIPTION OF DRAWINGS

[0029] Figure 1 The figure is a schematic diagram of the high-voltage equipment insulation performance test method based on deep learning involved in the present application.

[0030] Figure 2 The figure is a schematic diagram of the high-voltage equipment insulation performance test total process involved in the present application.

[0031] Figure 3 The figure is a schematic diagram of the deep learning model architecture synergy process involved in the present application.

[0032] Figure 4 This is a schematic diagram of the data processing flow involved in this invention. Detailed Implementation

[0033] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0034] like Figure 1 and Figure 2 As shown, this embodiment provides a deep learning-based method for testing the insulation performance of high-voltage equipment, including:

[0035] S1: Test Data Acquisition

[0036] 1.1 Determination of Acquisition Parameters

[0037] Based on the core influencing factors of high-voltage equipment insulation performance, three types of key parameters were collected:

[0038] Partial discharge parameters include partial discharge amplitude, pulse frequency, discharge phase distribution, and discharge energy.

[0039] Dielectric loss parameters include dielectric loss value (tanδ), dielectric loss rate of change (Δtanδ), and dielectric loss temperature coefficient (α_tanδ).

[0040] Leakage current parameters include the effective value of leakage current (I_leak), peak value (I_leak,peak), and waveform distortion rate (THD_leak).

[0041] 1.2 Data Acquisition Equipment and Deployment

[0042] Partial discharge acquisition equipment: It adopts ultra-high frequency partial discharge sensors (detection frequency band 300MHz-3GHz) and ultrasonic sensors (detection frequency band 20kHz-200kHz). The sensors are deployed in key parts such as bushings, flanges of high voltage equipment, and winding ends. Each device is equipped with 3 to 5 sensors.

[0043] Dielectric loss acquisition equipment: A high-precision dielectric loss tester (measurement accuracy ±0.0001) is used, which is connected to the equipment circuit through voltage transformers and current transformers to realize real-time acquisition of dielectric loss values;

[0044] Leakage current acquisition equipment: uses a Rogowski coil (measurement range 0-100mA, accuracy ±0.5%), connected in series in the equipment grounding circuit to avoid affecting the normal operation of the equipment;

[0045] Data synchronization equipment: A GPS synchronization module (synchronization accuracy ≤1μs) is used to ensure that the collection timestamps of the three types of parameters are completely consistent, laying the foundation for subsequent time series feature extraction.

[0046] 1.3 Data Acquisition Settings

[0047] Typical operating conditions of high-voltage equipment in actual operation:

[0048] Rated voltage operating condition: The rated operating voltage of the equipment (e.g., 110kV, 220kV, 500kV, determined according to the rated voltage of the test equipment) is continuously collected for 30 minutes;

[0049] 1.2 times rated voltage condition: Simulates an overvoltage scenario, continuously collecting data for 20 minutes;

[0050] 1.5 times rated voltage condition: Simulates a severe overvoltage scenario, continuously collecting data for 10 minutes;

[0051] Temperature change condition: Under rated voltage, the temperature changes sequentially from -20℃, 0℃, 25℃, 40℃, and 60℃. Data is collected after stabilizing at each temperature point for 15 minutes.

[0052] Humidity variation operating conditions: Under rated voltage and 25℃, the humidity changes sequentially from 30%RH, 50%RH, 70%RH and 90%RH. Data is collected after each humidity point has been running stably for 15 minutes.

[0053] 1.4 Acquisition Frequency and Data Storage

[0054] Acquisition frequency: The sampling frequency for partial discharge parameters is 1MHz (to ensure the capture of pulse details), and the sampling frequency for dielectric loss and leakage current parameters is 1kHz (to balance accuracy and storage capacity). A data slice is generated every 3 minutes according to the timestamp, and each slice contains data from all sampling points within that time period.

[0055] Data storage format: CSV format is used. Each data record contains the following fields: device number, acquisition timestamp, operating condition type, partial discharge amplitude, pulse frequency, discharge phase distribution, discharge energy, dielectric loss value, dielectric loss rate of change, dielectric loss temperature coefficient, effective value of leakage current, peak value of leakage current, and waveform distortion rate. The storage path is the raw data directory of the computing terminal, with folders named hierarchically according to device number and acquisition date.

[0056] 1.5 Data Collection Results

[0057] After the data collection is completed, the raw dataset is obtained. Its structure is Where N is the total number of data records, and 14 corresponds to the 14 fields mentioned above (excluding device number and collection timestamp, which are retained as an index in subsequent processing). The data processing flow is as follows: Figure 4 As shown.

[0058] S2: Data Preprocessing

[0059] 2.1 Data Cleaning

[0060] The moving average filtering method is used to remove high-frequency noise. The core idea is to use the mean in the local neighborhood to smooth the original data and retain the effective signal trend.

[0061] Define the moving average filter window size W for the original data. Each numerical feature in (j=1,2,...,12, corresponding to 12 core parameters), construct the filter function: ; is the cleaned value of the j-th feature of the i-th record; W is the size of the filtering window, which is 5 (through experiments, it has been verified that a window of 5 data points can effectively remove noise without destroying the local features of the signal). This is the floor function; when When <1, k starts from 1; when When the value is greater than N, the process continues from k to N (boundary handling); k is the index variable within the filtering window, used to traverse all data points within the window.

[0062] right The above filtering function is applied to each of the 12 numerical features to obtain the cleaned dataset. Its structure is The number of records N in the original data must be kept consistent to ensure data integrity.

[0063] 2.2 Missing Value Handling

[0064] 2.2.1 Missing Value Detection

[0065] Statistical methods were used to detect missing values ​​and calculate the missing proportion of each feature. : ;in Let be the number of missing records for the j-th feature.

[0066] 2.2.2 Missing Value Repair

[0067] when <5% (minor missing): Linear interpolation is used for repair. For the missing position i, the repair value is... for: If the missing position is the first record (i=1), then If it is the last record (i=N), then .

[0068] when (Moderate missing records): K-nearest neighbor interpolation was used for repair, selecting K=5 complete records most similar to the missing record and calculating the weighted average. ;in The index for the m-th nearest neighbor record. , Let m be the Euclidean distance between the missing record and its m-th nearest neighbor. .

[0069] when (Large amount of missing data): Remove the device data corresponding to this feature and re-collect the data for S1 (because the data for this feature is not reliable enough to support subsequent analysis).

[0070] After processing, a dataset with no missing values ​​is obtained. The number of records after removing a large amount of missing data. ; The original dataset after S1 data collection Total number of records; This represents the number of valid data records remaining after missing value processing (removing records with a large number of missing data).

[0071] 2.3 Outlier Handling

[0072] The 3σ criterion is used to detect and process outliers, and extreme values ​​that deviate from the normal range are identified based on the normal distribution characteristics of the data.

[0073] 2.3.1 Outlier Detection

[0074] right Each feature Calculate the mean and standard deviation : ; ;like Then determine This is an outlier.

[0075] 2.3.1 Outlier Repair

[0076] Replace outliers with boundary values ​​of the normal range for this feature: The preprocessed dataset is obtained after processing. The dataset will be passed to S3 for feature extraction.

[0077] S3: Feature Extraction

[0078] Based on preprocessing 256 feature parameters (corresponding to the 16×16 dimension of the CNN input feature map in S5) are extracted from three types of parameters: partial discharge, dielectric loss, and leakage current. This ensures comprehensive coverage of the local correlation and temporal trend information of the insulation state. It should be noted that the extraction or calculation of the following parameters can refer to existing technologies, and will not be elaborated here.

[0079] 3.1 Partial Discharge Feature Extraction (86 features)

[0080] 3.1.1 Time-domain features (40)

[0081] Amplitude-related parameters (9 parameters): maximum amplitude of partial discharge pulse, arithmetic mean of all pulse amplitudes, median of amplitude, standard deviation of amplitude, variance of amplitude, difference between maximum and minimum amplitude, interquartile range of amplitude, skewness of amplitude distribution, and kurtosis of amplitude distribution.

[0082] Frequency-related (14 items): total number of pulses detected within the acquisition time, average number of pulses per second, arithmetic mean of the number of pulses in each 3-minute acquisition slice, standard deviation of the number of pulses in each slice, skewness of the distribution of the number of pulses in each slice, kurtosis of the distribution of the number of pulses in each slice, after dividing the pulse amplitude into 5 intervals (0-0.2 times the maximum amplitude, 0.2-0.4 times the maximum amplitude, 0.4-0.6 times the maximum amplitude, 0.6-0.8 times the maximum amplitude, 0.8-1.0 times the maximum amplitude), the proportion of the number of pulses in each interval to the total number of pulses, the average number of pulses per minute, the acquisition time corresponding to when the cumulative number of pulses reaches 90% of the total number of pulses, and the number of abrupt changes in the number of pulses exceeding 2 times the standard deviation;

[0083] Energy-related (10): total energy of all partial discharge pulses, average discharge energy per second, arithmetic mean of single pulse energy, standard deviation of single pulse energy, skewness of pulse energy distribution, kurtosis of pulse energy distribution, and the proportion of pulse energy to total energy in each of the four equidistant intervals after dividing the total energy.

[0084] Time-related factors (7): the ratio of the amplitude difference between two adjacent pulses to the time interval, the maximum value of the ratio, the minimum value of the ratio, the arithmetic mean of the ratio, the standard deviation of the ratio, the arithmetic mean of the time intervals between adjacent pulses, and the coefficient of variation of the time intervals between adjacent pulses (the ratio of the standard deviation to the mean).

[0085] 3.1.2 Frequency Domain Features (30)

[0086] After performing a Fast Fourier Transform on the partial discharge amplitude sequence, the following features are extracted:

[0087] Basic frequency domain parameters (6): maximum amplitude in the frequency domain spectrum, frequency value corresponding to the maximum amplitude, mean frequency of the frequency domain amplitude, centroid frequency of the frequency domain amplitude, root mean square frequency of the frequency domain amplitude, and total energy of the frequency domain spectrum;

[0088] Frequency band energy percentage (5 bands): After dividing the frequency into 5 bands (0-10kHz, 10-50kHz, 50-100kHz, 100-500kHz, 500kHz-1MHz), the proportion of energy in each band to the total frequency domain energy.

[0089] Frequency domain distribution statistics (4): skewness, kurtosis, standard deviation, and interquartile range of frequency domain amplitude distribution;

[0090] Harmonic characteristics (15): The proportion of the amplitude of the first 15 harmonics in the frequency domain spectrum to the amplitude of the fundamental frequency.

[0091] 3.1.3 Phase characteristics (16)

[0092] Based on the phase distribution data of partial discharge pulses, the following features are extracted:

[0093] The core parameters of the phase distribution (4): the maximum peak value of the phase distribution curve, the phase position corresponding to the maximum peak value, the mean of the phase distribution, and the standard deviation of the phase distribution;

[0094] Quadrant distribution characteristics (8): Calculate the proportion of pulses in the four quadrants 0°-90°, 90°-180°, 180°-270°, and 270°-360° to the total number of pulses, and the proportion of discharge energy in the four quadrants to the total energy;

[0095] Phase distribution statistics (4 items): skewness, kurtosis, variance, and interquartile range of phase distribution.

[0096] 3.2 Extraction of dielectric loss features (85 features)

[0097] 3.2.1 Basic Features (30)

[0098] Related to dielectric loss values ​​(15): arithmetic mean, standard deviation, variance, skewness, kurtosis, maximum value, minimum value, median, interquartile range, mode of dielectric loss values ​​(the most frequently occurring dielectric loss value), and the percentage of data points in each interval after dividing the dielectric loss value into 5 intervals (0-0.001, 0.001-0.005, 0.005-0.01, 0.01-0.05, >0.05).

[0099] Related to the rate of change of dielectric loss (7 items): arithmetic mean, standard deviation, variance, maximum value, minimum value, skewness of the rate of change distribution, and kurtosis of the rate of change distribution of the absolute values ​​of the difference between adjacent dielectric loss values;

[0100] Temperature coefficient correlation (8): arithmetic mean, standard deviation, maximum value, minimum value of all temperature coefficients, and arithmetic mean of temperature coefficients within the set temperature range (0℃, 25℃, 40℃, 60℃) (a total of 4 range means).

[0101] 3.2.2 Time-series and operating condition correlation features (55 features)

[0102] Time-series trend characteristics (15): Using 5 data points as a sliding window, calculate the moving mean, moving standard deviation, and moving range of dielectric loss values ​​at 1min, 5min, and 10min time dimensions (5 each, 15 in total).

[0103] Operating condition response characteristics (28):

[0104] Voltage-related parameters (6): Difference between dielectric loss value at 1.2 times the rated voltage, Difference between dielectric loss value at 1.5 times the rated voltage, Difference between dielectric loss rate of change at 1.2 times the rated voltage, Difference between dielectric loss rate of change at 1.5 times the rated voltage, Standard deviation of dielectric loss value at 1.2 times the rated voltage, Standard deviation of dielectric loss value at 1.5 times the rated voltage;

[0105] Temperature-related factors (12): Slope and intercept of linear fitting between dielectric loss value and temperature in 4 temperature ranges (0℃, 25℃, 40℃, 60℃) (8 in total), Difference between dielectric loss value at 25℃ (rated temperature) and dielectric loss value in the other 3 temperature ranges (3 in total), and mean value of dielectric loss value in the 25℃ range (1 in total).

[0106] Humidity-related factors (10): Slope and intercept of linear fitting between dielectric loss value and humidity in 4 humidity ranges (30%RH, 50%RH, 70%RH, 90%RH) (8 in total), Difference between dielectric loss value at 50%RH (rated humidity) and at 90%RH (1), Overall standard deviation of dielectric loss value under humidity conditions (1).

[0107] Correlation characteristics (3): linear correlation coefficient between dielectric loss value sequence and corresponding temperature sequence, linear correlation coefficient between dielectric loss value sequence and corresponding humidity sequence, and linear correlation coefficient between dielectric loss change rate sequence and corresponding voltage sequence;

[0108] Mutation characteristics (9): The first 3 main mutation points are selected, and each mutation point corresponds to 3 types of characteristics: the specific acquisition time of the mutation, the change amplitude of dielectric loss value before and after the mutation, the time required for the dielectric loss value to stabilize after the mutation, the ambient temperature when the mutation occurs, the ambient humidity when the mutation occurs, the operating voltage level when the mutation occurs, the stable average value of dielectric loss value after the mutation, the maximum value of dielectric loss change rate before and after the mutation, and the duration of the mutation (the time from the occurrence of the mutation to the start of stabilization).

[0109] 3.3 Leakage current feature extraction (85 features)

[0110] 3.3.1 Basic Features (30)

[0111] RMS correlation (15 items): arithmetic mean, standard deviation, variance, skewness, kurtosis, maximum value, minimum value, median, interquartile range, mode of RMS leakage current, and percentage of data points in each of the 5 intervals (0-1mA, 1-5mA, 5-10mA, 10-50mA, >50mA) after dividing the RMS value into 5 intervals (0-1mA, 1-5mA, 5-10mA, 10-50mA, >50mA) (5 interval percentages, totaling 15 items);

[0112] Peak correlation (7): arithmetic mean, standard deviation, variance, skewness, kurtosis, maximum value, and minimum value of all leakage current peak values;

[0113] Waveform distortion rate related (8 items): arithmetic mean, standard deviation, maximum value, minimum value of all waveform distortion rates, percentage of data points in each interval after dividing the distortion rate into 3 intervals (0-5%, 5%-10%, >10%) (3 items), mean of distortion rate of all intervals (1 item).

[0114] 3.3.2 Time-series and operating condition correlation features (55 features)

[0115] Time series trend characteristics (15): Using 5 data points as a sliding window, calculate the moving mean, moving standard deviation, and moving range of the effective value of leakage current in each window (5 each, 15 in total).

[0116] Operating condition response characteristics (28):

[0117] Voltage-related parameters (6): 1.2 times the difference between the effective value of leakage current and the rated voltage, 1.5 times the difference between the effective value of leakage current and the rated voltage, 1.2 times the difference between the peak value of leakage current and the rated voltage, 1.5 times the difference between the peak value of leakage current and the rated voltage, 1.2 times the standard deviation of the effective value of leakage current at the rated voltage, and 1.5 times the standard deviation of the peak value of leakage current at the rated voltage.

[0118] Temperature-related factors (12): Slope and intercept of the linear fit between the effective value of leakage current and temperature in 4 temperature ranges (0℃, 25℃, 40℃, 60℃) (8 in total), difference between the effective value of leakage current at 25℃ and the other 3 temperature ranges (3 in total), and mean value of the effective value of leakage current in the 25℃ range (1 in total).

[0119] Humidity-related factors (10): The slope and intercept of the linear fit between the effective value of leakage current and humidity in 4 humidity ranges (30%RH, 50%RH, 70%RH, 90%RH) (8 in total), the difference between the effective value of leakage current at 50%RH and 90%RH (1), and the overall standard deviation of the effective value of leakage current under humidity conditions (1).

[0120] Correlation characteristics (3): linear correlation coefficient between the effective value sequence of leakage current and the corresponding voltage sequence, linear correlation coefficient with the corresponding temperature sequence, and linear correlation coefficient with the corresponding humidity sequence;

[0121] Mutation characteristics (9): The first 3 main mutation points are selected, and each mutation point corresponds to 3 types of characteristics: the specific collection time of the mutation, the change amplitude of the effective value of leakage current before and after the mutation, the time required for the effective value of leakage current to stabilize after the mutation, the ambient temperature when the mutation occurs, the ambient humidity when the mutation occurs, the operating voltage level when the mutation occurs, the stable average value of the effective value of leakage current after the mutation, the maximum value of the rate of change of leakage current before and after the mutation, and the duration of the mutation (the time from the occurrence of the mutation to the start of stabilization).

[0122] 3.4 Feature Extraction Results

[0123] The three types of features—partial discharge (86 features), dielectric loss (85 features), and leakage current (85 features)—are combined in a fixed order: partial discharge features → dielectric loss features → leakage current features, forming a feature vector with a unified format. Each row of this feature vector corresponds to one preprocessed original data record, and each column corresponds to one extracted feature, serving as input data for subsequent feature standardization processing.

[0124] 3.4 Feature Extraction Results

[0125] The above 86+85+85=256 features are combined in the order of partial discharge features → dielectric loss features → leakage current features to form a feature vector. Each row corresponds to a preprocessed original data record, and each column corresponds to an extracted feature.

[0126] S4: Feature Standardization and Label Construction

[0127] 4.1 Feature Standardization

[0128] The Z-score standardization method is used to eliminate the dimensional differences between different features, ensuring that the weights of each feature are balanced during model training.

[0129] 4.1.1 Standardized Formula

[0130] For eigenvectors Each feature in (k=1,2,...,256), standardized values for: ; (The mean of the kth feature); (Standard deviation of the k-th feature); This is the original value of the k-th feature in the i-th record.

[0131] 4.1.2 Standardization Process and Results

[0132] right The above formula is applied to each of the 256 features to obtain the standardized feature matrix. After standardization, each feature has a mean of 0 and a standard deviation of 1, ensuring a consistent data distribution.

[0133] 4.2 Tag Construction

[0134] Based on the actual state of insulation performance of high-voltage equipment, four types of labels are constructed, corresponding to different levels of insulation status. The label data comes from the results of manual inspection of the equipment and historical operation records (manual inspection uses traditional methods such as ultrasonic and infrared imaging, combined with the equipment's operating years and maintenance records to determine the insulation status).

[0135] 4.2.1 Tag Definition

[0136] Label 1 (Normal Condition): The equipment has good insulation performance, and the partial discharge, dielectric loss, and leakage current are all within the standard range, with no signs of aging;

[0137] Label 2 (Slight Aging): The equipment insulation shows slight aging, the frequency of partial discharge pulses has increased slightly, the dielectric loss value slightly exceeds the lower limit of the standard, and the leakage current is not obviously abnormal.

[0138] Label 3 (Moderate Aging): The equipment insulation is moderately aged, the amplitude and frequency of partial discharge have increased significantly, the dielectric loss value exceeds the standard range, and the effective value of leakage current has increased slightly.

[0139] Label 4 (Severe Aging Condition): The equipment insulation is severely aged with obvious defects, high partial discharge energy, dielectric loss value significantly exceeds the standard, and leakage current peak value has increased significantly, requiring immediate repair.

[0140] 4.2.2 Tag Matrix Construction

[0141] The tag matrix is ​​constructed using one-hot encoding. ,in:

[0142] If the insulation status of the device corresponding to the i-th record is normal (label 1), then =[1,0,0,0];

[0143] If it is in a state of slight aging (label 2), then =[0,1,0,0];

[0144] If it is in a moderate aging state (label 3), then =[0,0,1,0];

[0145] If it is in a severely aged state (label 4), then =[0,0,0,1].

[0146] 4.2.3 Dataset Partitioning

[0147] Standardize the feature matrix With the label matrix The dataset is divided into training, validation, and test sets in a 7:2:1 ratio.

[0148] Training set: Training set feature matrix Training set label matrix Number of training set samples ;

[0149] Validation set: Validation set feature matrix Validation set label matrix Number of validation set samples ;

[0150] Test set: Test set feature matrix Test set label matrix Number of test set samples .

[0151] After the partitioning, the training set is used for parameter training of the model in S5, the validation set is used for hyperparameter adjustment and early stopping judgment during the training process, and the test set is used for performance evaluation after the model training is completed.

[0152] S5: Deep Learning Model Training and Optimization

[0153] 5.1 Overall Model Architecture Design: This step focuses on constructing an algorithm fusion system based on the essential characteristics of high-voltage equipment insulation performance. The core logic is separation followed by fusion and bidirectional feedback. From the perspective of insulation data characteristics, the three features—partial discharge, dielectric loss, and leakage current—include both local correlation information of different parts of the equipment (such as the correlation between the frequency domain peak value of the partial discharge signal and the location of insulation defects) and temporal dynamic trends that change with test time (such as the slow increase of leakage current under long-term voltage application). Therefore, firstly, local correlation features are extracted using a convolutional neural network, and temporal dynamic features are captured using a long short-term memory network. These two are initially fused to form basic features. Then, the nonlinear mapping capability of quantum surface fitting is used to optimize the discriminative power of the basic features. This is because the change in insulation state from normal to severe aging is nonlinear, and traditional linear fusion is insufficient to characterize such subtle differences. Finally, a quantum random forest completes the classification decision, and the parameters of the preceding algorithms are adjusted through a feedback mechanism, forming a closed loop of extraction-optimization-classification-adjustment. This ensures that the four algorithms have no independent links and work collaboratively throughout the process through parameter transfer and interaction. The deep learning model architecture collaborative process is as follows: Figure 3 As shown.

[0154] 5.2 The process of building, training and applying each algorithm model:

[0155] 5.2.1 Convolutional Neural Network (CNN): The structure of the convolutional neural network is perfectly adapted to the local correlation characteristics of insulation features. It adopts a progressive structure of input layer - 2 convolutional layers - 2 pooling layers - 1 fully connected layer. The parameter settings of each layer are aimed at maximizing the extraction of local insulation features.

[0156] Input layer: Receives the S4-normalized feature vectors and reshapes them into a four-dimensional tensor form. Its dimension is represented as Where B is the batch size, which can be either 32 or 64. These two values ​​are chosen to ensure both training stability and computational efficiency. H is the feature height and W is the feature width, both set to 16. This setting corresponds to the 256 feature parameters (16×16) extracted by S3, ensuring that the feature vector can be completely mapped to a two-dimensional feature map. C is the number of feature channels, set to 3, which correspond to the three core insulation features of partial discharge, dielectric loss, and leakage current, respectively, to realize independent input and subsequent fusion of different types of features.

[0157] Convolutional layers: Two convolutional layers are used to progressively deepen the extraction of local features. The first convolutional layer uses... The second convolutional layer uses one convolutional kernel. There are 1 convolutional kernel, and the kernel size is uniform. and The 3×3 size effectively covers the correlation between adjacent features on the feature map (such as the correlation between partial discharge amplitude and pulse frequency) while avoiding overfitting due to excessive parameters. The convolution calculation is performed using the following formula: In the formula This is the feature map output by the l-th convolutional layer, where l can be either 1 or 2. For the activation function, the ReLU function is chosen, and its expression is: This function can effectively alleviate the gradient vanishing problem and adapt to the nonlinear extraction requirements of insulation features; This is the input to the l-th convolutional layer. When l=1 That is When l=2 The output of the first pooling layer ; * represents a convolution operation, used to calculate the local correlation between the convolution kernel and the input feature map; The weights of the convolution kernel of the l-th convolutional layer, and the weights of the first convolutional layer. Dimensions The second convolutional layer Dimensions ; The bias term of the l-th convolutional layer, the first convolutional layer Dimensions The second convolutional layer Dimensions ; The value is 32. The value is set to 64. This incremental setting is to allow subsequent convolutional layers to extract more complex local correlation features, while controlling the total number of parameters to avoid model redundancy.

[0158] Pooling layers: Alternating with convolutional layers, using max pooling with a 2×2 kernel size and a stride of 2. The core function of max pooling is to reduce the dimensionality of the feature map while preserving key local features (such as peak features of partial discharge signals and abrupt changes in dielectric loss data), avoiding irrelevant information from interfering with subsequent calculations. Pooling calculations are performed using the following formula: In the formula This is the feature map output by the l-th pooling layer, where l takes the value of 1 or 2; This represents the max pooling function, which takes the maximum value of all elements within a 2×2 window on the input feature map and uses it as the element at the corresponding position in the output feature map.

[0159] Fully connected layer: Flattens the feature map output from the second pooling layer into a one-dimensional vector. , The dimension is ,in The feature dimension after flattening is calculated as follows: Substitute After =64, H=16, W=16, =64 x 4 x 4 = 1024. The role of the fully connected layer is to compress high-dimensional local features into a fixed-dimensional vector so that it can be fused with the output features of the subsequent Long Short-Term Memory network. Its calculation is achieved through the following formula: In the formula This is the local feature vector output by the convolutional neural network, with dimension 1. , The value is 128. This dimension was chosen to balance the feature representation capability with the computational efficiency of subsequent algorithms. Represents matrix multiplication; Here is the weight matrix of the fully connected layer, with dimension 1. ; The bias term for the fully connected layer has a dimension of .

[0160] Model training process: The training of the convolutional neural network aims to minimize the deviation between local features and insulation state labels, and is optimized step by step in three stages: initialization, loss calculation, and parameter update.

[0161] Initialization: Select an appropriate initialization method for different parameters. Convolution kernel weights. With fully connected layer weights We initialize using a He normal distribution, the expression of which is: ,in The input dimension of the weight matrix is ​​used; this method is chosen because the ReLU activation function effectively avoids gradient vanishing under He normal distribution initialization; the bias term... and All vectors are initialized to zero to ensure that no additional bias is introduced during the initialization phase.

[0162] Loss calculation: temporarily using local feature vectors With insulation status label The cross-entropy loss is used as the training loss, denoted as . .in The label matrix constructed in S4 has dimensions of 4 corresponds to the four states of insulation performance of high-voltage equipment (normal state, slight aging state, moderate aging state, and severe aging state), in the matrix. A value of 1 indicates that the i-th sample belongs to the c-th class, while a value of 0 indicates that it does not. Cross-entropy loss effectively measures the difference between the predicted probability and the true label, and its calculation is achieved using the following formula: In the formula Let be the local feature vector of the i-th sample; The softmax function transforms a feature vector into a probability distribution. Its calculation expression for the c-th class is as follows: ; The function is the natural logarithm; the outer summation iterates through all samples and all state categories, and finally divides by the number of samples B to obtain the average loss.

[0163] Parameter Update: The Adam optimizer is used to iteratively update the parameters. The Adam optimizer combines the advantages of momentum and adaptive learning rate, which can stably accelerate training convergence. Setting the learning rate... =0.001, this value is the optimal value verified through multiple pre-training iterations, which ensures both convergence speed and avoids parameter oscillations caused by an excessively large learning rate. Parameter updates are achieved through the following formula: ; In the formula and These are the weights and biases at the t-th training round; and These are the weights and biases for the (t+1)th round after the update; For loss gradient with respect to weights, For loss The gradient of the bias term is calculated using the backpropagation algorithm.

[0164] Model application process: The S4-standardized feature vectors of the training, validation, and test sets are input into the constructed convolutional neural network, and then processed sequentially through convolution, pooling, and fully connected layers to output the corresponding local feature vector for each sample. Because subsequent long short-term memory networks need to extract dynamic features based on temporal information, The samples will be sorted according to the order of their test timestamps to ensure that the local features of each sample can be associated with the corresponding test time, providing a foundation for the temporal analysis of Long Short-Term Memory networks; simultaneously, It will also serve as a source of basic features for quantum surface fitting, providing original local correlation information for subsequent nonlinear optimization.

[0165] 5.2.2 Long Short-Term Memory Network (LSTM): The Long Short-Term Memory Network is designed for the temporal characteristics of insulation data. It adopts an input layer-2 LSTM unit layers-1 fully connected layer structure. Through a unique gating mechanism, it solves the gradient vanishing problem of traditional recurrent neural networks and accurately captures the dynamic changes of insulation features over time.

[0166] Input layer: Receives local feature vectors from the output of the convolutional neural network. And reorganized into a time sequence according to the order of the test timestamps. . The dimension is Where T is the time step, with a value of 10. This setting corresponds to 10 consecutive test time points (e.g., data is collected every 3 minutes, for a total test duration of 30 minutes), ensuring that the short-term dynamic trends of insulation characteristics can be covered. =128, and Maintaining consistent dimensions ensures that time-series sequences can fully inherit information from local features.

[0167] LSTM Unit Layer: Two LSTM unit layers are set to deepen temporal feature extraction. The number of LSTM units in the first layer is... =64, the number of units in the second LSTM layer is =128, the increasing number of cells is to allow subsequent layers to capture more complex timing dependencies (such as the difference in leakage current under different voltage conditions). Each LSTM cell contains four core components: forget gate, input gate, cell state, and output gate. These components work together to selectively retain and update timing information. The specific calculations are implemented using the following series of formulas:

[0168] 1. Forget Gate: Controls the discarding of useless information from the previous cell state; its output determines the proportion of the previous cell state retained. The calculation expression is: In the formula This is the forget gate output vector at time step t, with the same dimension as the current number of LSTM units; Let t be the input time series sequence segment at time step t; This is the output vector of the LSTM unit at time step t-1; The weight matrix from the input sequence to the forget gate has the following dimensions in the first layer of the LSTM: In the second layer of the LSTM, its dimension is ; The recursive weight matrix output to the forget gate in the previous time step has the following dimensions in the first layer of the LSTM: In the second layer of the LSTM, its dimension is ; This is the bias term for the forget gate, with the same dimension as the current number of LSTM units.

[0169] 2. Input Gate: Controls the update ratio of the input information at the current moment, consisting of two steps: first, the sigmoid function determines the position of the information to be updated, and then the tanh function generates candidate update information. The calculation expression is: ; In the formula This is the input gate output vector at time step t, with the same dimension as the current number of LSTM units; Let be the candidate cell state vector at time step t, with dimensions consistent with the current number of LSTM units; The hyperbolic tangent activation function is expressed as follows: Its output range is [-1, 1], which can effectively alleviate the gradient problem; , These are the weight matrices from the input sequence to the input gate and the candidate cell state, respectively. , These are the recursive weight matrices for the output to the input gate and the candidate cell state at the previous time step, respectively. , These are the bias terms for the input gate and the candidate cell state, respectively, and their dimensions are adapted to the current number of LSTM units.

[0170] 3. Cell State: The core memory component of the LSTM unit, integrating the discard decisions of the forget gate and the update information of the input gate to achieve the updating of long-term temporal information. The calculation expression is: In the formula Let be the cell state vector at time step t. Let be the cell state vector at time step t-1; This is element-wise multiplication, which means multiplying elements at corresponding positions. This operation enables the selective retention and updating of different information.

[0171] 4. Output Gate: Controls the output of useful information from the cell state, determining the temporal characteristics output at the current moment. The calculation expression is: ; In the formula This is the output vector of the output gate at time step t, with the dimension consistent with the current number of LSTM units; This is the output vector of the LSTM unit at time step t, which is the temporal feature vector at that time step. The weight matrix from the input sequence to the output gate. This is the recursive weight matrix output to the output gate in the previous time step. These are the bias terms for the output gates, and their dimensions are adapted to the current number of LSTM cells.

[0172] Fully connected layer: Since the LSTM unit layer outputs the temporal features of each time step, and subsequent fusion requires a fixed-dimensional vector, a fully connected layer is set to process the output of the last time step. The output of the second-layer LSTM unit at the last time step (t=T) is taken. , The dimension is Compressed into a form through a fully connected layer Consistent temporal feature vectors , The dimension is Fully connected computation is achieved using the following formula: In the formula Here is the weight matrix of the fully connected layer, with dimension 1. ; The bias term for the fully connected layer has a dimension of .

[0173] Model training process: The training of the Long Short-Term Memory Network aims to optimize the fusion effect of temporal features and local features, and the parameters are adjusted based on the deviation between the initial fused features and the labels.

[0174] Initialization: Select an appropriate initialization method for different parameters to avoid gradient problems. Input weight matrix. , , , With recursive weight matrix , , , Orthogonal initialization is used for all components. This method ensures that the column vectors of the initialized weight matrix are orthogonal, effectively mitigating gradient vanishing during time-series training; bias term , , , All are initialized to zero vectors; fully connected layer weights Initialization using a He normal distribution, bias term Initialize as a vector of zero.

[0175] Preliminary fusion and loss calculation: Since individual temporal features or local features cannot fully characterize the insulation state, we first fuse... and Initial fusion is carried out, and initial fusion characteristics are formed. The fusion method uses element-wise averaging, which ensures a balance of information between the two features while avoiding the introduction of too many parameters. Its calculation is achieved through the following formula: In the formula To initially fuse feature vectors, the dimension is... .by With tags The cross-entropy loss is used as the training loss of the Long Short-Term Memory network, denoted as... Its computational logic is the same as The process is consistent, except that the local feature vectors are replaced with the preliminary fused feature vectors. The specific expression is as follows: In the formula Let be the initial fusion feature vector of the i-th sample.

[0176] Parameter update: The Adam optimizer is used for parameter updates, and the learning rate is set. =0.001, this learning rate is consistent with the convolutional neural network to ensure coordinated training pace. The parameter update expression is the same as the convolutional neural network, only the learning rate and loss function are replaced with... and Specifically: ; In the formula For loss gradient with respect to weights, For loss The gradient with respect to the bias term.

[0177] Model application process: First, receive the data sorted by timestamp. Organize it into a time series The data is then input into an LSTM unit layer, where a series of calculations using a gating mechanism are performed to obtain the temporal features at each time step. Finally, a fully connected layer outputs the temporal feature vector. .Will and Fusion Then, directly The parameters are passed to the quantum surface fitting, providing a fusion feature basis for nonlinear optimization. At the same time, the long short-term memory network receives the parameters fed back from the subsequent quantum random forest and adjusts its own gating weights. This interaction can dynamically optimize the temporal feature extraction effect according to the classification results. The specific interaction process will be explained in detail in Section 2.4.

[0178] 5.2.3 Quantum Surface Fitting (QSF): Quantum surface fitting is designed based on the principle of quantum superposition. The core is to construct a nonlinear mapping function to solve the initial fusion characteristics. The problem of being unable to characterize the nonlinear changes in insulation state needs to be addressed by improving the ability of features to distinguish between different insulation states.

[0179] Definition of quantum state: Each feature dimension is mapped to the ground state of a qubit, realizing the conversion from classical features to quantum states. For The d-th dimension feature (d ranges from 1 to...) The corresponding ground state of the qubit is denoted as The expression is .in and These are the two fundamental states of a quantum bit. and For quantum amplitude, the normalization condition must be satisfied. =1; and The value is determined by The d-th eigenvalue Decision, due to After standardization, the range is [0,1], therefore set , This setting ensures that the quantum amplitude always satisfies the normalization condition, while achieving a smooth mapping from classical eigenvalues ​​to quantum states.

[0180] Fitting Surface Function: Construct a nonlinear fitting surface to optimize the features after quantum state mapping, and output the optimized feature vector. The fitted surface function is designed based on the measurement expectation value of quantum states, which can fully utilize the superposition property of quantum states to characterize nonlinear relationships. Its calculation expression is: In the formula This is the optimized eigenvector output by quantum surface fitting. , dimension ; and quantum states In the ground state and The measurement probability on the d-th dimension reflects the quantum state distribution of the d-th dimension feature; and These are the ground states. and The weights on the d-th dimension feature together constitute the core parameter matrix for quantum surface fitting. , The dimension is ; Let be a unit vector with the d-th dimension, meaning that the d-th dimension of this vector has a value of 1, and the other dimensions have values ​​of 0. Ensure that the optimization result for each feature dimension corresponds to the correct position in the output vector.

[0181] Optimize the objective function: Define the fitting error as the training objective, ensuring... To get as close as possible to the ideal characteristics Ideal characteristics By tag Reverse generation, the generation rule is: if the sample belongs to the c-th type of insulating state (c=1,2,3,4), then The feature values ​​of the c×32 to (c+1)×32 dimensions are set to 1, and the remaining dimensions are set to 0. This rule allows the ideal features to form a strong correlation with the insulation state category, which facilitates subsequent fitting and optimization. The fitting error is denoted as... The mean squared error calculation method effectively measures the difference between the optimized features and the ideal features. Its expression is: In the formula For the i-th sample The d-th eigenvalue; For the i-th sample The d-th eigenvalue; the outer summation iterates through all samples and all feature dimensions, and finally divides by . The mean square error is obtained.

[0182] Model training process: Training the quantum surface fitting to minimize the fitting error. To achieve this, the core weight parameters are optimized through backpropagation. .

[0183] Initialization: To ensure that the fitted surface maintains its linear mapping characteristics in the initial stage, facilitating subsequent nonlinear optimization, the core weights are... Initialize as a uniformly distributed vector, i.e. The value of d ranges from 1 to 1. This initialization method allows the initial fitting result to be... , This means that the initial optimization features are fixed values, which makes it easier to observe the subsequent optimization trend.

[0184] Error backpropagation: Calculating the fitting error using the chain rule. For core weights The gradient provides a basis for parameter updates. For and The gradients are calculated as follows:

[0185] for Its gradient The expression is: ;for Its gradient The expression is: In the formula The quantum amplitude of the d-th dimension feature of the i-th sample The square of, The quantum amplitude of the d-th dimension feature of the i-th sample The square of the gradient; the coefficient 2 in the gradient expression is determined by the derivative property of the mean square error, divided by... This is to obtain the average gradient and avoid the influence of the number of samples and feature dimensions on the gradient magnitude.

[0186] Parameter Update: The Adam optimizer is used for parameter updates. Considering that quantum surface fitting is more sensitive to nonlinear characteristics, the learning rate is set accordingly. =0.0005, this learning rate is lower than that of CNN and LSTM, which can avoid unstable fitting due to excessively rapid weight updates. The expression for parameter update is: In the formula, * represents 0 or 1, corresponding to respectively and ; The weights are for the t-th round of training. These are the weights updated during the (t+1)th training round.

[0187] Model application process and interaction with CNN: Quantum surface fitting simultaneously achieves feature output and parameter feedback, forming a two-way interaction with the convolutional neural network. First, it receives preliminary fused features. The quantum state is converted into the ground state of the qubit through quantum state mapping, and then the optimized features are calculated and output by fitting the surface function. ,Will The data is directly passed to the quantum random forest, providing a highly discriminative feature base for classification decisions; simultaneously, quantum surface fitting calculates mapping coefficients. It is used to provide feedback to adjust the parameters of the convolutional neural network.

[0188] The calculation method is based on all feature dimensions of all samples. The average value is expressed as: .because ,when When it is small (close to 0) Larger, therefore The larger, the more it means The higher the proportion of low and medium feature values, the weaker the feature discrimination. Based on this characteristic, The convolutional kernel weights of the CNN are adjusted as feedback parameters, and the specific interaction formula is as follows: In the formula To adjust the kernel weights of the l-th convolutional layer in the CNN after round t+1, This is to adjust the weights in the previous t-th round; 0.1 is a fixed interaction coefficient used to control the adjustment range and avoid excessive weight changes that could lead to instability in CNN training. The core function of this interaction is: when... When the feature discrimination is weak, increasing the weights of the CNN convolution kernel enhances the CNN's ability to extract locally related features, thereby improving its performance. The high degree of differentiation provides better local features for subsequent fusion.

[0189] 5.2.4 Quantum Random Forest (QRF): Quantum Random Forest is based on quantum decision tree ensemble design. It uses the randomness of quantum measurement to improve classification robustness, and at the same time reduces the risk of overfitting of a single decision tree through an ensemble voting mechanism, ultimately achieving accurate classification of insulation states.

[0190] Quantum decision tree construction: Each quantum decision tree adopts a tree structure with a fixed number of levels, and the number of levels is set. =5, this setting balances classification accuracy and computational efficiency. Too few layers will lead to insufficient classification ability, while too many layers will increase the risk of overfitting. Each node in the tree structure corresponds to... One feature dimension d (d ranges from 1 to...) Each node defines a quantum splitting threshold. Where l is the number of levels in the tree (ranging from 1 to 1). ), used to determine the splitting direction of the sample at that node.

[0191] The splitting rule is designed based on the principle of quantum measurement. The specific process is as follows: for the d-th eigenvalue of the sample... First, map it into a quantum state. The mapping method is consistent with the quantum surface fitting, that is Then measure the quantum state in the ground state. The probability of , The expression is Finally, a comparison. With quantum splitting threshold ,like If the sample splits from the current node to the left child node, it splits to the right child node. This splitting continues layer by layer until the leaf node of the tree is reached. Each leaf node corresponds to one of the four insulating states, which is the classification result of the sample under the current quantum decision tree.

[0192] Integrated voting mechanism: To improve classification robustness, multiple quantum decision trees are constructed to form an integrated model, and the number of decision trees is set. =100, this number is the optimal value verified through multiple experiments. Too few trees will result in insignificant ensemble effects, while too many trees will significantly increase the computational load. Each quantum decision tree outputs a classification result for a sample. Where m is the decision tree index (ranging from 1 to 1). ), The values ​​are 1, 2, 3, and 4, which correspond to normal state, slight aging state, moderate aging state, and severe aging state, respectively.

[0193] The final classification result of the quantum random forest is determined through a voting mechanism, first assigning weights to each decision tree. Then, the voting score for each insulation state is calculated, and finally, the category with the highest voting score is selected as the final classification result. The specific calculation is achieved through the following series of formulas:

[0194] 1. Voting score calculation: For insulation state c (c=1,2,3,4), its voting score is calculated as follows. The expression is: In the formula This is an indicator function; when the condition within the parentheses is met (i.e., the classification result of the m-th decision tree is class c), The value is 1 if it is set to 1, otherwise the value is 0. Let be the weight of the m-th decision tree, with an initial value of . The results will be dynamically adjusted based on the classification accuracy of the decision tree.

[0195] 2. Final classification result determination: The category with the highest voting score is selected as the final classification result for the sample. Its expression is: In the formula The function that returns the voting score is the function that takes the maximum value of the input variable. The maximum value of category c.

[0196] Model training process: The training of quantum random forest aims to optimize classification accuracy, and is divided into three stages: initialization, loss calculation, and parameter update, adjusting the split threshold and decision tree weights.

[0197] Initialization: Quantum splitting threshold Initialized to 0.5, this value represents the quantum measurement probability. The median value ( The range is [0,1], which ensures a balanced distribution of samples at the nodes in the initial stage; decision tree weights Initialize to That is, all decision trees have equal initial weights to ensure fair initial voting.

[0198] Classification loss calculation: based on the final classification result With real labels The cross-entropy loss is used as the training loss, denoted as . .because Based on voting scores Therefore, the loss calculation must be based on the probability distribution of the voting scores, and its expression is: In the formula Let be the voting score of the i-th sample in category c; The sum of vote scores for all categories for the i-th sample is used to normalize the vote scores into a probability distribution; the meanings of the remaining parameters are the same as those in the previous section. Consistent.

[0199] Parameter update: Parameter update is divided into two parts: decision tree weight adjustment and split threshold adjustment.

[0200] 1. Decision Tree Weight Adjustment: The weights of each decision tree are dynamically adjusted based on its classification accuracy. Decision trees with higher accuracy are assigned higher weights to increase their influence in the voting process. First, the classification accuracy of the m-th decision tree is calculated. , The expression is ,in Let m be the classification result of the m-th decision tree for the i-th sample. The true class of the i-th sample (by...) Determined, that is Then update the weights based on accuracy. The update expression is This expression ensures that the sum of the weights is 1, while giving greater weights to decision trees with higher accuracy.

[0201] 2. Splitting Threshold Adjustment: Optimize the quantum splitting threshold using gradient descent. Set the learning rate =0.001, the update expression is: In the formula To adjust the splitting threshold in round t+1, To adjust the splitting threshold in the previous t-th round; For loss The gradient of the split threshold is calculated using the chain rule, reflecting the impact of changes in the split threshold on the classification loss.

[0202] Model application process and interaction with LSTM: Quantum Random Forest achieves classification result output and parameter feedback, forming a bidirectional interaction with the Long Short-Term Memory network. It first receives the optimized features from the quantum surface fitting output. ,Will Input each quantum decision tree, obtain the classification result of each tree through the quantum splitting rule, and then output the final classification result through an ensemble voting mechanism. Voting scores by category Meanwhile, the quantum random forest calculates the feedback coefficients. It is used to adjust the gating weights of the Long Short-Term Memory network. The calculation method is the average of the percentages of the maximum votes across all samples. The percentage of the maximum votes refers to the ratio of the maximum vote score in the sample to the sum of all vote scores. The expression is . The value range is [0.25, 1]. When the value is close to 1, it indicates that the voting results of the sample are concentrated and the classification confidence is high; when... A value close to 0.25 indicates that the voting results are dispersed and the classification confidence is low. Low classification confidence is often due to insufficient capture of temporal features (such as the failure to effectively extract changes in insulation state over time). Based on this characteristic, [the following is a possible interpretation:] The LSTM's gate weights are adjusted as feedback parameters, and the specific interaction formula is as follows: ; In the formula and These are the adjusted LSTM forget gate and input gate weights for round t+1. and These are the weights for round t before adjustment; The adjustment coefficient for the input gate weights ensures that when When the classification confidence is low (hourly), the input gate weights can be increased. The core function of this interaction is: when... When the temporal features are insufficient, the weight of the forget gate is reduced to retain more historical temporal information, while the weight of the input gate is increased to absorb more current temporal information. This two-pronged approach enhances the temporal feature extraction capability of LSTM, providing higher-quality temporal features for subsequent fusion.

[0203] 5.2.5 General Formula for Collaborative Fusion of Four Algorithms

[0204] By integrating the input-output relationships, parameter passing paths, and bidirectional interaction logic of the four algorithms, a general formula for the synergistic integration of the four algorithms is formed, fully embodying the closed-loop process of extraction-optimization-classification-feedback:

[0205] 1. Locally correlated feature extraction (CNN): ;

[0206] 2. Temporal Dynamic Feature Extraction (LSTM): ,in , ;

[0207] 3. Preliminary Fusion (CNN and LSTM): ;

[0208] 4. Nonlinear Eigenvalue Optimization (QSF): ;

[0209] 5. Insulation Condition Classification (QRF): ;

[0210] 6. Two-way feedback interaction: , , ;

[0211] 5.3 Overall Model Training and Optimization Process

[0212] Training parameter settings:

[0213] Batch size B: It can take two values, 32 and 64, and is used alternately during training. Choosing 32 is to ensure the diversity of samples within the batch when the number of samples is small, while choosing 64 is to improve training efficiency when the number of samples is sufficient. Alternating between them can balance diversity and efficiency.

[0214] Training epochs: There are six possible values: 50, 60, 70, 80, 90, and 100. A fixed number of epochs is not set; instead, the training stops dynamically based on the validation set loss. Setting multiple epochs is to avoid undertraining (too few epochs) or overfitting (too many epochs). The validation set loss is used to determine whether the model has converged.

[0215] Learning rate: Learning rate of convolutional neural networks =0.001, Long Short-Term Memory Network Learning Rate =0.001, Quantum Surface Fitting Learning Rate =0.0005, learning rate of quantum random forest =0.001. The QSF learning rate is low because its nonlinear mapping is more sensitive to parameter changes, and a low learning rate can avoid unstable fitting; the other three have the same learning rate to ensure coordinated training rhythm.

[0216] Quantum decision tree parameters: number of trees =100, number of floors =5. 100 trees can ensure the ensemble effect and avoid overfitting of a single tree; the 5-layer structure can control the amount of computation while ensuring classification accuracy and avoid the tree structure being too complex.

[0217] Feature dimension parameters: Output dimensions of local features and temporal features =128, convolutional layer feature height and width H=W=16. =128 can balance feature representation ability and computational efficiency; H=W=16 corresponds to the 256 features extracted by S3, ensuring that the features can be completely mapped.

[0218] Training iteration process: The overall model training proceeds iteratively according to the process of data loading - forward computation - backpropagation - interactive update - early stopping judgment. Each iteration implements the collaborative optimization of four algorithms:

[0219] 1. Data Loading: Load the S4-converted training and validation set data from the model training directory on the computing terminal. The training set accounts for 70% of the total samples, and the validation set accounts for 20%. Divide the training set into multiple batches according to the currently set batch size B, and load the feature vectors of one batch in each iteration. With corresponding tags At the same time, the timestamp information of this batch is loaded for the time sequence organization of LSTM.

[0220] 2. Forward computation: The forward computation is performed in the order of CNN→LSTM→QSF→QRF. First, the forward computation is performed... Input CNN, output Then Sort by timestamp Input LSTM output , fusion Next, Input QSF, Output And calculate Finally, Input QRF, Output And calculate Voting scores for each category .

[0221] 3. Loss Calculation: Calculate the independent losses and the overall total loss for the four algorithms. Independent losses include those of the CNN algorithm. LSTM QSF's QRF Total loss The expression is obtained by weighted summation. The weights of QSF and QRF are set at 30%, and CNN and LSTM at 20% because the non-linear optimization of QSF and the classification decision of QRF directly determine the accuracy of the final test results. They are core components and should be given higher weights.

[0222] 4. Backpropagation and Interactive Update: The total loss is calculated using the backpropagation algorithm. For the gradients of all parameters, update the parameters in the order of QRF→QSF→LSTM→CNN to ensure that optimizations in subsequent algorithms can be propagated to preceding algorithms. Simultaneously, perform bidirectional feedback updates: based on the output of QSF... Adjust the convolutional kernel weights of the CNN based on the QRF output. Adjusting the weights of the forget gate and input gate in the LSTM enables dynamic interactive optimization of parameters.

[0223] 5. Early stopping judgment: After each training round (all training batches have been iterated), calculate the total loss of the validation set using the validation set data. .like If no decrease occurs for 5 consecutive rounds (i.e., the change in loss value is less than 0.0001), the model is considered to have converged and training is stopped; if the early stopping condition is not met, the next iteration continues until the maximum number of training rounds (100 rounds) is reached.

[0224] Model Evaluation and Optimization: After model training, performance needs to be evaluated using a test set. If the performance does not meet the requirements, iterative optimization should be performed to ensure that the model can meet the actual testing needs.

[0225] Evaluation metrics calculation: The test set (10% of the total sample) is divided using S4 for evaluation. Core evaluation metrics include Acc (Accuracy), Precision (Prec), Recall (Rec), and F1 score. Accuracy reflects the proportion of samples correctly classified by the model. Precision reflects the proportion of samples predicted as belonging to a certain class that actually belong to that class; the overall precision is the average of the precision of the four classes. Recall reflects the proportion of samples that actually belong to a certain class that were correctly predicted by the model; the overall recall is the average of the recall of the four classes. The F1 score is the harmonic mean of precision and recall to avoid bias from a single metric; the overall F1 score is the average of the F1 scores of the four classes.

[0226] Optimization Standards and Iterations: The evaluation criteria are set as test set accuracy ≥ 90% and overall F1 score ≥ 88%. If the criteria are met, the current model is saved as Insulation Performance Test Model_v1.0 and stored in the model library of the computing terminal for use in actual S6 testing; if the criteria are not met, the causes of deviations are analyzed and targeted optimizations are made.

[0227] 1. If the recall rate of a certain type of insulation condition is low (such as severe aging condition), return to S4 to supplement the sample data of that type of condition (such as adding severe aging equipment feature samples under 1.5 times rated voltage conditions) and reconstruct the dataset.

[0228] 2. If the overall accuracy is low, return to the S3 optimization feature extraction method (such as adding new features like partial discharge pulse density under overvoltage and diurnal variation of dielectric loss value) to improve feature discrimination.

[0229] 3. If the loss function converges slowly, return to 3.1 to adjust the learning rate (e.g., adjust the learning rate of QSF from 0.0005 to 0.0008) to speed up the convergence.

[0230] After optimization, return to step 3.2 and re-execute the training iteration process until the model evaluation meets the target.

[0231] S6: Actual Insulation Performance Test and Result Output

[0232] 6.1 Test Data Preprocessing: The newly acquired high-voltage equipment insulation data is processed according to the S2-S4 procedure: data cleaning yields... Missing and outlier handling results Extracting feature vectors from new data Use the mean of the training set and standard deviation (To prevent data breaches) Standardization .

[0233] 6.2 Model Loading and Prediction:

[0234] Model loading: Load the insulation performance test model_v1.0 after S5 training has reached the target from the model library of the computing terminal. This model contains the optimal parameters of CNN, LSTM, QSF, and QRF.

[0235] Prediction process: Input the model according to the S5 forward computation flow:

[0236] 1. CNN Feature Extraction: Remodeled into a four-dimensional tensor Input to a CNN model, output local feature vectors ;

[0237] 2. LSTM Feature Extraction: [The text abruptly ends here, likely due to an incomplete sentence or a format Sort by timestamp as a time series Input an LSTM model and output a time-series feature vector. ;

[0238] 3. Preliminary fusion: calculation ;

[0239] 4. Quantum Surface Fitting Optimization: Input Output optimized features ;

[0240] 5. Quantum Random Forest Classification: Input The final classification result is output through an integrated voting mechanism. (Values ​​range from 1 to 4, corresponding to four types of insulation states), and outputs the voting scores for each category. .

[0241] 6.3 Results Output and Feedback:

[0242] Output format: Output a test report, including the following core content:

[0243] Basic equipment information: equipment number, test time, test conditions;

[0244] Classification results: The insulation status corresponding to each data record (normal / slight aging / moderate aging / severe aging);

[0245] Confidence level: Calculated based on voting scores (i represents the i-th record), reflecting the reliability of the classification results;

[0246] Feature analysis: The top 10 features that contribute most to the classification results are highlighted (based on the importance of CNN convolutional kernel weights and QRF decision tree features) to provide a reference for equipment maintenance;

[0247] Maintenance recommendations: Based on the classification results, specific recommendations will be provided (normal condition: regular monitoring; slight aging: increase monitoring frequency; moderate aging: schedule planned maintenance; severe aging: immediately shut down for maintenance).

[0248] Results Feedback and Model Iteration: If the confidence level of severe aging in the test results is less than 80%, or if the test results of the same device are inconsistent (e.g., three consecutive tests show slight, moderate, and slight aging respectively), then the test data of the device will be added to the original dataset of S1, the model will be retrained according to the process of S2-S5, and the insulation performance test model_v1.0 in the model library will be updated to the new version.

[0249] If the test accuracy of a certain type of insulation state (based on the results of subsequent manual re-inspection) is less than 85%, then return to S3 to optimize the feature extraction method corresponding to that type of state (such as adding targeted features), reconstruct the dataset and train the model;

[0250] Regularly (every 6 months), collect the operating data and maintenance records of all test equipment, and perform batch iterative optimization of the model to ensure that the model adapts to the long-term changing trend of equipment aging.

[0251] 6.4 Closed-Loop Testing Process: The actual testing process forms a closed loop of data acquisition → preprocessing → feature extraction → standardization → model prediction → result feedback → model optimization, ensuring the accuracy and adaptability of the testing method. After each model optimization, performance must be re-evaluated using a new test set to ensure that the accuracy ≥ 90% and F1 score ≥ 88% are met before it can be used in the next round of actual testing.

Claims

1. A method for testing the insulation performance of high-voltage equipment based on deep learning, characterized in that, include: Step 1: Collect insulation parameter data of high-voltage equipment; Step 2: Preprocess the collected insulation parameter data to obtain preprocessed data; Step 3: Extract multi-dimensional features from the preprocessed data to form feature vectors; Step 4: Standardize the feature vectors to obtain a standardized feature matrix. At the same time, construct a label matrix based on the insulation status of the high-voltage equipment. Divide the standardized feature matrix and the label matrix into training set, validation set and test set. Step 5: Construct a deep learning model that integrates convolutional neural networks, long short-term memory networks, quantum surface fitting, and quantum random forests. Train the model using the training set and dynamically adjust the model parameters using the validation set during training. Evaluate the model performance using the test set. Step 6: Apply the qualified model to the newly collected insulation parameter data of high-voltage equipment and output the insulation status classification results; The core logic of model training in step 5 is as follows: Local correlation feature vectors are extracted from the standardized feature matrix using a convolutional neural network; the local correlation feature vectors sorted by timestamps are processed using a long short-term memory network to extract temporal dynamic feature vectors; the local correlation feature vectors and temporal dynamic feature vectors are initially fused to obtain a preliminary fused feature vector; the preliminary fused feature vectors are nonlinearly optimized using quantum surface fitting to obtain an optimized feature vector; the optimized feature vectors are classified using a quantum random forest to obtain the insulation state classification result; simultaneously, the quantum surface fitting outputs mapping coefficients to adjust the parameters of the convolutional neural network, and the quantum random forest outputs feedback coefficients to adjust the parameters of the long short-term memory network, forming a closed-loop optimization.

2. The method for testing the insulation performance of high-voltage equipment based on deep learning according to claim 1, characterized in that, Step 2 preprocessing includes data cleaning, missing value handling, and outlier handling. Data cleaning uses the moving average filtering method. Missing value handling uses linear interpolation, K-nearest neighbor interpolation, or re-collecting data depending on the missing value ratio. Outlier handling uses the 3σ criterion to detect and replace outliers with normal range boundary values.

3. The method for testing the insulation performance of high-voltage equipment based on deep learning according to claim 1, characterized in that, The multi-dimensional features in step 3 include the time-domain features, frequency-domain features, and phase features of partial discharge; the basic features, time-series trend features, operating condition response features, correlation features, and abrupt change features of dielectric loss; and the basic features, time-series trend features, operating condition response features, correlation features, abrupt change features, and harmonic features of leakage current.

4. The method for testing the insulation performance of high-voltage equipment based on deep learning according to claim 1, characterized in that, In step 4, the standardization process adopts the Z-score standardization method, and the label matrix is ​​constructed using one-hot encoding, corresponding to four insulation states of high-voltage equipment: normal state, slight aging state, moderate aging state, and severe aging state.

5. The method for testing the insulation performance of high-voltage equipment based on deep learning according to claim 1, characterized in that, Convolutional neural networks adopt a progressive structure of input layer - 2 convolutional layers - 2 pooling layers - 1 fully connected layer. The input layer receives the standardized feature matrix and reshapes it into a four-dimensional tensor. The convolutional layer calculates and captures the local correlation between features. The pooling layer retains key local features and reduces the dimensionality. The fully connected layer compresses the high-dimensional features into a fixed-dimensional local correlation feature vector.

6. The method for testing the insulation performance of high-voltage equipment based on deep learning according to claim 1, characterized in that, The Long Short-Term Memory (LSTM) network adopts an input layer-two LSTM unit layers-one fully connected layer structure. The input layer organizes the local correlation feature vectors into a temporal sequence according to timestamps. The LSTM unit layers capture temporal dynamic dependencies through the synergistic effect of forget gate, input gate, cell state and output gate. The fully connected layer compresses the output of the last time step into a fixed-dimensional temporal dynamic feature vector.

7. The method for testing the insulation performance of high-voltage equipment based on deep learning according to claim 1, characterized in that, The process of quantum surface fitting is as follows: each feature dimension of the initially fused feature vector is mapped to the ground state of a quantum bit; a nonlinear fitting surface is constructed based on the measurement expectation value of the quantum state; the initially fused feature vector is optimized to obtain the optimized feature vector; and the average value of the square of the quantum amplitude of all feature dimensions of all samples is calculated as the mapping coefficient.

8. The method for testing the insulation performance of high-voltage equipment based on deep learning according to claim 1, characterized in that, Quantum random forest is composed of multiple quantum decision trees. Each quantum decision tree determines the direction of sample splitting through quantum measurement principles. Quantum random forest determines the final classification result through an ensemble voting mechanism. During the voting process, the weights are dynamically adjusted based on the classification accuracy of each quantum decision tree. At the same time, the average of the maximum voting score percentage of all samples is calculated as the feedback coefficient.

9. The method for testing the insulation performance of high-voltage equipment based on deep learning according to claim 1, characterized in that, In step 5, the total loss of model training is obtained by weighted summation of the cross-entropy loss of the convolutional neural network, the cross-entropy loss of the long short-term memory network, the mean square error loss of the quantum surface fitting, and the cross-entropy loss of the quantum random forest. The loss weights of the quantum surface fitting and the quantum random forest are higher than those of the convolutional neural network and the long short-term memory network.

10. The method for testing the insulation performance of high-voltage equipment based on deep learning according to claim 1, characterized in that, In step 5, the evaluation metrics for model performance include accuracy, precision, recall, and F1 score. If the recall of a certain type of insulation state is lower than the preset value, the sample data of that type of state is supplemented and the dataset is reconstructed. If the overall precision is lower than the preset value, the feature extraction method is optimized. If the loss function converges slowly, the model learning rate is adjusted until the model evaluation meets the target.

Images