High-voltage motor stator winding partial discharge signal identification method and system

Abstract

The invention discloses a high-voltage motor stator winding partial discharge signal identification method and system, and relates to the technical field of high-voltage motor insulation state monitoring, and the method comprises the following steps: obtaining a partial discharge signal of a high-voltage motor stator winding; sequentially carrying out wavelet packet decomposition and feature extraction on the partial discharge signal to obtain a feature vector; inputting the feature vector into a pre-trained CS-RF model to obtain a final recognition result; the pre-trained CS-RF model is obtained through training by taking the minimum expected misclassification cost as a node splitting objective function, and the expected misclassification cost is obtained through calculation of a preset misclassification cost matrix. According to the invention, the problem of limited fault identification precision caused by misclassification equalization, insufficient feature extraction pertinence and unbalanced sample categories in the prior art can be solved.

Description

Technical Field

[0001] This invention relates to the field of high-voltage motor insulation condition monitoring technology, specifically to a method and system for identifying partial discharge signals in the stator winding of a high-voltage motor. Background Technology

[0002] High-voltage motors, as core equipment in power systems and industrial production, have their stator winding insulation performance directly determining their service life. Partial discharge (PD) is a key early indicator of insulation aging and defect development. Internal partial discharge (IPD) and surface partial discharge (SPD) correspond to different insulation degradation modes. IPD originates from air gaps within the winding insulation, while SPD originates from surface contamination or insufficient creepage distance. If these two types of faults are not identified in time, they will gradually deteriorate, leading to insulation breakdown and causing major accidents such as motor burnout and production line shutdown. In existing technologies, to achieve automatic identification and classification of partial discharge signals, the Random Forest (RF) algorithm is widely used for PD signal identification due to its strong anti-interference ability and good adaptability to high-dimensional features.

[0003] However, the aforementioned existing technical solutions still have inherent limitations in practical engineering applications. First, the "equal misclassification cost" strategy, meaning that traditional random forest models assume all types of misclassification have the same cost during training, fails to distinguish the difference in engineering risks caused by "misclassifying a fault discharge as a normal state (false negative)" and "misclassifying a normal state as a fault discharge (false positive)." False negatives lead to insulation defects not being detected in time and continuing to deteriorate, causing serious consequences, while false positives can usually be eliminated with only manual verification; the two have vastly different impacts on engineering safety. Second, traditional RF fails to fully optimize feature utilization based on the time-frequency distribution characteristics of PD signals (e.g., internal discharge and surface discharge are concentrated in specific frequency bands) during the feature extraction stage, resulting in insufficient recognition accuracy under strong power frequency interference and random noise environments. Third, there is an imbalance between the number of fault samples and normal samples, leading to insufficient recall of fault-type signals under complex operating conditions, making it difficult to meet the stringent requirements of early warning of high-reliability insulation conditions. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for identifying partial discharge signals in the stator windings of high-voltage motors, so as to solve the problems of limited fault identification accuracy caused by misclassification, insufficient feature extraction, and sample class imbalance in the existing technology.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, a method for identifying partial discharge signals in the stator winding of a high-voltage motor includes the following steps: Acquire the partial discharge signal of the stator winding of a high-voltage motor; The partial discharge signal is subjected to wavelet packet decomposition (WPD) and feature extraction in sequence to obtain a feature vector; The feature vectors are input into a pre-trained CS-RF (Cost Sensitive-Random Forest) model to obtain the final recognition result; The pre-trained CS-RF model is trained using the minimization of the expected misclassification cost as the node splitting objective function, and the expected misclassification cost is calculated through a preset misclassification cost matrix.

[0006] In some implementations, the step of sequentially performing wavelet packet decomposition and feature extraction on the partial discharge signal specifically includes: The partial discharge signal is decomposed using the db6 wavelet basis to obtain several equal-bandwidth sub-bands; Within the typical frequency range of a preset partial discharge type, several equal-bandwidth sub-bands that simultaneously satisfy the energy proportion condition and the distinguishability condition are selected as key sub-bands; the energy proportion condition is: the proportion of the energy of the equal-bandwidth sub-band in the total energy of the preset partial discharge type is greater than or equal to a first preset threshold; the distinguishability condition is: the frequency overlap of different equal-bandwidth sub-bands is less than or equal to a second preset threshold. Feature extraction is performed on the key subband to obtain feature vectors.

[0007] In some implementations, the feature vector includes PD energy ratio, wavelet packet entropy, PD pulse number, power frequency suppression ratio, PD average amplitude, and PD energy entropy.

[0008] In some implementations, the pre-trained CS-RF model is trained by the following steps: Acquire partial discharge signal samples of the stator winding of a high-voltage motor, and divide the partial discharge signal samples into a training set and a test set. The partial discharge signal samples include: no partial discharge signal, internal discharge signal and surface discharge signal. The Bootstrap method is used to generate multiple training subsets from the training set to train the random forest model. Then, the random forest model is cross-validated and validated on the test set to obtain the CS-RF model. The random forest model has 100-180 decision trees, and the number of training subsets is the same as the number of training sets. The random forest model selects nodes for splitting. One characteristic, The number of features in the feature vector.

[0009] In some implementations, the feature vector is input into a pre-trained CS-RF model to obtain the final recognition result, specifically including: The feature vector is input into a pre-trained CS-RF model, and the category of the feature vector is predicted through a decision tree voting mechanism. The category with the highest number of votes is used as the initial identification result. If the vote rate of the preliminary identification result is greater than or equal to the preset vote rate threshold, then the preliminary identification result is taken as the final identification result; otherwise, a verification indication is output.

[0010] In some implementations, in the preset misclassification cost matrix, rows represent the true category of the partial discharge signal, and columns represent the predicted category of the partial discharge signal. The true category includes: no partial discharge, internal discharge, and surface discharge, and the predicted category includes: no partial discharge, internal discharge, and surface discharge. When the true category is internal discharge or surface discharge, and the predicted category is no partial discharge, the corresponding element in the preset misclassification cost matrix is ​​the underreporting cost. When the true category and the predicted category are the same, the element at the corresponding position in the preset misclassification cost matrix is ​​the correct classification cost, and the correct classification cost is zero. In other cases, the element at the corresponding position in the preset misclassification cost matrix is ​​the false alarm value; The value of the missed report is greater than the value of the false report.

[0011] Secondly, a system for identifying partial discharge signals in the stator winding of a high-voltage motor includes: The data acquisition module is used to acquire the partial discharge signal of the stator winding of the high-voltage motor; The feature extraction module is used to sequentially perform wavelet packet decomposition and feature extraction on the partial discharge signal to obtain a feature vector. The signal category recognition module is used to input the feature vector into a pre-trained CS-RF model to obtain the final recognition result. The pre-trained CS-RF model is trained with minimizing the expected misclassification cost as the node splitting objective function. The expected misclassification cost is calculated by a preset misclassification cost matrix.

[0012] Thirdly, an electronic device includes a memory, a processor, and a computer program stored in the memory and executable in the processor, wherein the processor executes the computer program to implement the steps of the method for identifying partial discharge signals of a high-voltage motor stator winding.

[0013] Fourthly, a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method for identifying partial discharge signals in a high-voltage motor stator winding.

[0014] Fifthly, a computer program product comprising a computer program that, when executed by a processor, implements the steps of the method for identifying partial discharge signals of a high-voltage motor stator winding.

[0015] Compared with the prior art, the present invention has the following beneficial effects: This invention provides a method for identifying partial discharge signals in the stator windings of a high-voltage motor. The method trains the CS (System-Controlled Component) algorithm by minimizing the expected misclassification cost as the objective function for node splitting. The RF model is used, and the cost is calculated using a pre-set misclassification cost matrix. This allows the model to explicitly distinguish between different risk levels of missed faults and normal false alarms during training. As a result, the trained model will actively tend to classify suspected fault samples as fault categories when making identifications, thereby significantly reducing the missed fault rate, improving the reliability of early warning of insulation faults in high-voltage motors, and solving the problem of insufficient adaptability of traditional random forests in safety-sensitive scenarios.

[0016] Furthermore, key sub-bands were screened using energy proportion and discriminative criteria. The energy proportion condition ensured that the extracted features contained the most important energy information for that discharge type, resulting in strong characterization and good resistance to random noise. The discriminative criterion, by controlling frequency overlap, ensured that the key sub-bands selected for different discharge types were separated as much as possible in the frequency domain, fundamentally enhancing the class discrimination ability of the extracted features, thus laying the foundation for subsequent CS (Computer Science) analysis. The RF model provides high-quality, high-purity input, laying a solid foundation for solving the problem of insufficient recognition accuracy.

[0017] Furthermore, a bootstrap method is employed to generate training subsets. By sampling with replacement, the data subset used for each training iteration is slightly different. This injects necessary diversity into the random forest, improving the model's generalization ability and preventing overfitting. It also naturally alleviates the class imbalance problem caused by the scarcity of on-site fault samples, as minority class samples may be repeatedly sampled in some bootstrap subsets, thus increasing the model's learning opportunities for fault categories. Finally, cross-validation and test set validation are combined to ensure the final CS (Comparative Stability) is optimized. The RF model maintains stable and superior recognition performance even when faced with real, imbalanced field data.

[0018] Furthermore, when the vote rate of the highest-voting category exceeds the preset vote rate threshold, it indicates that the decision-making consistency within the forest is high and the result is reliable, so it can be directly output. When the vote rate is lower than the threshold, it indicates that the model's judgment of the sample has a large degree of uncertainty. At this time, an instruction to be reviewed is output, which cleverly balances the efficiency and security of automated identification. It makes quick decisions for high-confidence samples and triggers manual or auxiliary means to review low-confidence samples, further cutting off the risk of misjudgment that may be caused by the uncertainty of the model itself, and improving the engineering practicality and safety margin of the whole method.

[0019] Furthermore, by pre-setting the false negative cost matrix so that the value of missed false positives is greater than the value of false positives, the engineering requirement that the risk of missed false positives far outweighs the cost of false positives is precisely encoded into the control parameters for model training. This enables CS While pursuing the best overall accuracy, the RF model is forced to prioritize minimizing high-risk false negatives, thus ensuring from the source of parameter design that the model behavior is highly consistent with the goal of equalizing the cost of misclassification. Attached Figure Description

[0020] Figure 1 A flowchart illustrating a method for identifying partial discharge signals in the stator winding of a high-voltage motor, provided in an embodiment of the present invention; Figure 2 This is a structural diagram of a high-voltage motor stator winding partial discharge signal identification system provided in an embodiment of the present invention; Figure 3 The diagram shows a comparison of the effects of a method for identifying partial discharge signals in the stator winding of a high-voltage motor, as provided in an embodiment of the present invention. Detailed Implementation

[0021] To enable those skilled in the art to better understand the present invention, the technical solution of the present invention will be further described in detail below with reference to the accompanying drawings. The content described herein is for explanation rather than limitation of the present invention.

[0022] It should be noted that the terms "comprising" and "having" and any variations thereof in the specification of this invention are intended to cover a non-exclusive inclusion, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such processes, methods, systems, products or devices.

[0023] First, let's explain the English abbreviations used in this embodiment: Application-Specific Integrated Circuit (ASIC); Bootstrap Sampling Method (Bootstrap); Central Processing Unit (CPU); Cost-Sensitive Random Forest (CS-RF); Daubechies 6 (Wavelet), db6; Digital Signal Processor (DSP); Field-Programmable Gate Array (FPGA); Internal partial discharge (IPD); K-fold Cross Validation (K-fold); Partial discharge (PD); Random Access Memory (RAM); Random Forest (RF); Surface partial discharge (SPD); Wavelet Packet Decomposition (WPD).

[0024] This invention provides a method for identifying partial discharge signals in the stator windings of high-voltage motors based on cost-sensitive random forests. The method comprises five core steps: PD signal acquisition, WPD feature extraction, CS-RF model construction, model training and iterative verification, and PD signal identification. The specific process is as follows: (1) Acquisition of PD signal and initial parameters: Acquire the partial discharge signal of the stator winding of the high-voltage motor. The partial discharge signal includes no partial discharge signal, internal discharge signal, surface discharge signal and associated parameters. The associated parameters include signal sampling parameters (sampling frequency, sampling duration), motor structural parameters (stator winding insulation type, air gap size), and monitoring environment parameters (power frequency interference intensity, environmental noise level). The no partial discharge signal is generated by the following formula:

[0025] Where A=5V is the power frequency signal amplitude, f=50Hz is the power frequency, σ=0.5V is the Gaussian noise intensity, and ε(t) is the standard Gaussian white noise. S 无PD There is no partial discharge signal; The internal discharge signal is given by the following formula:

[0026] in =95~105kHz is the center frequency of the internal discharge pulse. =1.5~2.5V is the pulse amplitude, α=8000~12000s - ¹ represents the attenuation coefficient. The time when the random pulse occurs. ∈[0,2π] represents the initial phase, and N=40~60 represents the number of pulses. t For time variables, S IPD ( t () is an internal discharge signal.

[0027] The surface discharge signal is given by the following formula:

[0028] in =290~310kHz is the center frequency of the surface discharge pulse. S spD ( t The signal is the surface discharge difference signal, and the other parameters are the same as the internal discharge signal.

[0029] (2) WPD Feature Extraction: Select a preset wavelet basis, such as the Daubechies Wavelets (db) series, the Symlets Wavelets (sym) series, etc., and perform N-level decomposition on the PD signal (N is a positive integer, determined by the highest analysis frequency F_max and the sub-band accuracy requirements), dividing the 0~F_max frequency band into 2 Each sub-band has equal bandwidth. Then, based on the typical frequency range of the faulty PD signal, sub-bands with an energy ratio of ≥60% are selected as key sub-bands. The typical frequency range includes, but is not limited to, the frequency range of internal discharge (approximately 50~200kHz) and the frequency range of surface discharge (approximately 200~500kHz).

[0030] M-dimensional features (M≥3) are extracted from the key sub-bands. The features include energy type (sub-band energy ratio, energy proportion), entropy type (wavelet packet entropy, energy entropy), pulse type (pulse number, average amplitude), and interference suppression type (power frequency interference ratio). The M-dimensional feature is preferably 6-dimensional, and the specific calculation method is as follows: F1 (PD Energy Ratio): Sum of energy in key PD frequency bands / Total energy in non-power frequency bands; F2 (wavelet packet entropy) = , Let i be the proportion of the energy of the i-th subband to the total energy. H The wavelet packet entropy; F3 (PD pulse count): The number of pulses in which the absolute value of the wavelet coefficients in the key subband is greater than the preset threshold; F4 (Power Frequency Rejection Ratio): Low-frequency band (including 50Hz) energy / total signal energy; F5 (PD Average Amplitude): The average of the absolute values ​​of the wavelet coefficients in the key subband; F6 (PD Energy Entropy): ,in, H ’ Let M be the energy entropy of the PD and the number of critical PD subbands. The value represents the proportion of the energy of the j-th critical subband to the total energy of the critical subband. Its diagnostic significance is to reflect the concentration of PD energy in the critical frequency band.

[0031] (3) CS-RF model construction: The CS-RF model is constructed based on the traditional random forest framework. The traditional random forest consists of 100-200 decision trees. When splitting a decision tree node, the number of features randomly selected is the square root of the total number of features. The total number of features is 6, so 2 features are randomly selected. A k×k misclassification cost matrix is ​​introduced, where k is the number of PD signal categories. The principle of fault missed reporting cost > false alarm cost > correct classification cost is followed. The correct classification cost C(i,i)=0, the fault is the no-PD missed reporting cost C(i,1)=5~10, and the no-PD is the fault false alarm cost C(1,i)=1~3. The misclassification cost matrix is ​​embedded into the decision tree training process, and the node splitting objective function is changed from minimizing the Gini coefficient to minimizing the expected misclassification cost. The formula for calculating the expected misclassification cost is as follows:

[0032] In the formula, N is the total number of samples in the current node. , These represent the number of samples for the left and right child nodes, respectively. , These are the misclassification costs for the left and right child nodes, respectively. Cost The cost of expected misclassification.

[0033] The k×k misclassification cost matrix, where k is the number of PD signal categories, follows the principle of fault underreporting cost > false alarm cost > correct classification cost. The correct classification cost C(i,i)=0, the underreporting cost of fault judgment as no PD C(i,1)=5~10, and the false alarm cost of no PD judgment as fault C(1,i)=1~3.

[0034] The basic parameters of the CS-RF model are: number of decision trees = 100~180, preferably 100; number of features randomly selected when splitting a node = 2; total number of features M = 6. An integer, Bootstrap sampling ratio = 1, indicates that the number of samples in the training subset is consistent with the original training set; the objective function for splitting decision tree nodes is to minimize the expected misclassification cost, calculated using the following formula:

[0035] In the formula, N is the total number of samples in the current node. , These represent the number of samples for the left and right child nodes, respectively. , These are the misclassification costs for the left and right child nodes, respectively.

[0036] (4) Model training and iterative verification: Divide the M-dimensional features into training and test sets in a ratio of 7:3 to 8:2, and train the CS-RF model using Bootstrap sampling with replacement; evaluate the model performance using K-fold cross-validation, set a verification threshold, and if the model performance does not meet the standard, adjust the misclassification cost matrix parameters or the range of key feature sub-bands, and repeat steps (1) and (2); if it meets the standard, proceed to the signal recognition stage. The core evaluation indicators of the 10-fold cross-validation include: Accuracy = (Number of correctly classified non-PD samples + Number of correctly classified internal discharge samples + Number of correctly classified surface discharge samples) / Total number of samples; Internal discharge recall = Number of correctly classified internal discharge samples / Total number of true internal discharge samples; Surface discharge recall = Number of correctly classified surface discharge samples / Total number of true surface discharge samples; Overall score = 0.3 × Precision + 0.25 × (Internal discharge recall + Surface discharge recall) / 2 + 0.2 × No-PD recall, where No-PD recall = Number of correctly classified No-PD samples / Total number of real No-PD samples.

[0037] (5) PD signal identification: For the stator winding partial discharge signal to be identified, complete the WPD decomposition and feature extraction according to step (2), input it into the CS-RF model after training, and calculate the highest category vote rate through decision tree voting. The vote rate = (number of votes for the highest vote category ÷ total number of decision trees) × 100%; if the vote rate is ≥60%~70%, output the result directly; if it is lower than the threshold, output the verification indication and combine infrared imaging and ultrasonic positioning for supplementary verification. The decision tree voting mechanism is: all decision trees count the predicted categories of the same sample to be identified, and the category with the most votes is the final identification result.

[0038] Based on the above, such as Figure 1 As shown, the cost-sensitive random forest method for identifying partial discharge signals in the stator windings of high-voltage motors, as described in this invention, comprises the following specific steps: S1, acquire the partial discharge signal of the stator winding of the high-voltage motor; Specifically, three types of partial discharge signals from the stator winding of the high-voltage motor are acquired: no partial discharge signal, internal discharge signal, and surface discharge signal. These three types of signals are generated by a field high-frequency current sensor (bandwidth 10kHz~1MHz).

[0039] S2, perform wavelet packet decomposition and feature extraction on the partial discharge signal in sequence to obtain the feature vector; Wavelet packet decomposition: The db6 wavelet basis was used. Experiments verified that the db6 wavelet basis has the best decomposition accuracy for PD transient pulses. N-level wavelet packet decomposition was performed on the PD signal, where N is a positive integer, determined by the highest analysis frequency of the signal. The accuracy requirements of the sub-band dictate that 0~ The frequency band is divided into 2 Each equal-bandwidth sub-band, sub-band bandwidth = / 2 Based on the typical frequencies of the PD type of the target device, such as internal discharge typically concentrated in 50~200kHz and surface discharge concentrated in 200~500kHz, sub-bands in the corresponding frequency ranges are selected as key sub-bands. When selecting sub-bands, the core principle is that the sub-band energy accounts for ≥60% of the total energy of that type of PD. For example, the sub-band energy proportion of 100kHz±5kHz for internal discharge is 68%, and the sub-band energy proportion of 300kHz±5kHz for surface discharge is 72%, both meeting the requirements. Next, the frequency overlap of different types of candidate sub-bands is verified. The frequency overlap is defined as intersection bandwidth / total bandwidth ≤30%. If the overlap between the two sub-bands is 0%, the distinguishability is satisfied. Finally, the key sub-bands must simultaneously meet the criteria, containing core information and effectively distinguishing discharge types.

[0040] Feature Calculation: Extract M-dimensional features (M≥3, set according to recognition accuracy requirements) from the key sub-band, including but not limited to: energy features: energy proportion of key sub-band, energy ratio of different sub-bands, etc.; entropy features: wavelet packet entropy, energy distribution entropy, etc.; pulse features: pulse number, average amplitude, pulse interval, etc.; interference suppression features: energy proportion of power frequency interference sub-band, noise suppression ratio, etc. The M-dimensional features are preferably 6-dimensional, and the specific calculation method is as follows: F1 (PD Energy Ratio): Sum of energy in key PD frequency bands / Total energy in non-power frequency bands; F2 (wavelet packet entropy) = , The proportion of the energy of the i-th subband to the total energy; F3 (PD pulse count): The number of pulses in which the absolute value of the wavelet coefficients in the key subband is greater than the preset threshold; F4 (Power Frequency Rejection Ratio): Low-frequency band (including 50Hz) energy / total signal energy; F5 (PD Average Amplitude): The average of the absolute values ​​of the wavelet coefficients in the key subband; F6 (PD energy entropy) = Where M is the number of critical PD sub-bands. The value represents the proportion of the energy of the j-th critical subband to the total energy of the critical subband. Its diagnostic significance is to reflect the concentration of PD energy in the critical frequency band.

[0041] S3, input the feature vector into the pre-trained CS-RF model to obtain the final recognition result; the pre-trained CS-RF model is trained with minimizing the expected misclassification cost as the node splitting objective function, and the expected misclassification cost is calculated through a preset misclassification cost matrix, as follows: Cost-Sensitive Random Forest (CS-RF) Model Construction: First, the basic framework parameters are set. The number of decision trees is set to 100. Experiments have verified that 100 trees can balance accuracy and efficiency: fewer than 100 trees result in insufficient stability, while more than 200 trees significantly increase training time. The bootstrap sampling ratio is 1, indicating that the number of samples in the training subset is consistent with the original training set. When splitting a node, two features are randomly selected, where the total number of features is M. Integers are used to avoid feature redundancy affecting splitting accuracy.

[0042] The design basis of the misclassification cost matrix is ​​as follows: Through multiple sets of comparative experiments, it was verified that when the cost of missing a fault that is judged as normal is set to 8, the model can achieve the optimal balance between the false alarm rate and the false alarm rate, that is, the internal discharge false alarm rate is ≤8.8%, the surface discharge false alarm rate is ≤8%, and the false alarm rate is ≤12%. If the cost is <6, the false alarm rate is >12%, which is significantly higher. If the cost is >10, it indicates that the false alarm rate is too high, and the false alarm rate is >15%. Therefore, the cost of missing a fault is determined to be 8.

[0043] The cost mechanism embeds the following logic: When splitting a decision tree node, the expected misclassification cost for each candidate split point is calculated, as shown in the following formula:

[0044] Choose the split point with the lowest cost. For example, if a node contains 50 internal discharge samples and 50 samples without PD, and after splitting, the left subband contains 40 internal discharge samples and 10 samples without PD, and the right subband contains 10 internal discharge samples and 40 samples without PD, then the expected cost of the left subband is (40×8+10×1) / 50=6.6, the expected cost of the right subband is (10×8+40×1) / 50=2.4, and the total cost is (50 / 100)×6.6 + (50 / 100)×2.4=4.5. This split point is the optimal choice.

[0045] Model training and iterative validation: First, the data is partitioned. The total sample is divided into training and test sets proportionally, such as 7:3 to 8:2. The ratio can be adjusted according to the total sample size; when the sample size is small, the proportion of the training set can be increased. The training set is used for model parameter optimization, and the test set is used to verify generalization ability. Then, cross-validation is performed using K-fold cross-validation, where K = 5 to 10. When the sample size is small, the smaller K is used. Key metrics are calculated: accuracy, recall for each class, F1 score, etc., with a focus on the recall of the faulty class.

[0046] Test set validation: Use the qualified model to identify test samples, requiring test set accuracy ≥ 93% and fault class recall ≥ 90% to ensure the stability of the model on non-training data.

[0047] Signal Recognition: The signal to be recognized is preprocessed using the same wavelet basis, the same number of decomposition layers N, and the same feature extraction rules as the training samples to ensure consistency of input features. Model Prediction: The input features are fed into the trained CS-RF model, which outputs the recognition result through a decision tree voting mechanism. The recognition result is either a PD signal of a certain type or no PD.

[0048] Result verification mechanism: If the highest category's vote rate is greater than or equal to the threshold, such as 60%~70%, it can be adaptively adjusted and the result can be output directly; if the vote rate is lower than the threshold, it will be output as pending verification and confirmed by combining other detection methods, such as infrared imaging and ultrasonic positioning.

[0049] The invention will now be described in further detail with reference to the accompanying drawings and experimental data: PD signal and initial parameter acquisition: Using simulation software, for a specific scenario, a sampling frequency of 1MHz was selected to ensure coverage of the key frequency bands where internal and surface discharges occur. The sampling duration was 0.02s, with each sample containing 20,000 data points. 200 samples of each of the three signal types were generated, as follows (no partial discharge signal):

[0050] Where A=5V is the power frequency signal amplitude, f=50Hz is the power frequency, σ=0.5V is the Gaussian noise intensity, and ε(t) is the standard Gaussian white noise. Internal discharge signal:

[0051] in =95~105kHz is the center frequency of the internal discharge pulse. =1.5~2.5V is the pulse amplitude, α=8000~12000s - ¹ is the attenuation coefficient. The time when the random pulse occurs. ∈[0,2π] represents the initial phase, and N=40~60 represents the number of pulses. Surface discharge signal:

[0052] in =290~310kHz is the center frequency of the surface discharge pulse, and the other parameters are consistent with the internal discharge signal.

[0053] WPD decomposition: Using db6 as the wavelet basis, each sample is decomposed into 6 levels, resulting in 64 subbands. Based on the Nyquist theorem decomposition at a sampling frequency of 1MHz, the subband frequency range covers 0-500kHz, and the bandwidth of a single subband is approximately 7.8125kHz. The coefficients of each subband are extracted using the wpcoef function, and the subband energy is calculated as the sum of squares of the coefficients. Among them, the subbands near the internal discharge frequency of 100kHz±5kHz are numbered 12-15, and the subbands near the surface discharge frequency of 300kHz±5kHz are numbered 35-38.

[0054] Feature calculation: Taking a certain internal discharge sample as an example: F1 (PD energy ratio): The sum of the energies of the 5 sub-bands around 100kHz = 12.5V², and the total energy of the non-power frequency band = 18.4V², so F1 = 12.5 / 18.4 ≈ 0.679; F2 (wavelet packet entropy): relative energy of 64 subbands After calculation, ≈1.86; F3 (PD pulse count): Preset threshold = 3 × standard deviation of wavelet coefficients without PD signal = 0.12V, count the number of pulses with absolute value of coefficients > 0.12V in key subband = 28; F4 (Power Frequency Rejection Ratio): Low-frequency band (including 50Hz) energy = 3.2V², total energy = 21.6V², therefore F4 = 3.2 / 21.6≈0.148; F5 (PD Average Amplitude): The average absolute value of the key subband coefficients = 0.18V; F6 (PD Energy Entropy): After calculating the relative energy of the key subband, ≈0.72.

[0055] CS-RF model training: Dataset: 600 PD signal samples are used, including 200 samples without PD signal, 200 samples of internal discharge signal, and 200 samples of surface discharge signal. The samples are divided into a training set of 420 samples and a test set of 180 samples in a 7:3 ratio.

[0056] Comparison: Traditional RF: a cost-free sensitive mechanism that applies equal costs to misclassification behaviors such as classifying a fault as a fault and a fault as a faultless RF; CS-RF: in the asymmetric cost matrix, the penalty coefficient FP for classifying a fault as a faultless RF takes values ​​of 1, 2, 4, 6, and 8 respectively. This embodiment ultimately uses FP=8. The 3×3 misclassification cost matrix C is shown below:

[0057] The matrix rows represent the true categories (1 = no partial discharge, 2 = internal discharge, 3 = surface discharge), and the columns represent the predicted categories. C(2,1)=C(3,1)=8 indicates that the cost of a false negative when a fault is classified as normal is 8 times the cost of other misclassifications. C(1,1)=C(2,2)=C(3,3)=0 indicates a correct classification with a cost of 0 and no penalty. C(1,2)=C(1,3)=C(2,3)=C(3,2)=1 indicates false positives and misclassifications, i.e., a normal fault is classified as a fault, and an internal discharge is classified as a surface discharge, with a cost of 1 and a light penalty. Then, the misclassification cost matrix is ​​embedded into the decision tree training process. The node splitting objective function is changed from minimizing the Gini coefficient to minimizing the expected misclassification cost. The training subset of each decision tree is obtained by sampling with replacement using Bootstrap. Core evaluation metric: false negative rate, which is the proportion of internal discharge / surface discharge samples that are mistakenly identified as having no PD out of the total number of fault samples.

[0058] Performance comparison analysis: Figure 3 The horizontal axis represents the penalty coefficient FP of the random forest cost matrix, and the vertical axis represents the false negative rate (%) of stator winding partial discharge identification. The continuous curve in the figure shows the overall trend of a monotonically decreasing false negative rate as the penalty coefficient increases; the red dots mark the false negative rate when the penalty coefficient = 1, corresponding to the traditional RF result; the black dashed line indicates the operating point where the penalty coefficient = 8 is selected in this invention, and its false negative rate is reduced to 5%, meeting the industrial requirement of ≤8%. This figure also provides a direct comparison between this invention and traditional RF, verifying the significant technical effect obtained after introducing the asymmetric cost matrix.

[0059] The experiment yielded the effect of the penalty coefficient on the false negative rate and compared the invention with traditional RF, such as... Figure 3 As shown, through Figure 3 It is known that traditional RF has limitations. When FP=1, the false alarm rate of traditional RF is as high as about 18%. FP=1 can be equivalent to the case of equal cost, which is far higher than the engineering requirement of false alarm rate ≤10% for high voltage motor insulation monitoring. The optimization trend of CS-RF: As FP increases, that is, the penalty intensity for the false alarm behavior of the fault being judged as without PD increases, the false alarm rate of CS-RF decreases monotonically. This is because the higher penalty coefficient forces the model to be more inclined to judge the suspected fault as a fault when making decisions, thereby reducing false alarms. When FP=8, which is the cost of internal discharge being judged as having no PD and surface discharge being judged as having no PD in the cost matrix of this invention, the false alarm rate of CS-RF drops to about 5%, which is about 13 percentage points lower than that of traditional RF, and meets the requirement of engineering false alarm threshold ≤10%, thus verifying the suppressive effect of cost-sensitive mechanism on false alarm risk.

[0060] like Figure 2 As shown, this embodiment provides a partial discharge signal identification system for the stator winding of a high-voltage motor, including: The data acquisition module is used to acquire the partial discharge signal of the stator winding of the high-voltage motor; The feature extraction module is used to sequentially perform wavelet packet decomposition and feature extraction on the partial discharge signal to obtain a feature vector. The signal category recognition module is used to input the feature vector into a pre-trained CS-RF model to obtain the final recognition result. The pre-trained CS-RF model is trained with minimizing the expected misclassification cost as the node splitting objective function. The expected misclassification cost is calculated by a preset misclassification cost matrix.

[0061] This embodiment has the following advantages: (1) Precise control of false negative rate: By quantifying the high cost matrix that judges the fault as normal, the internal discharge recall rate of the CS-RF model reaches 91.2±1.8% and the surface discharge recall rate reaches 92.0±1.6%, which is 3%-4% higher than the traditional random forest, effectively avoiding the safety risks caused by false negatives; (2) High feature discrimination: Focusing on the 100kHz and 300kHz key subbands to extract 6-dimensional features, combined with indicators with clear physical meaning such as wavelet packet entropy and energy ratio, the feature discrimination is more than 15% higher than the full-band features, and the anti-interference ability is significantly enhanced; (3) Strong engineering adaptability: The model training efficiency is high, the single-fold training time is ≤5 minutes, and it supports real-time on-site monitoring; The iterative verification mechanism can be adapted to high-voltage motors of different power levels, and the parameter adjustment is flexible, which can guide the formulation of motor insulation maintenance plans; (4) Stable recognition accuracy: The standard deviation of the accuracy of 10-fold cross-validation is ≤1.5%, and the generalization accuracy of the test set is ≥94%, which is better than the traditional random forest algorithm, and the stability meets the long-term monitoring requirements.

[0062] The module division in this embodiment of the invention is illustrative and represents only one logical functional division. In actual implementation, other division methods may be used. Furthermore, the functional modules in the various embodiments of the invention can be integrated into a single processor, exist as separate physical entities, or be integrated into a single module. The integrated modules described above can be implemented in hardware or as software functional modules.

[0063] This embodiment also provides a computer device, which includes a processor and a memory. The memory is used to store a computer program (in this embodiment, the computer program includes a calculation component and an iterative component, capable of model calculation and model updating). The computer program includes program instructions, and the processor is used to execute the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, and is suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions in the computer storage medium to realize the corresponding method flow or corresponding function. The processor described in this embodiment can be used for the operation of a method for identifying partial discharge signals in the stator winding of a high-voltage motor.

[0064] This embodiment also provides a storage medium, specifically a computer-readable storage medium (Memory), which is a memory device in a computer device used to store programs and data. It is understood that the computer-readable storage medium here can include both the built-in storage medium in the computer device and extended storage media supported by the computer device. The computer-readable storage medium provides storage space that stores the terminal's operating system. Furthermore, this storage space also stores one or more instructions suitable for loading and execution by a processor. These instructions can be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here can be high-speed RAM or non-volatile memory, such as at least one disk storage device. The processor can load and execute one or more instructions stored in the computer-readable storage medium to implement the corresponding steps of the high-voltage motor stator winding partial discharge signal identification method in the above embodiment.

[0065] This embodiment also provides a computer program product, which includes a computer program. When the computer program is executed by a processor, it implements the corresponding steps of the high-voltage motor stator winding partial discharge signal identification method described in the above embodiment.

[0066] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0067] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0068] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0069] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0070] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for identifying partial discharge signals in the stator winding of a high-voltage motor, characterized in that, Includes the following steps: Acquire the partial discharge signal of the stator winding of a high-voltage motor; The partial discharge signal is subjected to wavelet packet decomposition and feature extraction in sequence to obtain a feature vector; The feature vector is input into a pre-trained CS-RF model to obtain the final recognition result; The pre-trained CS-RF model is trained using the minimization of the expected misclassification cost as the node splitting objective function, and the expected misclassification cost is calculated through a preset misclassification cost matrix.

2. The method for identifying partial discharge signals in the stator winding of a high-voltage motor according to claim 1, characterized in that, The step of sequentially performing wavelet packet decomposition and feature extraction on the partial discharge signal specifically includes: The partial discharge signal is decomposed using the db6 wavelet basis to obtain several equal-bandwidth sub-bands; Within the typical frequency range of a preset partial discharge type, several equal-bandwidth sub-bands that simultaneously satisfy the energy proportion condition and the distinguishability condition are selected as key sub-bands; the energy proportion condition is: the proportion of the energy of the equal-bandwidth sub-band in the total energy of the preset partial discharge type is greater than or equal to a first preset threshold; the distinguishability condition is: the frequency overlap of different equal-bandwidth sub-bands is less than or equal to a second preset threshold. Feature extraction is performed on the key subband to obtain feature vectors.

3. The method for identifying partial discharge signals in the stator winding of a high-voltage motor according to claim 2, characterized in that, The feature vector includes PD energy ratio, wavelet packet entropy, PD pulse number, power frequency suppression ratio, PD average amplitude, and PD energy entropy.

4. The method for identifying partial discharge signals in the stator winding of a high-voltage motor according to claim 1, characterized in that, The pre-trained CS-RF model is trained through the following steps: Acquire partial discharge signal samples of the stator winding of a high-voltage motor, and divide the partial discharge signal samples into a training set and a test set. The partial discharge signal samples include: no partial discharge signal, internal discharge signal and surface discharge signal. The Bootstrap method is used to generate multiple training subsets from the training set to train the random forest model. Then, the random forest model is cross-validated and validated on the test set to obtain the CS-RF model. The random forest model has 100-180 decision trees, and the number of training subsets is the same as the number of training sets. The random forest model selects nodes for splitting. One characteristic, The number of features in the feature vector.

5. The method for identifying partial discharge signals in the stator winding of a high-voltage motor according to claim 1, characterized in that, The feature vector is input into a pre-trained CS-RF model to obtain the final recognition result, specifically including: The feature vector is input into a pre-trained CS-RF model, and the category of the feature vector is predicted through a decision tree voting mechanism. The category with the highest number of votes is used as the initial identification result. If the vote rate of the preliminary identification result is greater than or equal to the preset vote rate threshold, then the preliminary identification result is taken as the final identification result; otherwise, a verification indication is output.

6. The method for identifying partial discharge signals in the stator winding of a high-voltage motor according to claim 1, characterized in that, In the preset misclassification cost matrix, rows represent the true category of the partial discharge signal, and columns represent the predicted category of the partial discharge signal. The true categories include: no partial discharge, internal discharge, and surface discharge, and the predicted categories include: no partial discharge, internal discharge, and surface discharge. When the true category is internal discharge or surface discharge, and the predicted category is no partial discharge, the corresponding element in the preset misclassification cost matrix is ​​the underreporting cost. When the true category and the predicted category are the same, the element at the corresponding position in the preset misclassification cost matrix is ​​the correct classification cost, and the correct classification cost is zero. In other cases, the element at the corresponding position in the preset misclassification cost matrix is ​​the false alarm value; The value of the missed report is greater than the value of the false report.

7. A system for identifying partial discharge signals in the stator winding of a high-voltage motor, characterized in that, include: The data acquisition module is used to acquire the partial discharge signal of the stator winding of the high-voltage motor; The feature extraction module is used to sequentially perform wavelet packet decomposition and feature extraction on the partial discharge signal to obtain a feature vector; The signal category recognition module is used to input the feature vector into a pre-trained CS-RF model to obtain the final recognition result. The pre-trained CS-RF model is trained with minimizing the expected misclassification cost as the node splitting objective function. The expected misclassification cost is calculated by a preset misclassification cost matrix.

8. An electronic device, characterized in that, The device includes a memory, a processor, and a computer program stored in the memory and executable in the processor. When the processor executes the computer program, it implements the method for identifying partial discharge signals of a high-voltage motor stator winding as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the method for identifying partial discharge signals of a high-voltage motor stator winding as described in any one of claims 1 to 6.

10. A computer program product, the computer program product comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the method for identifying partial discharge signals of the stator winding of a high-voltage motor as described in any one of claims 1 to 6.

Images