Multi-physical-quantity-based turn-to-turn arc severity lightweight discrimination system and method

Abstract

The invention discloses a turn-to-turn arc severity lightweight discrimination system and method based on multiple physical quantities. The system comprises a multi-source signal acquisition and windowing module which is used for acquiring a multi-source physical signal of a multi-turn winding and dividing the multi-source physical signal into a plurality of sliding time windows; the full feature vector construction module is used for extracting multi-dimensional feature parameters from each physical signal for each sliding time window and splicing the multi-dimensional feature parameters into a full feature vector; the feature standardization and dimension reduction module is used for processing the full feature vector to obtain a multi-physical-quantity low-dimensional feature subset; the model training module is used for constructing and training a support vector machine discrimination model based on the multi-physical-quantity low-dimensional feature subset; and the lightweight discrimination module is used for completing lightweight discrimination of the arc discharge severity of the to-be-determined multi-source physical signal. The system has high identification accuracy and a stable discrimination result.

Description

Technical Field

[0001] This invention belongs to the technical field of online monitoring and intelligent fault severity judgment of inter-turn arc faults in power equipment windings in power systems, and relates to a lightweight judgment system and method for the severity of inter-turn arc faults based on multiple physical quantities. Background Technology

[0002] During long-term operation, the inter-turn insulation of oil-paper insulated electrical equipment (such as oil-immersed transformers and reactors) is simultaneously subjected to the coupled effects of multiple factors, including electric field stress, thermal stress, mechanical vibration, and aging of the oil-paper dielectric. This makes it prone to weak points in the insulation and induces inter-turn discharges. These discharges often exhibit an evolutionary pattern from weak to strong and from localized to widespread: initially, they may manifest as intermittent discharges or short-duration arcs, gradually transitioning to sustained arc discharges that may extend along the inter-turn path to multiple turns, ultimately leading to larger-scale inter-turn breakdowns or short circuits. Because multi-turn arc discharges are characterized by concentrated energy release, rapid development, and uncertain expansion paths, once they reach a moderate to severe stage, they significantly accelerate the deterioration of the oil-paper insulation and cause irreversible damage. Therefore, classifying the severity of inter-turn arc discharges is crucial for maintenance and early warning systems.

[0003] Furthermore, existing technologies for monitoring and identifying winding discharge faults often rely on single physical quantity signals (such as voltage / current only, acoustic signals only, or electromagnetic radiation signals only) to construct discrimination methods. However, arc discharge in oil is a typical multi-physical process, and different sensing links exhibit varying sensitivities to discharge energy release, channel morphology changes, and dielectric response. Single signals are easily affected by factors such as noise, installation location, propagation attenuation, and operating condition fluctuations, resulting in insufficient feature stability and weak stage separability, making it difficult to maintain reliable discrimination performance in complex scenarios such as multi-turn continuous breakdown. Meanwhile, while joint monitoring of multiple physical quantities helps improve discrimination accuracy, it also brings problems such as high feature dimensionality, excessive redundant information, and a large number of model parameters, making it difficult to deploy algorithms in real time at field terminals or edge devices. Especially in online early warning scenarios, excessive computational overhead of the discrimination model will affect the timeliness of alarms and engineering usability.

[0004] Therefore, how to construct a lightweight severity discrimination method with small model size and high discrimination accuracy by removing redundant information through feature selection and dimensionality reduction based on joint detection of multiple physical quantities is an urgent technical problem to be solved. Summary of the Invention

[0005] To overcome the above problems, this invention proposes a lightweight discrimination system and method for the severity of inter-turn arc based on multiple physical quantities.

[0006] Specifically, the object of the present invention is to provide the following aspects:

[0007] Firstly, a lightweight system for determining the severity of inter-turn arcs based on multiple physical quantities is provided, the system comprising:

[0008] The multi-source signal acquisition and windowing module is used to acquire multi-source physical signals of multi-turn windings during arc discharge in oil and divide them into multiple sliding time windows.

[0009] The full feature vector construction module is used to extract multi-dimensional feature parameters from each physical signal for each sliding time window to characterize the intensity and evolution state of arc discharge, and then concatenate them into a full feature vector.

[0010] The feature standardization and dimensionality reduction module is used to perform dimensionality-wise standardization on the full feature vector, and to perform feature filtering and dimensionality reduction on the standardized feature vector to obtain a low-dimensional feature subset of multiple physical quantities.

[0011] The model training module is used to construct and train a support vector machine discriminant model based on the low-dimensional feature subset of the multi-physical quantities.

[0012] The lightweight discrimination module is used to input the low-dimensional feature subset of multiple physical quantities obtained after processing the multi-source physical signals to be judged into the trained support vector machine discrimination model to complete the lightweight discrimination of the severity of arc discharge.

[0013] Optionally, in the multi-source signal acquisition and windowing module, the number of turns in the multi-turn winding is at least 4 turns and the last turn is grounded.

[0014] Optionally, in the multi-source signal acquisition and windowing module, the multi-source physical signals are transmitted to the computer via a data acquisition card, and the sampling rate of the data acquisition card is 60-65 MHz.

[0015] Optionally, in the multi-source signal acquisition and windowing module, the multi-source physical signals include at least the voltage signal of the ungrounded winding, the arc circuit current signal, the acoustic emission signal, and the ultra-high frequency electromagnetic radiation signal.

[0016] Optionally, in the full feature vector construction module, the multi-dimensional feature parameters include maximum discharge amplitude, standard deviation of discharge amplitude, average discharge amplitude, skewness, kurtosis, pulse repetition rate, spectral center frequency, spectral bandwidth, frequency band energy ratio, and frequency band energy ratio change rate.

[0017] Optionally, in the feature standardization and dimensionality reduction module, the dimension-by-dimensional standardization process includes: performing zero-mean, unit-variance standardization on each feature parameter in the full feature vector, so that the mean of each feature in the sample set is 0 and the standard deviation is 1.

[0018] Optionally, in the model training module, the input of the support vector machine discriminant model is a dataset composed of feature vectors from the low-dimensional feature subsets of multiple physical quantities corresponding to each sliding time window.

[0019] Secondly, a lightweight method for determining the severity of inter-turn arcs based on multiple physical quantities is provided, the method comprising:

[0020] Step 1: Obtain the multi-source physical signals of the multi-turn winding during arc discharge in oil, and divide them into multiple sliding time windows;

[0021] Step 2: For each sliding time window, extract multi-dimensional feature parameters from each physical signal to characterize the arc discharge intensity and evolution state, and concatenate them into a full feature vector;

[0022] Step 3: Perform dimension-wise standardization on the full feature vector, and perform feature filtering and dimensionality reduction on the standardized feature vector to obtain a low-dimensional feature subset of multiple physical quantities.

[0023] Step 4: Based on the low-dimensional feature subset of the multi-physical quantities, construct and train a support vector machine discriminant model;

[0024] Step 5: Input the low-dimensional feature subset of multiple physical quantities obtained after processing in steps 1-3 into the trained support vector machine discrimination model to complete the lightweight discrimination of arc discharge severity.

[0025] Optionally, step 1 includes:

[0026] Step 1-1: Obtain multi-source physical signals of the multi-turn winding during arc discharge in oil;

[0027] Steps 1-2 involve denoising the multi-source physical signals and dividing each physical signal into multiple sliding time windows based on a unified time reference.

[0028] Thirdly, a computer-readable storage medium is provided, including a stored complete computer program that, when the computer program is run, implements the method described in the first aspect.

[0029] The beneficial effects of this invention include: It simultaneously acquires monitoring signals of multiple physical quantities such as voltage, current, acoustic emission, and ultra-high frequency electromagnetic radiation; it employs a sliding time window to segment and extract multi-dimensional feature parameters; based on this, it performs feature filtering and dimensionality reduction to form a low-dimensional feature subset; and trains a support vector machine model to achieve graded discrimination of the severity of multi-turn arc discharge. Because the information from multiple physical quantities is complementary and redundancy is effectively suppressed, this invention can achieve higher recognition accuracy and more stable discrimination results in multi-turn evolution scenarios, while also achieving a lightweight discrimination process, facilitating engineering deployment and application promotion. Attached Figure Description

[0030] Various other advantages and benefits of the present invention will become apparent to those skilled in the art upon reading the detailed description of the preferred embodiments below. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. It is obvious that the drawings described below are merely some embodiments of the invention, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0031] In the attached diagram:

[0032] Figure 1 The flowchart of the present invention, which is a lightweight method for determining the severity of inter-turn arcs based on multiple physical quantities, is shown.

[0033] Figure 2 The Pearson correlation coefficient matrix obtained in Example 1 is shown in the figure.

[0034] Figure 3 This shows a confusion matrix diagram illustrating the accuracy of discharge severity identification under the input of a low-dimensional feature subset of multiple physical quantities in Example 1.

[0035] Figure 4 This shows a confusion matrix diagram illustrating the accuracy of discharge severity identification under electrical signal input conditions in Example 1.

[0036] Figure 5 This shows a confusion matrix diagram illustrating the accuracy of discharge severity identification under acoustic emission signal input conditions in Example 1.

[0037] Figure 6 This shows a confusion matrix diagram illustrating the accuracy of discharge severity identification under UHF electromagnetic radiation signal input conditions in Example 1.

[0038] Figure 7 This invention illustrates a lightweight discrimination system for the severity of inter-turn arcs based on multiple physical quantities.

[0039] Figure 8 The diagram shows the structure of the support vector machine discriminant model of the present invention. Detailed Implementation

[0040] The following will refer to the appendix. Figures 1 to 8 Specific embodiments of the invention will be described in more detail below. While specific embodiments of the invention are shown in the accompanying drawings, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the invention and to fully convey the scope of the invention to those skilled in the art.

[0041] It should be noted that certain terms are used in the specification and claims to refer to specific components. Those skilled in the art will understand that different terms may be used to refer to the same component. This specification and claims do not distinguish components based on differences in terminology, but rather on differences in function. The terms "comprising" or "including" used throughout the specification and claims are open-ended and should be interpreted as "comprising but not limited to." The following descriptions are preferred embodiments for carrying out the invention; however, these descriptions are for the purpose of understanding the general principles of the specification and are not intended to limit the scope of the invention. The scope of protection of this invention is determined by the appended claims.

[0042] In the description of this invention, it should be noted that the terms "upper," "lower," "inner," "outer," "front," and "rear," etc., indicate the orientation or positional relationship based on the orientation or positional relationship in the working state of this invention, and are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this invention. Furthermore, the terms "first," "second," "third," and "fourth" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0043] To facilitate understanding of the embodiments of the present invention, further explanations and descriptions will be provided below with reference to the accompanying drawings and specific embodiments. The accompanying drawings do not constitute a limitation on the embodiments of the present invention.

[0044] On the one hand, the present invention provides a lightweight method for determining the severity of inter-turn arcs based on multiple physical quantities, such as... Figure 1 As shown, the method includes:

[0045] Step 1: Obtain the multi-source physical signals of the multi-turn winding during arc discharge in oil, and divide them into multiple sliding time windows;

[0046] Step 2: For each sliding time window, extract multi-dimensional feature parameters from each physical signal to characterize the arc discharge intensity and evolution state, and concatenate them into a full feature vector;

[0047] Step 3: Perform dimension-wise standardization on the full feature vector, and perform feature filtering and dimensionality reduction on the standardized feature vector to obtain a low-dimensional feature subset of multiple physical quantities.

[0048] Step 4: Based on the low-dimensional feature subset of the multi-physical quantities, construct and train a support vector machine discriminant model;

[0049] Step 5: Input the low-dimensional feature subset of multiple physical quantities obtained after processing the multi-source physical signals to be judged in steps 1-3 into the trained support vector machine discrimination model to complete the lightweight discrimination of the severity of inter-turn arc.

[0050] The above methods will be described in detail below.

[0051] Step 1: Obtain the multi-source physical signals of the multi-turn winding during arc discharge in oil and divide them into multiple sliding time windows.

[0052] According to a preferred embodiment, step 1 includes:

[0053] Step 1-1: Obtain multi-source physical signals of the multi-turn winding during arc discharge in oil;

[0054] Steps 1-2 involve denoising the multi-source physical signals and dividing each physical signal into multiple sliding time windows based on a unified time reference.

[0055] In step 1-1, the multi-turn winding is made of metal wire, with at least 4 turns and the last turn grounded, and the turn gap is 3-5mm; the number of turn gaps is the number of turns minus 1; preferably, if the number of turns is 4-7, then the corresponding number of turn gaps is 3-6.

[0056] In one embodiment, the number of turns is 4, resulting in 3 inter-turn gaps. Different discharge severity levels are defined based on the breakdown status of these inter-turn gaps. Definitions: 0 inter-turn gaps breaking down indicates no breakdown; 1 inter-turn gap breaking down indicates single-gap breakdown; 2 inter-turn gaps breaking down indicates double-gap breakdown; and 3 inter-turn gaps breaking down indicates through-breakdown. Of course, the processing approach remains the same when the number of turns is not 4.

[0057] In step 1-1, all multi-source physical signals are transmitted to the computer via a data acquisition card. The sampling rate of the data acquisition card is 60-65 MHz, for example, 62.5 MHz, to ensure that transient characteristics during the arc discharge process can be effectively captured.

[0058] In step 1-1, the multi-source physical signals include at least the voltage signal of the ungrounded winding, the arc circuit current signal, the acoustic emission signal, and the ultra-high frequency electromagnetic radiation signal.

[0059] Specifically, the voltage signal is measured using a high-voltage probe, the arc discharge circuit current signal is measured using a Rogowski coil, the acoustic emission signal is collected using an acoustic emission sensor (coupled with an acoustic coupling agent and the transformer shell / test tank wall), and the ultra-high frequency electromagnetic radiation signal is collected using an ultra-high frequency / ultra-high frequency antenna. The voltage of the high-voltage probe is typically 0 (excluding) - 100kV, and the frequency of the ultra-high frequency electromagnetic radiation signal is typically 1.5MHz - 30MHz.

[0060] In steps 1-2, the multi-source physical signals are denoised, and each physical signal is divided into multiple sliding time windows based on a unified time reference. This denoising process effectively suppresses the effects of power frequency interference, random noise, and sensor drift while preserving the transient characteristics of arc discharge, thereby improving the accuracy of subsequent feature extraction and the robustness of model discrimination.

[0061] Specifically, using the trigger time or synchronization clock as a unified time reference, all physical signals are aligned to the same zero time to ensure precise synchronization of multi-source physical signals in the time dimension; subsequently, wavelet thresholding is used for noise reduction.

[0062] For voltage signals and arc circuit current signals, due to DC offset during acquisition, DC component removal processing is performed first, followed by wavelet threshold denoising. For acoustic emission signals and UHF electromagnetic radiation signals, since they are AC transient signals without significant DC components, wavelet threshold denoising is performed directly. This differentiated processing avoids frequency domain feature distortion caused by DC component differences between different signal types, ensures time alignment and amplitude consistency of multi-physical quantity signals in subsequent feature extraction and fusion analysis, and improves the reliability of multi-modal data collaborative discrimination.

[0063] Furthermore, wavelet thresholding is applied to each physical signal for denoising, preferably using Daubechies wavelets, such as db4 or db6, with a corresponding decomposition layer of 4 or 6 layers to balance time resolution and frequency resolution.

[0064] In one implementation, wavelet transform processing is performed on voltage signals, arc circuit current signals, acoustic emission signals, and ultra-high frequency electromagnetic radiation signals, respectively. Each signal type corresponds to an independent channel, for a total of four channels.

[0065] For voltage signals and arc circuit current signals, DC component removal is first performed to obtain discrete signal sequences after DC removal. For acoustic emission signals and ultra-high frequency electromagnetic radiation signals, their original discrete signal sequences are denoted as... That is, the input signals of all channels are uniformly represented as discrete sequences. , N represents the number of data points involved in denoising; when using a sliding time window, N is the number of sampling points within a single sliding time window.

[0066] right Based on the db4-level discrete wavelet decomposition, a set of wavelet coefficients is obtained: the fourth-level approximation coefficients. and the detail coefficients of each layer ,in This represents the low-frequency smoothing component of the signal (i.e., the baseline trend portion). Let represent the detail coefficients of the j-th level, and The corresponding highest frequency detail subband mainly contains the transient pulse energy generated by arc discharge.

[0067] In steps 1-2, a soft thresholding function (threshold) is used. ) detail coefficient ( Threshold shrinkage is performed to obtain the detail coefficients after denoising. .

[0068] The threshold can be either a general threshold or an adaptive threshold based on noise variance estimation.

[0069] In one implementation, a universal threshold method is used to determine the threshold. This is to achieve effective denoising of arc discharge signals under different signal-to-noise ratio conditions. Specifically:

[0070] First, using the level 1 detail factor Estimating the standard deviation of noise .set up Indicates to Take the absolute value of each coefficient. This indicates that the noise standard deviation is calculated by taking the median. Median absolute deviation estimation: In the formula, the constant 0.6745 is a conversion factor under the condition that the noise approximately follows a zero-mean Gaussian distribution, used to convert the absolute deviation of the median into the standard deviation. This method does not rely on a priori noise model, but estimates the noise level only through the statistical properties of the signal itself, and has good robustness and universality.

[0071] Then, according to the above Calculate threshold : In the formula, N represents the number of data points involved in denoising, and in this invention, the number of sampling points within a single sliding time window is fixed. This threshold setting method comprehensively considers data length and noise energy, effectively suppressing random noise while preserving the transient characteristics of useful signals.

[0072] Furthermore, the soft threshold function is defined as: In the formula, The sign function is set to 1 when the input is greater than 0, -1 when it is less than 0, and 0 when it is equal to 0.

[0073] Finally, the fourth-level approximation coefficients are preserved. Unchanged, will and Perform inverse wavelet reconstruction to obtain the denoised signal. and will This serves as the input signal for subsequent sliding time window feature extraction.

[0074] In steps 1-2, the sliding time window adopts a window length. With step size The sliding window method, and This is used to create a window sequence where adjacent windows overlap. Window length The preferred step size is 5-15ms to cover the duration of a typical arc discharge pulse, ensuring that a single discharge event falls completely within a window; The preferred timeframe is 2-10 ms, which is less than the window length. This allows adjacent windows to overlap, improving temporal resolution and preventing the omission of short-term discharge events due to window misalignment. For example... Set to 10ms. The time interval is set to 4ms, and a window sequence is generated accordingly.

[0075] Step 2: For each sliding time window, extract multi-dimensional feature parameters from each physical signal to characterize the arc discharge intensity and evolution state, and concatenate them into a full feature vector.

[0076] In step 2, the multi-dimensional feature parameters include maximum discharge amplitude, standard deviation of discharge amplitude, average discharge amplitude, skewness, kurtosis, pulse repetition rate, spectral center frequency, spectral bandwidth, frequency band energy ratio, and frequency band energy ratio change rate.

[0077] For the voltage signal, since the voltage of the windings other than the ground turn is measured, at least three characteristic parameters are extracted for each voltage signal within the sliding time window, namely the maximum discharge amplitude, the standard deviation of the discharge amplitude, and the kurtosis.

[0078] For the arc discharge circuit current signal, only the maximum discharge amplitude is extracted as a 1-dimensional feature parameter within the sliding time window. This parameter directly reflects the instantaneous current intensity in the arc channel and is an important basis for judging the severity of the discharge.

[0079] For the acoustic emission signal, general time-domain and frequency-domain features and multi-band energy distribution features are extracted within a sliding time window. The general time-domain and frequency-domain features include maximum discharge amplitude, standard deviation of discharge amplitude, average discharge amplitude, skewness, kurtosis, pulse repetition rate, spectral center frequency, and spectral bandwidth. The multi-band energy distribution features include band energy proportion and band energy proportion change rate: the acoustic emission signal spectrum is divided into three sub-bands: the audible band, the 75-100kHz band, and the remaining bands within the 0-150kHz range excluding the audible band and the 75-100kHz band. For each sub-band, the band energy proportion and band energy proportion change rate are extracted.

[0080] For ultra-high frequency electromagnetic radiation signals, general time-domain and frequency-domain features and multi-band energy distribution features are extracted within a sliding time window. The general time-domain and frequency-domain features include maximum discharge amplitude, standard deviation of discharge amplitude, average discharge amplitude, skewness, kurtosis, pulse repetition rate, spectral center frequency, and spectral bandwidth. The multi-band energy distribution features include band energy proportion and band energy proportion change rate: the acoustic emission signal spectrum is divided into three sub-bands, namely the 2-5MHz band, the 9-17MHz band, and the remaining bands within the 0-30MHz range excluding the 2-5MHz and 9-17MHz bands. For each sub-band, the band energy proportion and band energy proportion change rate are extracted.

[0081] In step 2, all extracted multi-dimensional feature parameters are aligned according to time and concatenated into a full feature vector.

[0082] Step 3: Perform dimension-wise standardization on the full feature vector, and perform feature filtering and dimensionality reduction on the standardized feature vector to obtain a low-dimensional feature subset of multiple physical quantities.

[0083] In step 3, the dimension-by-dimensional standardization process includes: performing zero-mean, unit-variance standardization on each feature parameter in the full feature vector, so that the mean of each feature in the sample set is 0 and the standard deviation is 1.

[0084] Specifically, for the first The feature in the first The values ​​in each sample Its standardized value The calculation formula is: In the formula, Let k be the value of the k-th feature in the i-th sample; are the mean and standard deviation of the entire sample for the k-th feature, respectively; This represents the standardized value of the k-th feature in the i-th sample. This process eliminates the dimensional differences between different features, improving the stability of subsequent model training.

[0085] In step 3, the feature selection includes: calculating the separability score for each feature dimension using the Fisher discriminant criterion, sorting the features by score, and retaining the top K features. The specific formula is: In the formula, k is the feature number, c corresponds to the discharge severity category, and C is the total number of discharge severity stages (for example, if 4 types of discharge severity for inter-turn breakdown are defined, then C=4). The number of samples in class c. These are the mean and standard deviation of the k-th feature in the c-th class, respectively. This is the mean of the feature across all samples. The Fisher score for this feature.

[0086] When using Fisher's discrimination criterion for feature selection, K typically retains 70%-90% of the total feature dimension to balance feature information content and redundancy suppression. For example, if the original feature count is 38 dimensions, after sorting by Fisher scores, the top 32 features are retained, which is equivalent to retaining approximately 85% of the highest-scoring features.

[0087] According to Fisher's score principle, the greater the inter-class variance and the smaller the intra-class variance, the more discriminative the feature. Features are determined based on... Sort the features from largest to smallest to obtain a list of feature priorities that are most sensitive to severity.

[0088] In step 3, although the Fisher criterion can measure the discriminative power of each feature, there may still be a high correlation between different features. Therefore, in order to reduce the input dimension and avoid multicollinearity, this invention uses the Pearson correlation coefficient to analyze the linear correlation between features, thereby achieving feature dimensionality reduction.

[0089] Specifically, the feature dimensionality reduction includes: calculating the Pearson correlation coefficient for the retained features; when the absolute value of the correlation coefficient between any two features is greater than or equal to a preset threshold (e.g., 0.9), it is considered that there is a high linear correlation, and only the features with higher Fisher scores are retained, while the features with lower Fisher scores are removed to eliminate redundant information. The specific formula is: In the formula, Let N be the Pearson correlation coefficient between the k-th and l-th features, and N be the sample size. , Let represent the mean values ​​of the k-th and l-th feature parameters across all N samples, respectively.

[0090] In step 3, if the absolute correlation coefficient between a certain feature and the selected features meets a preset threshold, that is... If the information is considered highly redundant, only features with relatively higher Fisher scores are retained, while features with relatively lower Fisher scores are discarded. Through a dual selection strategy of "prioritizing high discriminative ability and eliminating strongly correlated features," a subset of features with low dimensionality but rich information can be obtained, providing high-quality input for subsequent severity identification models.

[0091] In step 3, after the standardization, feature selection, and dimensionality reduction processes described above, a low-dimensional feature subset of multiple physical quantities is obtained for subsequent model input. This subset has a lower dimension than the original full feature vector and retains feature information with high discriminative power regarding the severity of discharge. The low-dimensional feature subset includes at least the maximum discharge amplitude, standard deviation of discharge amplitude, average discharge amplitude, skewness, kurtosis, pulse repetition rate, spectral center frequency, spectral bandwidth, frequency band energy proportion, and frequency band energy proportion change rate. The feature vectors from all windows are aggregated to form a dataset for model training. Ultimately, the feature parameters involved in the identification effectively compress the original feature space, improving the model's lightweight nature.

[0092] In one embodiment, the multi-turn winding has 4 turns and 3 turn gaps. For the voltage signal, the maximum discharge amplitude, standard deviation of discharge amplitude, and kurtosis of each winding voltage are extracted, resulting in a total of 9-dimensional feature parameters. For the arc discharge circuit current signal, the maximum discharge amplitude is extracted, resulting in a total of 1-dimensional feature parameter. For the acoustic emission signal, in addition to extracting 8 dimensions such as maximum discharge amplitude, standard deviation of discharge amplitude, average discharge amplitude, skewness, kurtosis, pulse repetition rate, spectral center frequency, and spectral bandwidth, the acoustic emission signal spectrum is divided into three sub-bands: the audible band, the 75-100kHz band, and the remaining bands in the 0-150kHz range excluding the audible band and the 75-100kHz band. For each sub-band, the band energy percentage and the band energy percentage change rate are extracted, resulting in a total of 6 dimensions. Therefore, a total of 14-dimensional feature parameters are extracted from the acoustic emission signal. For UHF electromagnetic radiation signals, in addition to extracting eight dimensions—maximum discharge amplitude, standard deviation of discharge amplitude, average discharge amplitude, skewness, kurtosis, pulse repetition rate, spectral center frequency, and spectral bandwidth—the acoustic emission signal spectrum is divided into three sub-bands: the 2-5MHz band, the 9-17MHz band, and the remaining band within the 0-30MHz range excluding the 2-5MHz and 9-17MHz bands. For each sub-band, the band energy percentage and the rate of change of the band energy percentage are extracted, resulting in six dimensions. Therefore, a total of 14 characteristic parameters are extracted for UHF electromagnetic radiation signals. In summary, a total of 38 characteristic parameters are initially extracted.After dimensional standardization, Fisher's discrimination criterion, and Pearson correlation analysis, its dimensions are lower than the original full feature vector, totaling 23 dimensions, including: UA_max (maximum discharge amplitude of voltage UA signal), UHF_std (standard deviation of electromagnetic radiation signal), UHF_bw (spectral bandwidth of electromagnetic radiation), AE_pctA (energy proportion of A-band of acoustic signal), UHF_fc (center frequency of electromagnetic radiation signal spectrum), AE_fc (center frequency of acoustic signal spectrum), UHF_pctR (energy proportion of other frequency bands of electromagnetic radiation), UHF_PRR (pulse repetition rate of electromagnetic radiation), UA_std (standard deviation of voltage UA signal), UC_ max (maximum discharge amplitude of voltage UC signal), UC_std (standard deviation of voltage UC signal), AE_pctL (energy percentage of L band of acoustic signal), UHF_pctB (energy percentage of B band of electromagnetic radiation signal), AE_bw (spectral bandwidth of acoustic signal), AE_max (maximum discharge amplitude of acoustic signal), I_max (current amplitude), UHF_kurt (kurtosis of electromagnetic radiation signal), AE_kurt (kurtosis of acoustic signal), UHF_skew (skewness of electromagnetic radiation signal), UHF_pctA (energy percentage of A band of electromagnetic radiation signal), AE_pctR (energy percentage of R band of acoustic signal), UH_max (maximum discharge amplitude of voltage UH signal), UHF_dpctR (energy change rate of R band of electromagnetic radiation signal).

[0093] Wherein, max corresponds to the maximum discharge amplitude, mean corresponds to the average discharge amplitude, std corresponds to the standard deviation of the discharge amplitude, PRR corresponds to the pulse repetition rate, kurt corresponds to the kurtosis, skew corresponds to the skewness, fc corresponds to the center frequency of the spectrum, bw corresponds to the bandwidth of the spectrum, pctX corresponds to the energy proportion of the preset frequency band X, dpctX corresponds to the rate of change of the energy proportion of the preset frequency band X, UHF corresponds to the ultra-high frequency electromagnetic radiation signal, AE corresponds to the acoustic emission signal, I corresponds to the arc discharge circuit current signal, and UA, UC, and UH correspond to the voltage signals of the remaining 3 turns (excluding the grounding turn) when the number of turns is 4. For example, UA_max represents the maximum amplitude of the voltage signal of the first winding, and UHF_std represents the standard deviation of the ultra-high frequency electromagnetic radiation signal.

[0094] In pctX, AE_pctL represents the energy percentage of the audible frequency band of the sound signal, AE_pctH represents the energy percentage of the 75-100kHz frequency band of the sound signal, AE_pctR represents the energy percentage of the remaining frequency band of the sound signal within the 0-150kHz range excluding the audible and 75-100kHz ranges, UHF_pctA represents the energy percentage of the electromagnetic radiation signal within the 2-5MHz frequency band, UHF_pctB represents the energy percentage of the electromagnetic radiation signal within the 9-17MHz frequency band, and UHF_pctR represents the energy percentage of the remaining frequency band of the electromagnetic radiation signal within the 0-30MHz range excluding the 2-5MHz and 9-17MHz frequency bands. Similarly, in dpctX, the letters for different frequency bands of sound signals and electromagnetic radiation signals have the same meaning.

[0095] Step 4: Based on the low-dimensional feature subset of the multi-physical quantities, construct and train a support vector machine discriminant model.

[0096] In step 4, the input to the support vector machine discriminant model is a dataset composed of feature vectors from the low-dimensional feature subsets of multiple physical quantities corresponding to each sliding time window.

[0097] First, the dataset is divided into training and validation sets in an 8:2 or 7:3 ratio, and a stratified sampling strategy is used to ensure that the proportion of each discharge severity category is consistent between the training and validation sets. Then, k-fold cross-validation is used on the training set to find the optimal hyperparameters, with 5-fold cross-validation being preferred. Finally, using these optimal hyperparameters, a support vector machine (SVM) discriminant model is constructed and trained on the complete training set. In practical engineering implementation, to prevent overfitting, the model performance is periodically evaluated on the validation set during training (every 8-12 iterations). Training is terminated early when the validation set recognition accuracy no longer significantly improves after 15 consecutive parameter updates (the improvement is less than 0.0001) or when the maximum number of iterations (5000) is reached.

[0098] When using k-fold cross-validation on the training set to find the optimal hyperparameters, the preset penalty coefficient C and kernel parameter... A grid search is performed within the specified range, and the average recognition accuracy of cross-validation is used as the evaluation metric to select the optimal performance. The combination is used as the final hyperparameter. The range of the penalty coefficient C is preferably [defined here]. nuclear parameters The preferred range .

[0099] Furthermore, such as Figure 8 As shown, the SVM model includes an input layer, a kernel mapping layer, and an output discriminant layer. Specifically:

[0100] Input layer: Receives feature vectors (feature vectors in a low-dimensional feature subset of multiple physical quantities) and passes the vectors to the kernel mapping layer;

[0101] Kernel mapping layer: Based on the optimal kernel parameter γ selected in advance through k-fold cross-validation, it implicitly maps the input feature vector to the regeneration kernel Hilbert space using a radial basis function kernel, and constructs the kernel matrix through kernel function calculation; the final result obtained by this layer is a kernel matrix describing the similarity relationship between samples, making the originally linearly inseparable discharge severity categories approximately linearly separable in this space;

[0102] Output discriminant layer: Based on the optimal penalty coefficient C selected in advance through k-fold cross-validation, it uses the Sequence Minimum Optimization (SMO) algorithm to iteratively update the model parameters. It achieves model convergence by maximizing the classification margin (i.e., maximizing the minimum geometric distance between each category while ensuring that each discharge severity category sample is correctly classified) and suppressing training error, and finally trains the support vector machine discriminant model.

[0103] In step 4, after the model converges, the final output performance metrics include at least the accuracy of discharge severity discrimination, training time, single-window inference time, parameter size, model size, and computational cost.

[0104] Preferably, the severity of the discharge is manually labeled and includes at least the non-breakdown state, single-gap breakdown state, double-gap breakdown state, and through-breakdown state.

[0105] In contrast, this invention divides a dataset composed of all feature vectors into training and validation sets to construct and train a support vector machine (SVM) discriminant model. The input is the dataset composed of all feature vectors corresponding to each sliding time window. The SVM model based on the full feature vector input and the SVM model based on the low-dimensional feature subset input of multiple physical quantities are structurally identical, differing only in input dimension, and are used to compare and analyze the impact of feature dimensionality reduction on recognition performance. Since the full feature vector contains multi-dimensional representation information of all original physical signals, its corresponding model can be considered as the performance ceiling under the feature system extracted in this invention. By comparing the recognition accuracy, single-window inference time, and model parameter size of the full feature model and the low-dimensional feature model of multiple physical quantities on the same test set, it can be objectively verified that the proposed feature selection and dimensionality reduction strategy effectively retains key discriminant information while significantly reducing input dimensionality and computational complexity, thus proving that the lightweight discriminant scheme proposed in this invention has good engineering practical value.

[0106] Step 5: Input the low-dimensional feature subset of multiple physical quantities obtained after processing in steps 1-3 into the trained support vector machine discrimination model to complete the lightweight discrimination of arc discharge severity.

[0107] In practical applications, newly acquired multi-source physical signals are aligned, denoised, windowed, feature-extracted, standardized, and dimensionality-reduced according to the methods described in steps 1 to 3 to obtain a low-dimensional feature subset of multi-physical quantities corresponding to each time window. Subsequently, this feature subset is input into the support vector machine discrimination model trained in step 4 based on the low-dimensional feature subset of multi-physical quantities. The model outputs the severity category of arc discharge within the time window, realizing real-time discrimination of the discharge state.

[0108] To verify the effectiveness of multi-physical quantity information fusion, a comparative analysis was further conducted: single physical quantity signals containing only electrical signals (voltage and current), only acoustic emission signals, and only ultra-high frequency electromagnetic radiation signals were processed using the same procedure as for multi-physical quantities, performing dimension-wise standardization, feature selection, and dimensionality reduction to obtain their respective low-dimensional feature subsets; and based on each subset, support vector machine discrimination models with the same structure were independently trained (using the same kernel function, hyperparameter optimization strategy, and training procedure as the main model) to obtain their respective discrimination results.

[0109] By comparing the discrimination performance under different input conditions, it can be seen that the joint input of multiple physical quantities can make full use of the complementarity of voltage and current, acoustic signals and electromagnetic radiation information in terms of discharge intensity characterization and stage evolution characterization. Under the same training process and the same low-dimensional feature screening and dimensionality reduction strategy, it can significantly improve the severity classification discrimination ability. Compared with single physical quantity input, the recognition accuracy of joint input of multiple physical quantities is improved by about 11-20 percentage points, and the confusion between stages is significantly reduced, especially the better differentiation effect of adjacent stages. This verifies the effectiveness and superiority of the present invention in using multiple physical quantity detection and lightweight feature subset for severity discrimination.

[0110] Meanwhile, this invention also supports the use of a low-dimensional subset of features of a single physical quantity for discrimination in scenarios where resources are limited or sensors are partially failed, so as to achieve degraded operation and system fault tolerance, and improve the engineering applicability and deployment flexibility of the solution.

[0111] In this invention, multiple physical quantity monitoring signals, such as voltage, current, acoustic emission, and ultra-high frequency electromagnetic radiation, are acquired synchronously. A sliding time window is used for segmentation, and multi-dimensional time-frequency domain feature parameters characterizing the intensity and evolution of arc discharge are extracted from each physical channel to construct a high-dimensional full feature vector. Based on this, a dual feature filtering mechanism based on Fisher's discrimination criterion and Pearson correlation analysis is introduced to effectively eliminate feature dimensions with large intra-class dispersion, low inter-class discriminability, and high redundancy, ultimately forming a low-dimensional feature subset that is information-dense and dimensionally concise. This subset is then used as input to train a lightweight support vector machine discrimination model, achieving high-precision classification and discrimination of the severity of arc discharge in multi-turn oil (e.g., non-breakdown, single-gap breakdown, double-gap breakdown, and through-breakdown).

[0112] By fully utilizing the complementary characteristics of different physical quantities in terms of discharge energy release, channel morphology evolution, and dielectric response, and by significantly suppressing information redundancy through a scientific feature dimensionality reduction strategy, this invention not only achieves higher recognition accuracy (up to 95.00%) in complex multi-turn evolution scenarios, but also exhibits stronger stage discrimination capability and judgment stability. Simultaneously, the model input dimension is significantly reduced, for example from 38 dimensions to 23 dimensions, resulting in a simultaneous reduction in inference computation and storage overhead. This achieves a lightweight and efficient discrimination process, meeting the deployment requirements of edge-side or embedded monitoring terminals for real-time performance, low power consumption, and small model size, demonstrating good engineering applicability and promising prospects for widespread application.

[0113] On the other hand, according to the present invention, a lightweight discrimination system for the severity of inter-turn arcs based on multiple physical quantities is provided, such as... Figure 7 As shown, the system includes:

[0114] The multi-source signal acquisition and windowing module is used to acquire multi-source physical signals of multi-turn windings during arc discharge in oil and divide them into multiple sliding time windows.

[0115] The full feature vector construction module is used to extract multi-dimensional feature parameters from each physical signal for each sliding time window to characterize the intensity and evolution state of arc discharge, and then concatenate them into a full feature vector.

[0116] The feature standardization and dimensionality reduction module is used to perform dimensionality-wise standardization on the full feature vector, and to perform feature filtering and dimensionality reduction on the standardized feature vector to obtain a low-dimensional feature subset of multiple physical quantities.

[0117] The model training module is used to construct and train a support vector machine discriminant model based on the low-dimensional feature subset of the multi-physical quantities.

[0118] The lightweight discrimination module is used to input the low-dimensional feature subset of multiple physical quantities obtained after processing the multi-source physical signals to be judged into the trained support vector machine discrimination model to complete the lightweight discrimination of the severity of inter-turn arc.

[0119] It should be understood that the system and the method are consistent in principle, and the specific implementation of the method is applicable to the system and produces the same technical effect, which will not be elaborated here.

[0120] In another aspect, the computer-readable storage medium provided by the present invention includes a stored complete computer program that, when the computer program is run, implements the method as described in the first aspect. Example

[0121] The present invention is further described below through specific examples; however, these examples are merely exemplary and do not constitute any limitation on the scope of protection of the present invention. Example 1

[0122] To verify the effectiveness of the present invention, a multi-turn winding inter-turn discharge simulation platform was built: a 4-turn arc discharge (the last turn is grounded) was used to form 3 turn gaps, and four levels of discharge severity were defined accordingly: 0 turn gaps breaking down is the non-breakdown state, 1 turn gap breaking down is the single gap breakdown state, 2 turn gaps breaking down is the double gap breakdown state, and 3 turn gaps breaking down is the through breakdown state.

[0123] During the discharge process, voltage signals, current signals, acoustic emission signals (frequency 1.5MHz-30MHz), and ultra-high frequency electromagnetic radiation signals are collected as multi-source physical signals. All signals are transmitted to a computer via a data acquisition card with a sampling rate of 62.5MHz. The voltage signal is acquired using a 25kV high-voltage probe, the current signal is acquired using a Hall coil, the acoustic emission signal is acquired using an acoustic emission sensor, and the ultra-high frequency electromagnetic radiation signal is acquired using an ultra-high frequency antenna.

[0124] All signals are decomposed into 4-level discrete wavelet decomposition using db4 wavelet, and the window length is used. Set to 10ms, step size Set the time to 4ms to obtain the window sequence.

[0125] For each window, the voltage signal is used to extract the maximum discharge amplitude, standard deviation of discharge amplitude, and kurtosis of each winding voltage. Since the voltage signals of turns other than the grounding turn are measured, there are three voltage signals in total, and each signal is used to acquire 3-dimensional feature parameters, for a total of 9-dimensional feature parameters. For the arc discharge circuit current signal, the maximum discharge amplitude is extracted, for a total of 1-dimensional feature parameter. For the acoustic emission signal, in addition to extracting 8 dimensions such as maximum discharge amplitude, standard deviation of discharge amplitude, average discharge amplitude, skewness, kurtosis, pulse repetition rate, spectral center frequency, and spectral bandwidth, the acoustic emission signal spectrum is divided into three sub-bands: the audible band, the 75-100kHz band, and the remaining band in the 0-150kHz range excluding the audible band and the 75-100kHz band. For each sub-band, the band energy percentage and the band energy percentage change rate are extracted, for a total of 6 dimensions, that is, a total of 14-dimensional feature parameters are extracted for the acoustic emission signal. For UHF electromagnetic radiation signals, in addition to extracting eight dimensions—maximum discharge amplitude, standard deviation of discharge amplitude, average discharge amplitude, skewness, kurtosis, pulse repetition rate, spectral center frequency, and spectral bandwidth—the acoustic emission signal spectrum is divided into three sub-bands: the 2-5MHz band, the 9-17MHz band, and the remaining band within the 0-30MHz range excluding the 2-5MHz and 9-17MHz bands. For each sub-band, the band energy percentage and the rate of change of the band energy percentage are extracted, resulting in six dimensions. Therefore, a total of 14 characteristic parameters are extracted for UHF electromagnetic radiation signals. In summary, this embodiment initially extracts 38 characteristic parameters.

[0126] All extracted multi-dimensional feature parameters were aligned by time and concatenated into a full feature vector. Then, dimension-wise standardization was performed, and the vectors were sorted in descending order based on the Fisher criterion. The results are shown in Table 1. Wherein, max corresponds to the maximum discharge amplitude, mean corresponds to the average discharge amplitude, std corresponds to the standard deviation of the discharge amplitude, PRR corresponds to the pulse repetition rate, kurt corresponds to the kurtosis, skew corresponds to the skewness, fc corresponds to the spectral center frequency, bw corresponds to the spectral bandwidth, pctX corresponds to the energy proportion of the preset frequency band X, dpctX corresponds to the rate of change of the frequency band energy proportion of the preset frequency band X, UHF corresponds to the ultra-high frequency electromagnetic radiation signal, AE corresponds to the acoustic emission signal, I corresponds to the arc discharge circuit current signal, and UA, UC, and UH correspond to the voltage signals of the remaining 3 turns (excluding the grounding turn) when the number of turns is 4 in this embodiment. For example, UA_max represents the maximum amplitude of the voltage signal of the first winding, and UHF_std represents the signal standard deviation of the ultra-high frequency electromagnetic radiation signal.

[0127] In pctX, AE_pctL represents the energy percentage of the audible frequency band of the sound signal, AE_pctH represents the energy percentage of the 75-100kHz frequency band of the sound signal, AE_pctR represents the energy percentage of the remaining frequency band of the sound signal within the 0-150kHz range excluding the audible and 75-100kHz ranges, UHF_pctA represents the energy percentage of the electromagnetic radiation signal within the 2-5MHz frequency band, UHF_pctB represents the energy percentage of the electromagnetic radiation signal within the 9-17MHz frequency band, and UHF_pctR represents the energy percentage of the remaining frequency band of the electromagnetic radiation signal within the 0-30MHz range excluding the 2-5MHz and 9-17MHz frequency bands. Similarly, in dpctX, the letters for different frequency bands of sound signals and electromagnetic radiation signals have the same meaning.

[0128] Table 1:

[0129]

[0130] Features with a discrimination score less than 0.05 were discarded, and 32 feature parameters with a score greater than 0.05 were selected for further calculation of the Pearson correlation coefficient, resulting in the correlation matrix as follows. Figure 2As shown in Table 2, the pairwise correlation coefficients between the 32 different feature parameters are different. Specifically, the coefficients and selection of highly correlated feature pairs are shown in Table 2. For example, the correlation coefficients between UA_max and UC_max, and I_max are 0.916 and -0.936, respectively, both with absolute values ​​greater than 0.9, while only UA_max, which has a higher Fisher score, is retained. In summary, the feature parameters ultimately involved in the discharge severity identification in this invention total 23 dimensions, including: UA_max (maximum discharge amplitude of voltage UA signal), UHF_std (standard deviation of electromagnetic radiation signal), UHF_bw (electromagnetic radiation spectrum bandwidth), AE_pctA (energy proportion of A-band of acoustic signal), UHF_fc (center frequency of electromagnetic radiation signal spectrum), AE_fc (center frequency of acoustic signal spectrum), UHF_pctR (energy proportion of other frequency bands of electromagnetic radiation), UHF_PRR (electromagnetic radiation pulse repetition rate), UA_std (standard deviation of voltage UA signal), UC_max (maximum discharge amplitude of voltage UC signal), and UHF_c (standard deviation of voltage UA signal). Discharge amplitude), UC_std (standard deviation of voltage UC signal), AE_pctL (energy percentage of L band of acoustic signal), UHF_pctB (energy percentage of B band of electromagnetic radiation signal), AE_bw (spectral bandwidth of acoustic signal), AE_max (maximum discharge amplitude of acoustic signal), I_max (current amplitude), UHF_kurt (kurtosis of electromagnetic radiation signal), AE_kurt (kurtosis of acoustic signal), UHF_skew (skewness of electromagnetic radiation signal), UHF_pctA (energy percentage of A band of electromagnetic radiation signal), AE_pctR (energy percentage of R band of acoustic signal), UH_max (maximum discharge amplitude of voltage UH signal), UHF_dpctR (energy change rate of R band of electromagnetic radiation signal).

[0131] Table 2:

[0132]

[0133] After the standardization, feature selection, and dimensionality reduction processes described above, a low-dimensional feature subset of multiple physical quantities is obtained for subsequent model input. The feature vectors of all windows are aggregated to form a dataset for model training. This dataset is divided into a training set and a validation set in an 8:2 ratio (800 training sets, 200 sets for each stage; 200 validation sets, 50 sets for each stage). Using the full feature vector (38 dimensions) and the low-dimensional feature subset (23 dimensions) as inputs, respectively, support vector machine models based on the full feature vector input and support vector machine models based on the low-dimensional feature subset of multiple physical quantities are constructed and trained.

[0134] During training, 5-fold cross-validation is performed on the training set to optimize the hyperparameters. A radial basis function kernel is chosen as the kernel type, with a penalty coefficient C of 100. The kernel parameters are... Using a value of 0.1, and based on the selected optimal hyperparameters, the Sequence Minimum Optimization (SMO) algorithm is used to iteratively update the model parameters on the complete training set. Model convergence is achieved by maximizing the classification margin and suppressing training error. To prevent overfitting, training stops when the validation set recognition accuracy no longer improves (the improvement is less than 0.0001) after 15 consecutive parameter updates. The output model performance parameters include discharge severity discrimination accuracy, training time, single-window inference time, parameter size, model size, and computational cost. The discharge severity is manually labeled and includes at least the non-breakdown state, single-gap breakdown state, double-gap breakdown state, and through-breakdown state.

[0135] An additional 160 test sets were used for the experiments (40 sets for each stage). Table 3 shows a performance comparison between the support vector machine model based on full feature vector input and the support vector machine model based on low-dimensional feature subset input, based on experimental verification.

[0136] Table 3:

[0137]

[0138] As can be seen, the support vector machine model based on low-dimensional feature subset input achieves a simultaneous reduction in training and computational costs while only slightly decreasing the recognition accuracy. Specifically, the recognition accuracy decreases by 1.88 percentage points, which is within an acceptable range in engineering practice. Training time decreases from 5.449s to 4.851s, and single-window inference time decreases from 0.284ms to 0.246ms. Simultaneously, the parameter size decreases from 7500 Params to 7100 Params, the model size decreases from 120KB to 114KB, and the computational cost decreases from 270Mflops to 259Mflops. This demonstrates that, after feature selection and dimensionality reduction, redundant features can be effectively compressed and the discriminative model can be lightweighted while maintaining the basic severity recognition capability.

[0139] The signals of a single physical quantity, consisting only of electrical signals (voltage and current), acoustic emission signals, and ultra-high frequency electromagnetic radiation signals, are processed using the same procedure as for multi-physical quantities. This process involves dimension-wise standardization, feature selection, and dimensionality reduction to obtain low-dimensional feature subsets of each single physical quantity. Based on each subset, a support vector machine discrimination model with the same structure is trained independently (using the same kernel function, hyperparameter optimization strategy, and training procedure as the main model) to obtain the discrimination results for each subset.

[0140] Figure 3 This shows the confusion matrix for the accuracy of discharge severity identification under the input of low-dimensional feature subsets of multiple physical quantities. Figure 4 The diagram shows the confusion matrix indicating the accuracy of discharge severity identification under electrical signal (voltage and current) input conditions. Figure 5The confusion matrix shows the accuracy of discharge severity identification under acoustic emission signal input conditions. Figure 6 The diagram shows the confusion matrix of discharge severity identification accuracy under UHF electromagnetic radiation signal input. It is evident that the accuracy of discharge severity identification using multiple physical quantities is as high as 95.00%, while the accuracy rates of identification relying solely on electrical signals, acoustic signals, and UHF electromagnetic radiation signals are 81.25%, 78.75%, and 83.13%, respectively, all below 90%. This demonstrates that the combined input of multiple physical quantities fully utilizes the complementarity of voltage / current, acoustic signals, and electromagnetic radiation information in representing discharge intensity and characterizing stage evolution. Under the same training process and the same low-dimensional feature selection and dimensionality reduction strategy, it significantly improves the severity classification and discrimination ability. Compared to single physical quantity input, the identification accuracy of combined multiple physical quantity input is improved by approximately 11.87-16.25 percentage points, and the confusion between stages is significantly reduced, especially showing better differentiation between adjacent stages. This verifies the effectiveness and superiority of the present invention using multi-physical quantity detection and lightweight feature subsets for severity discrimination.

[0141] The present invention has been described in detail above with reference to preferred embodiments and exemplary examples. However, it should be noted that these specific embodiments are merely illustrative explanations of the invention and do not constitute any limitation on the scope of protection of the invention. Various improvements, equivalent substitutions, or modifications can be made to the technical content and embodiments of the present invention without departing from the spirit and scope of protection of the invention, and all such modifications fall within the scope of protection of the present invention. The scope of protection of the present invention is defined by the appended claims.

Claims

1. A lightweight system for determining the severity of inter-turn arc based on multiple physical quantities, characterized in that, The system includes: The multi-source signal acquisition and windowing module is used to acquire multi-source physical signals of multi-turn windings during arc discharge in oil and divide them into multiple sliding time windows. The full feature vector construction module is used to extract multi-dimensional feature parameters from each physical signal for each sliding time window to characterize the intensity and evolution state of arc discharge, and then concatenate them into a full feature vector. The feature standardization and dimensionality reduction module is used to perform dimensionality-wise standardization on the full feature vector, and to perform feature filtering and dimensionality reduction on the standardized feature vector to obtain a low-dimensional feature subset of multiple physical quantities. The model training module is used to construct and train a support vector machine discriminant model based on the low-dimensional feature subset of the multi-physical quantities. The lightweight discrimination module is used to input the low-dimensional feature subset of multiple physical quantities obtained after processing the multi-source physical signals to be judged into the trained support vector machine discrimination model to complete the lightweight discrimination of the severity of inter-turn arc.

2. The system according to claim 1, characterized in that, Preferably, in the multi-source signal acquisition and windowing module, the number of turns in the multi-turn winding is at least 4 turns and the last turn is grounded.

3. The system according to claim 1, characterized in that, In the multi-source signal acquisition and windowing module, the multi-source physical signals are transmitted to the computer via a data acquisition card with a sampling rate of 60-65 MHz.

4. The system according to claim 1, characterized in that, In the multi-source signal acquisition and windowing module, the multi-source physical signals include at least the voltage signal of the ungrounded winding, the arc circuit current signal, the acoustic emission signal, and the ultra-high frequency electromagnetic radiation signal.

5. The system according to claim 1, characterized in that, In the full feature vector construction module, the multi-dimensional feature parameters include maximum discharge amplitude, standard deviation of discharge amplitude, average discharge amplitude, skewness, kurtosis, pulse repetition rate, spectral center frequency, spectral bandwidth, frequency band energy ratio, and frequency band energy ratio change rate.

6. The system according to claim 1, characterized in that, In the feature standardization and dimensionality reduction module, the dimension-by-dimensional standardization process includes: performing zero-mean, unit-variance standardization on each feature parameter in the full feature vector, so that the mean of each feature in the sample set is 0 and the standard deviation is 1.

7. The system according to claim 1, characterized in that, In the model training module, the input of the support vector machine discriminant model is a dataset composed of feature vectors from the low-dimensional feature subsets of multiple physical quantities corresponding to each sliding time window.

8. A lightweight method for determining the severity of inter-turn arc based on multiple physical quantities, characterized in that, The method includes: Step 1: Obtain the multi-source physical signals of the multi-turn winding during arc discharge in oil, and divide them into multiple sliding time windows; Step 2: For each sliding time window, extract multi-dimensional feature parameters from each physical signal to characterize the arc discharge intensity and evolution state, and concatenate them into a full feature vector; Step 3: Perform dimension-wise standardization on the full feature vector, and perform feature filtering and dimensionality reduction on the standardized feature vector to obtain a low-dimensional feature subset of multiple physical quantities. Step 4: Based on the low-dimensional feature subset of the multi-physical quantities, construct and train a support vector machine discriminant model; Step 5: Input the low-dimensional feature subset of multiple physical quantities obtained after processing the multi-source physical signals to be judged in steps 1-3 into the trained support vector machine discrimination model to complete the lightweight discrimination of the severity of inter-turn arc.

9. The method according to claim 8, characterized in that, Step 1 includes: Step 1-1: Obtain multi-source physical signals of the multi-turn winding during arc discharge in oil; Steps 1-2 involve denoising the multi-source physical signals and dividing each physical signal into multiple sliding time windows based on a unified time reference.

10. A computer-readable storage medium comprising a stored complete computer program, characterized in that, When the computer program is run, it implements the method as described in claim 7 or 8.

Images