A method and system for interference source analysis for partial discharge measurement of large oil-filled equipment

Abstract

This invention relates to the field of interference source analysis technology, and provides a method and system for interference source analysis in partial discharge measurement of large oil-filled equipment, comprising: Step 1, acquiring the original acoustic signature signal of the large oil-filled equipment in operation; Step 2, preprocessing the original acoustic signature signal, including filtering, multi-stage downsampling, and acoustic signature data enhancement; Step 3, extracting the acoustic features of the preprocessed acoustic signature signal, including time-domain features, frequency-domain features, and time-frequency-domain features; and then performing multi-domain acoustic feature fusion; Step 4, based on the acoustic features, constructing a multi-domain feature-driven interference source classification model to identify and classify different types of interference sources; Step 5, based on the interference source identification results, suppressing interference in the original acoustic signature signal and extracting effective partial discharge signals. This invention can perform interference source analysis more effectively.

Description

Technical Field

[0001] This invention relates to the field of interference source analysis technology, and more specifically, to an interference source analysis method and system for measuring partial discharge in large oil-filled equipment. Background Technology

[0002] Large oil-filled equipment (such as converter transformers and power transformers) is the core equipment of the power system, and its operating status directly affects the safety and stability of the power grid. Partial discharge is an important sign of insulation degradation. If it is not detected and dealt with in time, it may lead to insulation breakdown and cause serious accidents. Therefore, accurate detection and early warning of partial discharge are of great engineering significance.

[0003] Currently, partial discharge detection methods mainly include electrical, chemical, and ultrasonic methods. Among them, ultrasonic detection is widely used in actual field applications due to its advantages such as non-contact operation, strong resistance to electromagnetic interference, and high positioning accuracy. Its basic principle is to utilize the ultrasonic signals generated by partial discharge, collect and analyze their characteristics through sensors, and thus determine the presence, type, and location of the discharge.

[0004] However, in practical applications, ultrasonic detection of partial discharge faces severe challenges from complex interference environments. Various interference sources exist in sites such as UHV converter stations, including but not limited to: I. Mechanical vibration noise: generated by the magnetostriction of transformer core, winding vibration, operation of cooling fan, oil pump, etc., with complex frequency components, often overlapping with partial discharge signal; II. Electromagnetic interference: Sources of high-voltage equipment, switch operations, radio signals, etc., may enter the acquisition system through coupling. III. Environmental noise: such as wind noise, rain noise, birdsong, etc., which have randomness and broadband characteristics; IV. System inherent noise: such as sensor background noise, thermal noise of the acquisition circuit, etc.

[0005] The presence of these interference sources severely reduces the signal-to-noise ratio of partial discharge signals, making it difficult for traditional thresholding and spectral analysis methods to accurately identify partial discharge signals, and easily leading to false alarms and missed alarms. Existing methods mostly rely on single-domain features (such as time-domain amplitude or frequency-domain peak value), lacking the ability to systematically model and classify multi-source interference, making it difficult to achieve highly robust partial discharge identification in complex environments.

[0006] In recent years, deep learning technology has demonstrated powerful capabilities in fields such as voiceprint recognition and fault diagnosis, providing new ideas for partial discharge interference suppression and signal recognition. However, existing research mainly focuses on the classification of partial discharge signals themselves, lacking systematic modeling and classification methods for interference sources, and research on multi-domain feature fusion and the construction of interference source identification models is still insufficient.

[0007] Therefore, there is an urgent need for a method that can effectively identify and classify various interference sources in partial discharge measurement. By modeling and suppressing interference sources, the accuracy and robustness of partial discharge signal extraction can be improved, providing a reliable foundation for subsequent partial discharge diagnosis and early warning. Summary of the Invention

[0008] The present invention provides a method and system for interference source analysis in partial discharge measurement of large oil-filled equipment, which can overcome some or all defects of the prior art.

[0009] An interference source analysis method for partial discharge measurement of large oil-filled equipment according to the present invention includes the following steps: Step 1: Obtain the original acoustic signature signal of the large oil filling equipment in operation; Step 2: Preprocess the original voiceprint signal, including filtering, multi-stage downsampling, and voiceprint data enhancement; Step 3: Extract the acoustic features of the preprocessed speaker signal, including time-domain features, frequency-domain features, and time-frequency-domain features; then perform multi-domain acoustic feature fusion. Step 4: Based on the acoustic features, construct a multi-domain feature-driven interference source classification model to identify and classify different types of interference sources; Step 5: Based on the interference source identification results, perform interference suppression on the original voiceprint signal and extract the effective partial discharge signal.

[0010] Preferably, the filtering method in step 2 is as follows: Let the original voiceprint signal be x ( t First, the signal is subjected to a Fast Fourier Transform (FFT) to obtain its spectral distribution; Based on the typical frequency characteristics of partial discharge acoustic emission signals, an adaptive bandpass filter is constructed. The filter frequency range is obtained by statistically analyzing the energy distribution of partial discharge samples in the acoustic fingerprint database, and its expression is as follows: in: The transfer function of the bandpass filter. The lower limit frequency, The upper limit frequency is used; based on bandpass filtering, wavelet noise reduction is introduced; the signal is decomposed into multiple scales using discrete wavelet transform: in: This is a low-frequency approximation component. For high-frequency detail components; Soft thresholding is applied to high-frequency detail components: in, To process high-frequency detail components, For symbolic functions, For the original high-frequency detail components, T This is an adaptive threshold.

[0011] As a preferred option, the multi-stage downsampling method in step 2 is as follows: First, the filtered signal is subjected to anti-aliasing filtering: in, The signal after anti-aliasing filtering. It is a low-pass filter; A phased downsampling strategy was then adopted; the original sampling rate was set to... The target sampling rate is Then downsampling factor for: The downsampling process is divided into two stages: Phase 1: Coarse Sampling; This is the signal after coarse downsampling. This is the coarse downsampling factor; Indicates from signal every middle One sample is drawn from each point, and the location of the draw is determined by... Sure, The sample number; Phase Two: Fine-grained Sampling; This is the signal after fine-sampling. This refers to the sampling factor for fine reduction. Indicates signal every middle One sample is drawn from each point, and the location of the draw is determined by... Sure; in: An energy preservation mechanism is also introduced during the downsampling process, by comparing the signal energy before and after downsampling. : The amplitude at the i-th sampling point; This represents the total quantity; If the energy loss exceeds the preset threshold, the downsampling ratio will be automatically adjusted to ensure that the partial discharge acoustic signature is not weakened.

[0012] As a preferred option, the voiceprint data enhancement method in step 2 is as follows: Data augmentation strategies are constructed by combining the physical characteristics of partial discharge acoustic signals, including the following methods: 2.1) Noise superposition and enhancement, used to simulate ambient noise at the equipment site; Add low-amplitude random noise to the original signal: in: For the enhanced signal, It is Gaussian white noise. Noise intensity coefficient; 2.2) Time-shift enhancement, used to simulate acoustic emission signals at different trigger times; New samples are generated by time-shifting the signal: in: This is a random time offset; 2.3) Frequency perturbation enhancement, used to simulate frequency changes caused by different equipment structures and propagation paths; Frequency perturbation is achieved by slightly scaling the signal spectrum: in: The frequency domain signal after perturbation The original frequency domain signal, This is the frequency disturbance coefficient; 2.4) Energy scale enhancement, used to simulate acoustic emission signals under different discharge intensities; Energy changes are achieved by altering the signal amplitude: Where γ is the amplitude scaling factor.

[0013] Preferably, in step 2, after filtering, multi-stage downsampling, and voiceprint data enhancement, a high-quality voiceprint signal dataset is obtained, and the processed signal... Represented as: This represents the nth signal; this dataset is used for subsequent analysis of partial discharge interference sources in large oil-filled equipment and training of the voiceprint recognition model.

[0014] As a preferred option, the temporal feature extraction method in step 3 is as follows: Let the preprocessed voiceprint signal be First, the signal is divided into frames, with each frame being [length missing]. N Then, the following time-domain features are calculated: 3.11) Root mean square value ; This indicator can reflect the changes in acoustic emission energy generated by partial discharge; 3.12) Peak Factor ; Partial discharge acoustic signals exhibit distinct pulse characteristics, and their peak factor is sensitive to discharge events. 3.13) Kurtosis ; The higher the kurtosis value, the more obvious the transient impulse component in the signal; 3.14) Short-time energy change rate ; This indicator can describe the abrupt changes in the energy of acoustic emission signals and can distinguish between mechanical vibration noise and discharge pulses; 3.15) Multi-scale pulse density characteristics ; This feature can be used to describe the frequency of partial discharge events; Finally, the time-domain feature vector is obtained. : The frequency domain feature extraction method is as follows: First, perform a Fast Fourier Transform on the signal; 3.21) Spectral centroid ; The spectral centroid is used to represent the concentrated location of spectral energy, and this feature can reflect the location of the main frequency of the acoustic signal; 3.22) Spectral bandwidth ; Spectral bandwidth is used to describe the degree of dispersion of spectral energy, and changes in bandwidth can reflect the acoustic characteristics of different interference sources; 3.23) Spectral Entropy ; Spectral entropy is used to measure the complexity of the spectrum distribution. The higher the spectral entropy, the more dispersed the signal spectrum. Based on the typical frequency range of partial discharge acoustic emission signals, the spectrum is divided into multiple frequency bands: low frequency band B1, mid frequency band B2, and high frequency band B3. Calculate the energy of each frequency band : Indicates the frequency band range; Constructing frequency band energy ratio : This feature can reflect the differences in the frequency structure of different interference sources; Finally, the frequency domain feature vector is obtained. : The time-frequency domain feature extraction method is as follows: The acquired partial discharge acoustic emission signal is subjected to continuous wavelet transform to obtain the time-frequency distribution of the signal at different time and frequency scales; the formula for the continuous wavelet transform is: in: These are continuous wavelet transform coefficients. a For scale parameters, b For time displacement, ψ For the mother wavelet function; Based on the time-frequency distribution, wavelet energy features are extracted to reflect the acoustic signal intensity at different frequency scales; the wavelet energy features The calculation formula is: These are the nth wavelet transform coefficients; Based on the time-frequency distribution, time-frequency energy concentration features are extracted to describe the time-frequency concentration of the partial discharge signal, thus distinguishing partial discharge pulses from continuous mechanical noise; the time-frequency energy concentration index The calculation formula is: in: Maximum energy frequency band; Based on the time-frequency distribution, a wavelet time-frequency map is constructed and converted into a two-dimensional matrix. Time-frequency texture features are extracted from the two-dimensional matrix. The time-frequency texture features include at least the mean energy, the variance energy, and the local gradient, which are used to characterize the local changes in the time-frequency structure. The extracted wavelet energy features, time-frequency energy concentration features, and time-frequency texture features are combined to obtain a time-frequency domain feature vector that characterizes the acoustic emission signal properties of partial discharge. : The average energy value. Let V be the energy variance.

[0015] As a preferred option, the multi-domain acoustic feature fusion method in step 3 is as follows: By fusing time-domain, frequency-domain, and time-frequency-domain features, a comprehensive acoustic feature vector is formed. : Multi-domain feature fusion can more comprehensively describe the characteristic structure of partial discharge acoustic emission signals.

[0016] Preferably, in step 4, the method for constructing the multi-domain feature-driven interference source classification model is as follows: 4.1) Construction of the feature dataset; First, the extracted time-domain features, frequency-domain features, and time-frequency-domain features are uniformly organized to construct an acoustic feature dataset; Let the feature vector of a single sample be... for: in: For time-domain feature vectors, For frequency domain eigenvectors, These are time-frequency domain feature vectors; All samples were used to construct the training dataset. : in: This indicates the category label of the interference source corresponding to the sample; The interference sources include: partial discharge sound sources, mechanical vibration noise, electromagnetic interference noise, and environmental background noise; by constructing the above dataset, we can provide basic data for subsequent model training. 4.2) Multi-domain feature fusion mapping; First, all features are normalized; then, a feature fusion mapping function is constructed. : in: These are the feature weight coefficients; The weighting coefficients are automatically determined using a feature mutual information evaluation method, thereby enhancing the ability to express key features. 4.3) Construction of interference source classification model; After feature fusion is completed, an interference source classification model is constructed; a two-stage classification structure model is adopted, including a coarse classification model and a fine classification model. 4.31) Coarse classification model; The coarse classification model is used to distinguish between partial discharge signals and non-discharge interference signals; This stage uses Support Vector Machines (SVM) for classification, and its discriminant function... for: in: For kernel function, For SVM Lagrange multipliers, This is the SVM bias term; using a coarse classification model, suspected partial discharge signals are quickly screened out. 4.32) Fine classification model; Based on the coarse classification, the identified partial discharge signals are further classified; The fine classification model uses a convolutional neural network (CNN) for feature learning, and its convolution operation is as follows: in: For the first i Feature maps output by the layer For convolution kernel, For the input feature matrix, For the i-th layer bias term; Finally, the softmax classifier outputs the category of the interference source: in: For the first i The probability of a type of interference source, Input values ​​for Softmax. This represents the number of interference source categories; 4.4) Model training methods; The classification model was trained using the constructed acoustic feature dataset. The dataset was first divided into a training set, a validation set, and a test set, with a ratio of 70% : 15% : 15%. The cross-entropy loss function is used during training. : To predict probabilities; The Adam optimization algorithm is used for parameter updates; through multiple rounds of iterative training, the model parameters gradually converge. 4.5) Model optimization methods, including: 4.51) Feature importance feedback mechanism; During model training, feature contribution is calculated. : For the j-th feature, the feature weight is adjusted according to its contribution, thereby strengthening the key feature; 4.52) Multi-scale sample training; By training with acoustic fingerprint samples of different time window lengths, the model can adapt to acoustic emission signals of different durations. 4.53) Model integration strategy; Combine multiple trained classification models: in: The final predicted probability of the model integration. For model weights, This refers to the number of models integrated; model ensemble can further improve the stability of classification results. 4.6) Output the classification results; Once trained, the interference source classification model can automatically identify the input voiceprint signal and output the corresponding interference source category; the output information includes the interference source category, classification probability, and feature contribution.

[0017] This invention provides an interference source analysis system for partial discharge measurement of large oil-filled equipment, which adopts the above-mentioned interference source analysis method for partial discharge measurement of large oil-filled equipment.

[0018] The beneficial effects of this invention are as follows: This invention fully explores the physical characteristics of partial discharge acoustic emission signals in different domains, extracting complementary features from the time domain (such as peak factor, kurtosis, and short-time energy change rate), frequency domain (such as spectral centroid, spectral entropy, and frequency band energy ratio), and time-frequency domain (such as wavelet energy and time-frequency texture) to construct a multi-dimensional fused feature vector. Compared to traditional identification methods that rely on single-domain features, this invention can more comprehensively and precisely characterize the acoustic fingerprint differences of different interference sources, significantly improving the accuracy and robustness of interference source classification.

[0019] This invention constructs a two-stage interference source identification model of "coarse classification + fine classification". First, a support vector machine (SVM) is used to quickly distinguish between partial discharge signals and non-discharge interference to screen suspected samples; then, a convolutional neural network (CNN) is used to perform fine classification of the partial discharge signals. This structure ensures the real-time requirements of online monitoring scenarios while improving the ability to identify complex interference sources, avoiding the trade-offs between accuracy and speed inherent in single models.

[0020] Based on actual field conditions, this invention designs a variety of data augmentation strategies with clear physical meanings, including noise superposition, time shifting, frequency perturbation, and energy scale enhancement. These methods can simulate signal changes under different environmental noise, propagation paths, and discharge intensities, effectively expanding the diversity of training samples. This allows the model to maintain good generalization performance when facing unseen conditions or novel interferences, reducing the risk of false alarms and false negatives.

[0021] By evaluating feature mutual information and analyzing gradient contribution, this invention dynamically adjusts the fusion weights of multi-domain features during model training, strengthening key features that contribute significantly to the classification task while weakening redundant or noisy features. This mechanism not only improves the model's classification performance but also provides a quantitative basis for subsequent feature selection and model interpretation.

[0022] This invention can not only identify and classify interference source types, but also perform targeted interference suppression on the original acoustic signature signal based on the identification results, thereby extracting high-quality partial discharge signals. This provides a reliable data foundation for subsequent partial discharge type identification, severity assessment, fault location, and early warning, significantly improving the intelligence level of condition monitoring and intelligent operation and maintenance of large oil-filled equipment. Attached Figure Description

[0023] Figure 1 This is a flowchart of an interference source analysis method for partial discharge measurement of a large oil-filled device, as described in Example 1. Detailed Implementation

[0024] To further understand the content of this invention, a detailed description of the invention will be provided in conjunction with the accompanying drawings and embodiments. It should be understood that the embodiments are merely illustrative and not limiting of the invention.

[0025] Example 1 like Figure 1 As shown, this embodiment provides an interference source analysis method for partial discharge measurement of large oil-filled equipment, which includes the following steps: Step 1: Obtain the original acoustic signature signal of the large oil filling equipment in operation.

[0026] Step 2: Preprocess the original voiceprint signal, including filtering, multi-stage downsampling, and voiceprint data enhancement.

[0027] In step 2, the filtering method is as follows: Let the original voiceprint signal be x ( t First, perform a Fast Fourier Transform (FFT) on the signal to obtain its spectral distribution: The frequency domain signal after Fourier transform. The number of sampling points. For frequency, The imaginary unit, The amplitude of the nth sampling point within a single frame; Based on the typical frequency characteristics of partial discharge acoustic emission signals, an adaptive bandpass filter is constructed. The filter frequency range is obtained by statistically analyzing the energy distribution of partial discharge samples in the acoustic fingerprint database, and its expression is as follows: in: The transfer function of the bandpass filter. The lower limit frequency, The upper limit frequency is used; based on bandpass filtering, wavelet noise reduction is introduced; the signal is decomposed into multiple scales using discrete wavelet transform: in: This is a low-frequency approximation component. For high-frequency detail components; Soft thresholding is applied to high-frequency detail components: in, To process high-frequency detail components, For symbolic functions, For the original high-frequency detail components, T This is an adaptive threshold.

[0028] In step 2, the multi-stage downsampling method is as follows: First, the filtered signal is subjected to anti-aliasing filtering: in, The signal after anti-aliasing filtering. It is a low-pass filter; A phased downsampling strategy was then adopted; the original sampling rate was set to... The target sampling rate is Then downsampling factor for: The downsampling process is divided into two stages: Phase 1: Coarse Sampling; This is the signal after coarse downsampling. This is the coarse downsampling factor; Indicates from signal every middle One sample is drawn from each point, and the location of the draw is determined by... Sure, The sample number; Phase Two: Fine-grained Sampling; This is the signal after fine-sampling. This refers to the sampling factor for fine reduction. Indicates signal every middle One sample is drawn from each point, and the location of the draw is determined by... Sure; in: An energy preservation mechanism is also introduced during the downsampling process, by comparing the signal energy before and after downsampling. : The amplitude at the i-th sampling point; This represents the total quantity; If the energy loss exceeds the preset threshold, the downsampling ratio will be automatically adjusted to ensure that the partial discharge acoustic signature is not weakened.

[0029] In step 2, the voiceprint data enhancement method is as follows: Data augmentation strategies are constructed by combining the physical characteristics of partial discharge acoustic signals, including the following methods: 2.1) Noise superposition and enhancement, used to simulate ambient noise at the equipment site; Add low-amplitude random noise to the original signal: in: For the enhanced signal, It is Gaussian white noise. Noise intensity coefficient; 2.2) Time-shift enhancement, used to simulate acoustic emission signals at different trigger times; New samples are generated by time-shifting the signal: in: This is a random time offset; 2.3) Frequency perturbation enhancement, used to simulate frequency changes caused by different equipment structures and propagation paths; Frequency perturbation is achieved by slightly scaling the signal spectrum: in: The frequency domain signal after perturbation The original frequency domain signal, This is the frequency disturbance coefficient; 2.4) Energy scale enhancement, used to simulate acoustic emission signals under different discharge intensities; Energy changes are achieved by altering the signal amplitude: Where γ is the amplitude scaling factor.

[0030] In step 2, after filtering, multi-stage downsampling, and voiceprint data enhancement, a high-quality voiceprint signal dataset is obtained. The processed signal... Represented as: This represents the nth signal; this dataset is used for subsequent analysis of partial discharge interference sources in large oil-filled equipment and training of the voiceprint recognition model.

[0031] Step 3: Extract the acoustic features of the preprocessed voiceprint signal, including time-domain features, frequency-domain features, and time-frequency-domain features; then perform multi-domain acoustic feature fusion.

[0032] In step 3, the temporal feature extraction method is as follows: Let the preprocessed voiceprint signal be First, the signal is divided into frames, with each frame being [length missing]. N Then, the following time-domain features are calculated: 3.11) Root mean square value ; The root mean square (RMS) value is used to reflect the energy level of the acoustic signal. This represents the number of sampling points for a single frame of signal. For the first i The signal amplitude at each sampling point; This indicator can reflect the changes in acoustic emission energy generated by partial discharge; 3.12) Peak Factor ; The peak factor is used to characterize the signal impulse characteristics: The peak value of the signal; Partial discharge acoustic signals exhibit distinct pulse characteristics, and their peak factor is sensitive to discharge events. 3.13) Kurtosis ; Kurtosis is used to reflect the sharpness of a signal distribution, and its calculation formula is as follows: in: For mathematical expectation, It is a sequence of signal amplitudes within a single frame. m The mean of the signal. s The standard deviation of the signal is represented by the kurtosis value; the higher the kurtosis value, the more pronounced the transient impulse component in the signal. 3.14) Short-time energy change rate ; To improve the ability to identify partial discharge pulses, a short-time energy change rate is introduced: in: For the first n Frame signal energy; this indicator can describe the abrupt changes in acoustic emission signal energy and can distinguish between mechanical vibration noise and discharge pulses; 3.15) Multi-scale pulse density characteristics ; Divide the signal into multiple time windows and count the number of pulse peaks per unit time: in: The number of pulses detected. T This is a statistical time window; this feature can be used to describe the frequency of partial discharge events. Finally, the time-domain feature vector is obtained. : The frequency domain feature extraction method is as follows: First, perform a Fast Fourier Transform on the signal: Obtain the spectral information of the signal; For frequency, The imaginary unit, The amplitude of the nth sampling point within a single frame; 3.21) Spectral centroid ; The spectral centroid is used to represent the concentrated location of spectral energy. Let i be the frequency value of the i-th frequency point; This feature can reflect the main frequency position of the acoustic signal; 3.22) Spectral bandwidth ; Spectral bandwidth is used to describe the degree of dispersion of spectral energy: Variations in bandwidth can reflect the acoustic characteristics of different interference sources; 3.23) Spectral Entropy ; Spectral entropy is used to measure the complexity of a spectral distribution. in: Normalized spectral probability; Indicates Logarithm to base 0; The higher the spectral entropy, the more dispersed the signal spectrum; Based on the typical frequency range of partial discharge acoustic emission signals, the spectrum is divided into multiple frequency bands: low frequency band B1, mid frequency band B2, and high frequency band B3. Calculate the energy of each frequency band : Indicates the frequency band range; Constructing frequency band energy ratio : This feature can reflect the differences in the frequency structure of different interference sources; Finally, the frequency domain feature vector is obtained. : The time-frequency domain feature extraction method is as follows: The acquired partial discharge acoustic emission signal is subjected to continuous wavelet transform to obtain the time-frequency distribution of the signal at different time and frequency scales; the formula for the continuous wavelet transform is: in: These are continuous wavelet transform coefficients. a For scale parameters, b For time displacement, ψ For the mother wavelet function; Based on the time-frequency distribution, wavelet energy features are extracted to reflect the acoustic signal intensity at different frequency scales; the wavelet energy features The calculation formula is: These are the nth wavelet transform coefficients; Based on the time-frequency distribution, time-frequency energy concentration features are extracted to describe the time-frequency concentration of the partial discharge signal, thus distinguishing partial discharge pulses from continuous mechanical noise; the time-frequency energy concentration index The calculation formula is: in: Maximum energy frequency band; Based on the time-frequency distribution, a wavelet time-frequency map is constructed and converted into a two-dimensional matrix. Time-frequency texture features are extracted from the two-dimensional matrix. The time-frequency texture features include at least the mean energy, the variance energy, and the local gradient, which are used to characterize the local changes in the time-frequency structure. The extracted wavelet energy features, time-frequency energy concentration features, and time-frequency texture features are combined to obtain a time-frequency domain feature vector that characterizes the acoustic emission signal properties of partial discharge. : The average energy value. Let V be the energy variance.

[0033] In step 3, the multi-domain acoustic feature fusion method is as follows: By fusing time-domain, frequency-domain, and time-frequency-domain features, a comprehensive acoustic feature vector is formed. : Multi-domain feature fusion can more comprehensively describe the characteristic structure of partial discharge acoustic emission signals.

[0034] Step 4: Based on the acoustic features, construct a multi-domain feature-driven interference source classification model to identify and classify different types of interference sources.

[0035] In step 4, the method for constructing the multi-domain feature-driven interference source classification model is as follows: 4.1) Construction of the feature dataset; First, the extracted time-domain features, frequency-domain features, and time-frequency-domain features are uniformly organized to construct an acoustic feature dataset; Let the feature vector of a single sample be... for: in: For time-domain feature vectors, For frequency domain eigenvectors, These are time-frequency domain feature vectors; All samples were used to construct the training dataset. : in: This indicates the category label of the interference source corresponding to the sample; The interference sources include: partial discharge sound sources, mechanical vibration noise, electromagnetic interference noise, and environmental background noise; by constructing the above dataset, we can provide basic data for subsequent model training. 4.2) Multi-domain feature fusion mapping; First, normalize all features: in: The normalized feature vectors, m The characteristic mean, s Standard deviation; Subsequently, a feature fusion mapping function is constructed. : in: These are the feature weight coefficients; The weighting coefficients are automatically determined using a feature mutual information evaluation method, thereby enhancing the ability to express key features. 4.3) Construction of interference source classification model; After feature fusion is completed, an interference source classification model is constructed; a two-stage classification structure model is adopted, including a coarse classification model and a fine classification model. 4.31) Coarse classification model; The coarse classification model is used to distinguish between partial discharge signals and non-discharge interference signals; This stage uses Support Vector Machines (SVM) for classification, and its discriminant function... for: in: For kernel function, For SVM Lagrange multipliers, This is the SVM bias term; using a coarse classification model, suspected partial discharge signals are quickly screened out. 4.32) Fine classification model; Based on the coarse classification, the identified partial discharge signals are further classified; The fine classification model uses a convolutional neural network (CNN) for feature learning, and its convolution operation is as follows: in: For the first i Feature maps output by the layer For convolution kernel, For the input feature matrix, For the i-th layer bias term; Finally, the softmax classifier outputs the category of the interference source: in: For the first i The probability of a type of interference source, Input values ​​for Softmax. This represents the number of interference source categories; 4.4) Model training methods; The classification model was trained using the constructed acoustic feature dataset. The dataset was first divided into a training set, a validation set, and a test set, with a ratio of 70% : 15% : 15%. The cross-entropy loss function is used during training. : To predict probabilities; The Adam optimization algorithm is used for parameter updates; through multiple rounds of iterative training, the model parameters gradually converge. 4.5) Model optimization methods, including: 4.51) Feature importance feedback mechanism; During model training, feature contribution is calculated. : For the j-th feature, the feature weight is adjusted according to its contribution, thereby strengthening the key feature; 4.52) Multi-scale sample training; By training with acoustic fingerprint samples of different time window lengths, the model can adapt to acoustic emission signals of different durations. 4.53) Model integration strategy; Combine multiple trained classification models: in: The final predicted probability of the model integration. For model weights, This refers to the number of models integrated; model ensemble can further improve the stability of classification results. 4.6) Output the classification results; Once trained, the interference source classification model can automatically identify the input voiceprint signal and output the corresponding interference source category; the output information includes the interference source category, classification probability, and feature contribution.

[0036] Step 5: Based on the interference source identification results, perform interference suppression on the original voiceprint signal and extract the effective partial discharge signal.

[0037] This embodiment provides an interference source analysis system for partial discharge measurement of large oil-filled equipment, which adopts the interference source analysis method for partial discharge measurement of large oil-filled equipment described above.

[0038] This embodiment can perform interference source analysis better.

[0039] Example 2 This embodiment uses a main transformer of model SFSZ10-180000 / 220 in a 220kV substation as the test object. The specific implementation steps are as follows: Step 1: Acquisition of raw voiceprint signal; Under normal transformer operation, three wideband ultrasonic sensors (frequency response range: 20kHz-200kHz) are installed on the outer wall of the transformer tank. The sensors are magnetically fixed and positioned at the center of the front, center of the side, and top of the transformer, respectively. The sampling system uses a 24-bit high-precision data acquisition card, with a sampling rate set to 500kHz and a continuous acquisition time of 60 minutes to acquire the raw acoustic signature signal. x ( t ).

[0040] Step 2: Signal preprocessing; 2.1 Filtering process; Perform a Fast Fourier Transform on the original voiceprint signal to obtain the spectral distribution. Based on the typical frequency characteristics of partial discharge (30-160kHz), an adaptive bandpass filter is constructed, and a lower limit frequency is set. =30kHz, upper frequency limit =160kHz, out-of-band interference is filtered out. Then, discrete wavelet transform is performed, and a 3-level decomposition is conducted using the db4 wavelet basis to extract high-frequency detail components. Perform soft thresholding.

[0041] 2.2 Multi-stage downsampling; Original sampling rate =500kHz, target sampling rate =50kHz, downsampling factor M=10. Downsampling is divided into two stages: the first stage is coarse downsampling. =5, signal obtained Second stage fine-tuning sampling Receive signal Anti-aliasing filtering is performed before downsampling using a low-pass filter with a cutoff frequency of 25kHz. After downsampling, the signal energy E is calculated and compared with the original signal energy. The energy loss is 2.3%, which is lower than the preset threshold of 5%, and meets the requirements.

[0042] 2.3 Voiceprint data enhancement; Four enhancement strategies were employed to expand the sample set: Noise amplification: Added to 10% of the samples =0.05 Gaussian white noise to simulate ambient background noise; Time shift enhancement: Random time shift of Δt = 50-200μs is applied to 10% of the samples; Frequency perturbation enhancement: Frequency scaling of 10% of the samples with β=0.95-1.05; Energy scale enhancement: Amplitude scaling of γ = 0.8-1.2 is applied to 10% of the samples.

[0043] The enhanced voiceprint signal dataset is obtained after processing. It contains a total of 12,000 samples.

[0044] Step 3: Multi-domain acoustic feature extraction; 3.1 Temporal feature extraction; The preprocessed signal is then divided into frames, with a frame length of N = 1024 points and a frame shift of 512 points. Calculations are performed for each frame: Root mean square (RMS): Reflects the signal energy level Peak factor CF: Characterizes the impact characteristics. The calculated CF is ≈3.2 in the normal state and can reach 8.5 in the partial discharge state. Kurtosis K: K≈3.1 in normal state, K>10 in partial discharge state; Short-term energy change rate (ER): ER < 0.2 in normal state, ER can reach 0.8-1.5 in partial discharge state; Multiscale pulse density (PD): Statistical window T=1s, normal state PD≈0-5 times / s, partial discharge state PD>20 times / s; Obtain the time-domain feature vector .

[0045] 3.2 Frequency domain feature extraction; Perform an FFT transform on each frame of the signal and calculate: Spectral centroid SC: Reflects the center of gravity of the frequency. Under normal conditions, SC≈2.1kHz, and under partial discharge conditions, SC≈45-80kHz. Spectral bandwidth (BW): Reflects the degree of frequency dispersion. Under normal conditions, BW ≈ 1.5kHz, and under partial discharge conditions, BW ≈ 25-40kHz. Spectral entropy H: H≈0.3 in normal state, H≈0.6-0.8 in partial discharge state; The spectrum was divided into three frequency bands: B1 (0-20kHz), B2 (20-80kHz), and B3 (80-160kHz). The energy ratios R1, R2, and R3 of each frequency band were calculated. Under normal conditions, R1 > 0.9, and R2 and R3 < 0.1; under partial discharge conditions, R2 or R3 increased significantly.

[0046] Obtain the frequency domain eigenvector .

[0047] 3.3 Time-frequency domain feature extraction; Continuous wavelet transform is performed on the signal using the cmor3-3 mother wavelet with a scale range of a = 1-128. Extraction: Wavelet energy characteristics : Calculate the total energy at each scale; Time-frequency energy concentration TC: Normal state TC≈0.15, partial discharge state TC≈0.35-0.5; Time-frequency texture features: Extracting the mean energy from the time-frequency map and energy variance ; Obtain the time-frequency domain eigenvector .

[0048] 3.4 Multi-domain feature fusion; The features from the three domains are concatenated to form a comprehensive feature vector. The dimensions are 5+6+4=15.

[0049] Step 4: Construction and training of the interference source classification model; 4.1 Construction of the feature dataset; The 12,000 samples were divided into training, validation, and test sets in a 7:2:1 ratio. Interference sources included partial discharge noise, mechanical vibration noise, electromagnetic interference noise, and ambient background noise, with 3,000 samples in each category. Features were then normalized.

[0050] 4.2 Model Training; Coarse classification model: SVM classifier is used, RBF kernel is selected as kernel function, penalty parameter C=10, gamma=0.1, and training distinguishes between partial discharge signal and non-discharge interference signal.

[0051] The fine classification model employs a CNN classifier with the following network structure: Input layer (15-dimensional) → Fully connected layer (64 nodes, ReLU activation) → Dropout layer (0.3) → Fully connected layer (32 nodes, ReLU activation) → Output layer (4 nodes, Softmax activation). It uses the cross-entropy loss function, Adam optimizer, learning rate 0.001, batch size 64, and 50 iterations.

[0052] 4.3 Model Optimization; Calculate feature contribution during training. The feature fusion weights α, β, and γ were dynamically adjusted. After optimization, the final weight coefficients were determined to be α=0.3 (time domain), β=0.4 (frequency domain), and γ=0.3 (time-frequency domain). Simultaneously, a model ensemble strategy was employed to integrate five models with different initializations.

[0053] Step 5: Interference source identification and signal extraction; The test set samples were input into the trained model, and the recognition accuracy reached 97.2%. The recognition results for various interference sources are as follows: Partial discharge sound source: recall rate 96.8%, precision rate 97.5%; Mechanical vibration and noise: recall rate 98.2%, precision rate 97.9%; Electromagnetic interference noise: recall rate 95.7%, precision rate 96.3%; Background noise: Recall 97.9%, Precision 96.8%; Based on the identification results, targeted interference suppression is applied to the original signal to extract the effective partial discharge signal. For example, when electromagnetic interference is identified, a notch filter is used to filter out the corresponding frequency band; when mechanical vibration noise is identified, adaptive filtering is used to eliminate periodic components.

[0054] Verification Experiment To verify the effectiveness of the method in this embodiment, an artificial discharge defect (point discharge, 4mm gap, 20kV voltage) was set on the same transformer. Signals were collected under three conditions: no interference, wind noise interference, and electromagnetic interference. The recognition performance of the method in this embodiment was compared with that of the conventional threshold method. The results are shown in Table 1: Table 1. Recognition performance of the method in this embodiment compared to the conventional threshold method. Experimental results show that the method in this embodiment can maintain a high recognition accuracy under various interference conditions, which is significantly better than the conventional threshold method. At the same time, it greatly improves the signal-to-noise ratio and provides a high-quality input signal for subsequent partial discharge diagnosis.

[0055] The present invention and its embodiments have been described above illustratively. This description is not restrictive, and the figures shown are only one embodiment of the present invention; the actual structure is not limited thereto. Therefore, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the present invention, such designs should fall within the protection scope of the present invention.

Claims

1. A method for analyzing interference sources in partial discharge measurements of large oil-filled equipment, characterized in that, Includes the following steps: Step 1: Obtain the original acoustic signature signal of the large oil filling equipment in operation; Step 2: Preprocess the original voiceprint signal, including filtering, multi-stage downsampling, and voiceprint data enhancement; Step 3: Extract the acoustic features of the preprocessed speaker signal, including time-domain features, frequency-domain features, and time-frequency-domain features; Then, multi-domain acoustic feature fusion is performed; Step 4: Based on the acoustic features, construct a multi-domain feature-driven interference source classification model to identify and classify different types of interference sources; Step 5: Based on the interference source identification results, perform interference suppression on the original voiceprint signal and extract the effective partial discharge signal.

2. The interference source analysis method for partial discharge measurement of large oil-filled equipment according to claim 1, characterized in that, In step 2, the filtering method is as follows: Let the original voiceprint signal be x ( t First, the signal is subjected to a Fast Fourier Transform (FFT) to obtain its spectral distribution; Based on the typical frequency characteristics of partial discharge acoustic emission signals, an adaptive bandpass filter is constructed. The filter frequency range is obtained by statistically analyzing the energy distribution of partial discharge samples in the acoustic fingerprint database, and its expression is as follows: in: The transfer function of the bandpass filter. The lower limit frequency, The upper limit frequency is used; based on bandpass filtering, wavelet noise reduction is introduced; the signal is decomposed into multiple scales using discrete wavelet transform: in: This is a low-frequency approximation component. For high-frequency detail components; Soft thresholding is applied to high-frequency detail components: in, To process high-frequency detail components, For symbolic functions, For the original high-frequency detail components, T This is an adaptive threshold.

3. The interference source analysis method for partial discharge measurement of large oil-filled equipment according to claim 2, characterized in that, In step 2, the multi-stage downsampling method is as follows: First, the filtered signal is subjected to anti-aliasing filtering: in, The signal after anti-aliasing filtering. It is a low-pass filter; A phased downsampling strategy was then adopted; the original sampling rate was set to... The target sampling rate is Then downsampling factor for: The downsampling process is divided into two stages: Phase 1: Coarse Sampling; This is the signal after coarse downsampling. This is the coarse downsampling factor; Indicates from signal every middle One sample is drawn from each point, and the location of the draw is determined by... Sure, The sample number; Phase Two: Fine-grained Sampling; This is the signal after fine-sampling. This refers to the sampling factor for fine reduction. Indicates signal every middle One sample is drawn from each point, and the location of the draw is determined by... Sure; in: An energy preservation mechanism is also introduced during the downsampling process, by comparing the signal energy before and after downsampling. : The amplitude at the i-th sampling point; This represents the total quantity; If the energy loss exceeds the preset threshold, the downsampling ratio will be automatically adjusted to ensure that the partial discharge acoustic signature is not weakened.

4. The interference source analysis method for partial discharge measurement of large oil-filled equipment according to claim 3, characterized in that, In step 2, the voiceprint data enhancement method is as follows: Data augmentation strategies are constructed by combining the physical characteristics of partial discharge acoustic signals, including the following methods: 2.1) Noise superposition and enhancement, used to simulate ambient noise at the equipment site; Add low-amplitude random noise to the original signal: in: For the enhanced signal, It is Gaussian white noise. Noise intensity coefficient; 2.2) Time-shift enhancement, used to simulate acoustic emission signals at different trigger times; New samples are generated by time-shifting the signal: in: This is a random time offset; 2.3) Frequency perturbation enhancement, used to simulate frequency changes caused by different equipment structures and propagation paths; Frequency perturbation is achieved by slightly scaling the signal spectrum: in: The frequency domain signal after perturbation The original frequency domain signal, This is the frequency disturbance coefficient; 2.4) Energy scale enhancement, used to simulate acoustic emission signals under different discharge intensities; Energy changes are achieved by altering the signal amplitude: Where γ is the amplitude scaling factor.

5. The interference source analysis method for partial discharge measurement of large oil-filled equipment according to claim 4, characterized in that, In step 2, after filtering, multi-stage downsampling, and voiceprint data enhancement, a high-quality voiceprint signal dataset is obtained. The processed signal... Represented as: This represents the nth signal; this dataset is used for subsequent analysis of partial discharge interference sources in large oil-filled equipment and training of the voiceprint recognition model.

6. The interference source analysis method for partial discharge measurement of large oil-filled equipment according to claim 5, characterized in that, In step 3, the temporal feature extraction method is as follows: Let the preprocessed voiceprint signal be First, the signal is divided into frames, with each frame being [length missing]. N Then, the following time-domain features are calculated: 3.11) Root mean square value ; This indicator can reflect the changes in acoustic emission energy generated by partial discharge; 3.12) Peak Factor ; Partial discharge acoustic signals exhibit distinct pulse characteristics, and their peak factor is sensitive to discharge events. 3.13) Kurtosis ; The higher the kurtosis value, the more obvious the transient impulse component in the signal; 3.14) Short-time energy change rate ; This indicator can describe the abrupt changes in the energy of acoustic emission signals and can distinguish between mechanical vibration noise and discharge pulses; 3.15) Multi-scale pulse density characteristics ; This feature can be used to describe the frequency of partial discharge events; Finally, the time-domain feature vector is obtained. : The frequency domain feature extraction method is as follows: First, perform a Fast Fourier Transform on the signal; 3.21) Spectral centroid ; The spectral centroid is used to represent the concentrated location of spectral energy, and this feature can reflect the location of the main frequency of the acoustic signal; 3.22) Spectral bandwidth ; Spectral bandwidth is used to describe the degree of dispersion of spectral energy, and changes in bandwidth can reflect the acoustic characteristics of different interference sources; 3.23) Spectral Entropy ; Spectral entropy is used to measure the complexity of the spectrum distribution. The higher the spectral entropy, the more dispersed the signal spectrum. Based on the typical frequency range of partial discharge acoustic emission signals, the spectrum is divided into multiple frequency bands: low frequency band B1, mid frequency band B2, and high frequency band B3. Calculate the energy of each frequency band : Indicates the frequency band range; Constructing frequency band energy ratio : This feature can reflect the differences in the frequency structure of different interference sources; Finally, the frequency domain feature vector is obtained. : The time-frequency domain feature extraction method is as follows: The acquired partial discharge acoustic emission signal is subjected to continuous wavelet transform to obtain the time-frequency distribution of the signal at different time and frequency scales; the formula for the continuous wavelet transform is: in: These are continuous wavelet transform coefficients. a For scale parameters, b For time displacement, ψ For the mother wavelet function; Based on the time-frequency distribution, wavelet energy features are extracted to reflect the acoustic signal intensity at different frequency scales; the wavelet energy features The calculation formula is: These are the nth wavelet transform coefficients; Based on the time-frequency distribution, time-frequency energy concentration features are extracted to describe the time-frequency concentration of the partial discharge signal, thus distinguishing partial discharge pulses from continuous mechanical noise; the time-frequency energy concentration index The calculation formula is: in: Maximum energy frequency band; Based on the time-frequency distribution, a wavelet time-frequency map is constructed and converted into a two-dimensional matrix. Time-frequency texture features are extracted from the two-dimensional matrix. The time-frequency texture features include at least the mean energy, the variance energy, and the local gradient, which are used to characterize the local changes in the time-frequency structure. The extracted wavelet energy features, time-frequency energy concentration features, and time-frequency texture features are combined to obtain a time-frequency domain feature vector that characterizes the acoustic emission signal properties of partial discharge. : The average energy value. Let V be the energy variance.

7. The interference source analysis method for partial discharge measurement of large oil-filled equipment according to claim 6, characterized in that, In step 3, the multi-domain acoustic feature fusion method is as follows: By fusing time-domain, frequency-domain, and time-frequency-domain features, a comprehensive acoustic feature vector is formed. : Multi-domain feature fusion can more comprehensively describe the characteristic structure of partial discharge acoustic emission signals.

8. The interference source analysis method for partial discharge measurement of large oil-filled equipment according to claim 7, characterized in that, In step 4, the method for constructing the multi-domain feature-driven interference source classification model is as follows: 4.1) Construction of the feature dataset; First, the extracted time-domain features, frequency-domain features, and time-frequency-domain features are uniformly organized to construct an acoustic feature dataset; Let the feature vector of a single sample be... for: in: For time-domain feature vectors, For frequency domain eigenvectors, These are time-frequency domain feature vectors; All samples were used to construct the training dataset. : in: This indicates the category label of the interference source corresponding to the sample; The interference sources include: partial discharge sound sources, mechanical vibration noise, electromagnetic interference noise, and environmental background noise; by constructing the above dataset, we can provide basic data for subsequent model training. 4.2) Multi-domain feature fusion mapping; First, all features are normalized; then, a feature fusion mapping function is constructed. : in: These are the feature weight coefficients; The weighting coefficients are automatically determined using a feature mutual information evaluation method, thereby enhancing the ability to express key features. 4.3) Construction of interference source classification model; After feature fusion is completed, an interference source classification model is constructed; a two-stage classification structure model is adopted, including a coarse classification model and a fine classification model. 4.31) Coarse classification model; The coarse classification model is used to distinguish between partial discharge signals and non-discharge interference signals; This stage uses Support Vector Machines (SVM) for classification, and its discriminant function... for: in: For kernel function, For SVM Lagrange multipliers, This is the SVM bias term; using a coarse classification model, suspected partial discharge signals are quickly screened out. 4.32) Fine classification model; Based on the coarse classification, the identified partial discharge signals are further classified; The fine classification model uses a convolutional neural network (CNN) for feature learning, and its convolution operation is as follows: in: For the first i Feature maps output by the layer For convolution kernel, For the input feature matrix, For the i-th layer bias term; Finally, the softmax classifier outputs the category of the interference source: in: For the first i The probability of a type of interference source, Input values ​​for Softmax. This represents the number of interference source categories; 4.4) Model training methods; The classification model was trained using the constructed acoustic feature dataset. The dataset was first divided into a training set, a validation set, and a test set, with a ratio of 70% : 15% : 15%. The cross-entropy loss function is used during training. : To predict probabilities; The Adam optimization algorithm is used for parameter updates; through multiple rounds of iterative training, the model parameters gradually converge. 4.5) Model optimization methods, including: 4.51) Feature importance feedback mechanism; During model training, feature contribution is calculated. : For the j-th feature, the feature weight is adjusted according to its contribution, thereby strengthening the key feature; 4.52) Multi-scale sample training; By training with acoustic fingerprint samples of different time window lengths, the model can adapt to acoustic emission signals of different durations. 4.53) Model integration strategy; Combine multiple trained classification models: in: The final predicted probability of the model integration. For model weights, This refers to the number of models integrated; model ensemble can further improve the stability of classification results. 4.6) Output the classification results; Once trained, the interference source classification model can automatically identify the input voiceprint signal and output the corresponding interference source category; the output information includes the interference source category, classification probability, and feature contribution.

9. An interference source analysis system for partial discharge measurement of large oil-filled equipment, characterized in that, It employs an interference source analysis method for partial discharge measurement of large oil-filled equipment as described in any one of claims 1-8.

Images