A portable corona detection method fusing TEV and ultrasonic multi-band signals

Abstract

The application discloses a portable partial discharge detection method fusing TEV and ultrasonic multi-band signals, comprising the following steps: S1, using a transient ground voltage sensor and an ultrasonic sensor array, synchronously collecting transient ground voltage signals and multi-band ultrasonic signals at a power equipment to be measured, thereby, the TEV sensor and the ultrasonic multi-band sensor array are used to cooperatively collect, the characteristic signal bands of different types and different severity partial discharges are covered, the comprehensiveness and accuracy of detection are greatly improved, the amplitude statistics and pulse phase analysis are performed on the TEV signals, the in-depth time-frequency analysis is performed on the ultrasonic signals of each band, the operation efficiency of a subsequent fusion diagnosis model is improved, thereby, the partial discharge type recognition accuracy is effectively improved, the weak partial discharge signals can be effectively captured, the early warning demand is met, the fusion diagnosis model with stronger generalization ability is used, combined with multi-feature correlation judgment logic, the model parameters can be calibrated through online fine adjustment, and the reliability of detection results is ensured.

Description

Technical Field

[0001] This application relates to the technical field of partial discharge detection, and in particular to a portable partial discharge detection method that integrates TEV and ultrasonic multi-band signals. Background Technology

[0002] During long-term operation, power equipment is prone to partial discharge due to factors such as insulation aging, manufacturing defects, and environmental corrosion. Partial discharge is an important sign of insulation deterioration in power equipment. If it is not detected and addressed in time, it will gradually exacerbate insulation damage, eventually leading to equipment failure or even power outages, seriously affecting the safe and stable operation of the power system.

[0003] Currently, transient ground voltage (TEV) detection is the most widely used single-signal detection method for partial discharge detection in power equipment. This method employs a non-contact acquisition mode, making it convenient and efficient to operate, and it has strong anti-interference capabilities against electromagnetic interference commonly found in industrial environments. It is suitable for on-site detection of various enclosed power equipment such as switchgear and GIS equipment. However, this method has obvious inherent limitations; it can only capture the ground voltage pulse signal generated during partial discharge and cannot obtain the specific detailed characteristics of the partial discharge.

[0004] On the one hand, its positioning accuracy is extremely low. It can usually only roughly determine the overall area of ​​the equipment where the partial discharge is located, and cannot accurately locate the specific location of the partial discharge (such as a terminal in the switch cabinet or a winding inside the transformer). This causes maintenance personnel to conduct a comprehensive inspection of the entire equipment area, which not only increases the workload and cost of maintenance, but may also cause the fault point to be missed due to incomplete inspection, thus delaying the maintenance opportunity.

[0005] On the other hand, this method cannot effectively distinguish the specific types of partial discharge, such as corona discharge, surface discharge, and internal discharge. These different types of partial discharge have significantly different degrees of damage to equipment insulation and different rates of development. Without distinguishing the type of partial discharge, it is difficult to formulate targeted maintenance strategies. Furthermore, TEV detection lacks sensitivity to weak partial discharge signals. In the early stages of slight insulation degradation and low partial discharge intensity, the generated TEV signal amplitude is weak and easily masked by weak interference signals in the environment, leading to ineffective detection and missed detections. This results in failure to provide early warning of partial discharge, missing the optimal maintenance opportunity, and ultimately causing further insulation degradation. In addition, the propagation of TEV signals is easily affected by the material and thickness of the equipment casing and the site layout, potentially causing signal attenuation and distortion, further reducing the accuracy of detection.

[0006] Application content

[0007] This application aims to address, at least to some extent, the technical problems in the related art.

[0008] To achieve the above objectives, this application proposes a portable partial discharge detection method that integrates TEV and multi-band ultrasonic signals, comprising the following steps:

[0009] S1. Using a transient ground voltage sensor and an ultrasonic sensor array, the transient ground voltage signal and multi-band ultrasonic signal at the power equipment under test are collected simultaneously.

[0010] S2. Perform amplitude statistics and pulse phase analysis on the acquired transient ground voltage signal to extract the first feature parameter; perform time-frequency analysis and feature extraction on the acquired ultrasonic signals of each frequency band to obtain the second feature parameter set corresponding to each frequency band;

[0011] S3. Construct a multi-source feature vector based on the first feature parameter and each of the second feature parameter sets;

[0012] S4. Input the multi-source feature vector into the pre-trained fusion diagnostic model. The fusion diagnostic model outputs the partial discharge type identification result, severity level and location information according to the built-in joint judgment logic. The joint judgment logic includes at least the association rules based on the activity of transient ground voltage signal and the center frequency and spectral features of ultrasonic signal.

[0013] S5. Compare and verify the results output by the fusion diagnostic model with the actual on-site detection. If the deviation between the identification result and the actual situation exceeds the preset threshold, calibrate by fine-tuning the model parameters online. If the deviation is within the allowable range, the detection result is confirmed to be valid.

[0014] In addition, the application may also include the following additional technical features:

[0015] Specifically, in step S1, the synchronous acquisition is achieved through GPS timing or a high-precision clock synchronization module, and the time synchronization error between the transient voltage signal and the multi-band ultrasonic signal does not exceed 1μs.

[0016] Specifically, in step S2, the amplitude statistics include statistical analysis of the peak value, effective value, kurtosis, and waveform factor of the transient ground voltage signal, and the pulse phase analysis uses the phase-resolved partial discharge spectrum analysis method to extract the pulse phase distribution characteristics as the first feature parameter.

[0017] Specifically, in step S2, the time-frequency analysis uses wavelet transform or short-time Fourier transform to decompose the ultrasonic signals of each frequency band in the time-frequency domain. The second set of feature parameters includes the peak value, frequency centroid, number of spectral peaks, spectral peak width, and energy ratio of each frequency band signal.

[0018] Specifically, in step S3, before constructing the multi-source feature vector, the first feature parameter and each second feature parameter set are normalized. The normalization process uses min-max normalization or z-score normalization to eliminate the influence of different dimensions on the feature vector. After the normalization process, principal component analysis is used to screen features, remove redundant features, and retain the top k features with the highest correlation to the type and severity of partial discharge, thereby reducing the computational complexity of the model.

[0019] Specifically, in step S4, the fusion diagnostic model is a deep learning-based fusion model, including a feature fusion layer, a feature extraction layer, and a classification and regression layer. The feature fusion layer uses an attention mechanism to perform weighted fusion of the first feature parameter and each set of second feature parameters. The feature extraction layer is connected to the feature fusion layer, receives the fused features after attention-weighted fusion, and performs multi-layer nonlinear transformation, deep feature abstraction, and discriminative feature enhancement on them. The classification and regression layer is connected to the feature extraction layer, receives high-order deep features, and performs joint diagnostic output.

[0020] Specifically, in step S4, the positioning information is calculated using the time delay difference positioning algorithm of the ultrasonic sensor array. Specifically, based on the time difference between the acquisition of the same partial discharge signal by each ultrasonic sensor and the arrangement coordinates of the sensor array, the specific location of the partial discharge is calculated using the triangulation method.

[0021] Specifically, in step S4, the joint judgment logic also includes a hierarchical decision-making mechanism based on fuzzy reasoning or decision trees. It determines whether there is partial discharge activity based on the activity level of the transient ground voltage signal. If the activity level is lower than the background noise threshold, it outputs a no-discharge state. If there is discharge, it further determines the discharge type based on the center frequency offset and spectral characteristics of the ultrasonic signal, and assesses the severity level by combining the multi-band energy ratio and the transient ground voltage pulse phase characteristics.

[0022] Specifically, in step S5, the preset threshold is 5%-15%, the online fine-tuning of model parameters adopts the small sample transfer learning method, and only fine-tunes the output layer and feature fusion layer parameters of the fusion diagnostic model to avoid model overfitting. In the online fine-tuning process, an incremental learning mechanism is introduced to add the calibrated data samples to the training cache pool and periodically retrain the model asynchronously to continuously improve the model's adaptability to specific devices or environments.

[0023] In summary, the advantages of the portable partial discharge detection method integrating TEV and ultrasonic multi-band signals proposed in this application are as follows:

[0024] 1. The system employs a TEV sensor and an ultrasonic multi-band sensor array for collaborative acquisition, covering the characteristic signal frequency bands of different types and severity of partial discharge. It can capture the overall activity of partial discharge through TEV signals and uncover detailed features of partial discharge through multi-band ultrasonic signals, significantly improving the comprehensiveness and accuracy of detection.

[0025] 2. Amplitude statistics and pulse phase analysis are performed on TEV signals, and in-depth time-frequency analysis is conducted on ultrasound signals of each frequency band to fully explore the deep characteristics of partial discharge. At the same time, high-quality multi-source feature vectors are constructed to improve the computational efficiency of subsequent fusion diagnostic models, thereby effectively improving the accuracy of partial discharge type identification and enabling the effective capture of weak partial discharge signals to meet the needs of early warning.

[0026] 3. Adopting a more generalizable fusion diagnostic model, combined with multi-feature association judgment logic, it adapts to the testing needs of different types of power equipment and different field environments. At the same time, it sets up on-site verification and online calibration mechanisms. When there is a deviation between the test results and the actual situation, calibration can be achieved by fine-tuning the model parameters online, ensuring the reliability of the test results. Attached Figure Description

[0027] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0028] Figure 1 This is a flowchart of a portable partial discharge detection method that integrates TEV and ultrasonic multi-band signals according to this application;

[0029] Figure 2 This is a flowchart illustrating the feature processing of a portable partial discharge detection method that integrates TEV and ultrasonic multi-band signals according to this application.

[0030] Figure 3 This is a flowchart of the fusion diagnostic model for a portable partial discharge detection method that integrates TEV and multi-band ultrasound signals, as described in this application. Detailed Implementation

[0031] To make the technical means, inventive features, objectives, and effects of this application easier to understand, the application is further described below with reference to specific illustrations. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0032] The present application will now be described in further detail with reference to the accompanying drawings.

[0033] like Figures 1-3As shown in the figure, a portable partial discharge detection method integrating TEV and ultrasonic multi-band signals according to an embodiment of this application includes the following steps: S1, using a transient ground voltage sensor and an ultrasonic sensor array to simultaneously acquire transient ground voltage signals and multi-band ultrasonic signals at the power equipment under test.

[0034] It should be noted that the transient ground voltage sensor adopts a non-contact design and is arranged to fit the metal shell of the power equipment under test (such as switchgear, transformer, GIS equipment, etc.) to ensure good coupling between the sensor and the equipment surface. The acquisition frequency is set to 100MHz-1GHz, which can effectively capture the transient ground voltage pulse signal generated during partial discharge and avoid signal attenuation caused by poor contact.

[0035] The ultrasonic sensor array consists of 3-6 ultrasonic sensors arranged in a ring or linear pattern to cover critical areas where partial discharge may occur in the equipment. The acquisition frequency band is divided into multiple intervals (e.g., 20-40kHz, 40-80kHz, 80-150kHz) to achieve synchronous acquisition of multi-band ultrasonic signals. This allows for the capture of characteristic ultrasonic signals generated by different types of partial discharge (e.g., corona discharge, surface discharge) and provides data support for subsequent positioning through array layout. During acquisition, a synchronization trigger module controls the acquisition timing of the two types of sensors to ensure that the timestamps of the two signals are consistent.

[0036] S2. Perform amplitude statistics and pulse phase analysis on the acquired transient ground voltage signal to extract the first feature parameter; perform time-frequency analysis and feature extraction on the acquired ultrasonic signals of each frequency band to obtain the second feature parameter set corresponding to each frequency band.

[0037] It should be noted that the amplitude statistics of transient ground voltage signals include calculating indicators such as the peak value, peak mean, peak variance, number of pulses, and pulse amplitude distribution probability of the signal, which are used to characterize the intensity and activity of partial discharge; pulse phase analysis, on the other hand, extracts features such as phase distribution entropy, phase concentration coefficient, and positive and negative half-cycle pulse asymmetry by drawing pulse phase distribution maps, and combines them with the working voltage phase of the power equipment to preliminarily determine the approximate type of partial discharge.

[0038] For ultrasonic signals in each frequency band, wavelet transform and short-time Fourier transform are used for time-frequency analysis to convert the time-domain signal into a time-frequency domain spectrum. Characteristic parameters such as center frequency, peak energy, spectral width, time-domain peak value, pulse duration, and waveform distortion rate of each frequency band are extracted to form a second set of characteristic parameters for the corresponding frequency band. Considering the characteristic differences of different frequency bands, the characteristic parameters of each frequency band are weighted to highlight the characteristic weight of high-sensitivity frequency bands (such as the 40-80kHz frequency band in the partial discharge characteristic signal set) and reduce the interference of invalid frequency bands.

[0039] S3. Construct multi-source feature vectors based on the first feature parameter and each set of second feature parameters.

[0040] It should be noted that the first feature parameter and the second feature parameters of each frequency band are spliced ​​together in a preset order to form a multi-source feature vector. This vector includes both the overall activity characteristics of partial discharge reflected by the transient ground voltage signal and the detailed characteristics of partial discharge reflected by the ultrasound signals of each frequency band, realizing the complementary fusion of multi-source information and providing comprehensive feature support for subsequent accurate diagnosis.

[0041] S4. Input the multi-source feature vectors into the pre-trained fusion diagnostic model. The fusion diagnostic model outputs the partial discharge type identification result, severity level and location information according to the built-in joint judgment logic. The joint judgment logic includes at least the association rules based on the activity of transient ground voltage signal and the center frequency and spectral features of ultrasonic signal.

[0042] It should be noted that the joint judgment logic, based on the association rules between TEV and the core features of the ultrasound signal, also incorporates threshold judgment and a multi-feature voting mechanism. When the TEV signal activity is higher than a preset threshold, and the center frequency of the ultrasound signal is between 40-80kHz with concentrated peak energy, it is preferentially judged as surface discharge. When the number of TEV signal pulses is large, the phase distribution is dispersed, and weak signals appear in multiple frequency bands of the ultrasound signal, it is judged as corona discharge. The severity level of partial discharge is classified according to characteristic parameters such as TEV peak value and ultrasound peak energy (e.g., mild, moderate, severe). The location information is calculated by the signal arrival time difference algorithm of the ultrasound sensor array, combined with the intensity distribution of the TEV signal, to achieve accurate location of the partial discharge.

[0043] S5. Compare and verify the results output by the fusion diagnostic model with the actual on-site detection. If the deviation between the identification result and the actual situation exceeds the preset threshold, calibrate by fine-tuning the model parameters online. If the deviation is within the allowable range, the detection result is confirmed to be valid.

[0044] It should be noted that the actual on-site testing results are obtained through manual inspection (such as equipment appearance inspection and infrared temperature measurement assistance) and cross-verification using the partial discharge pulse current method to ensure the accuracy of the actual situation. The preset thresholds are set according to the equipment type, industry standards and on-site testing experience. Typically, the type identification accuracy deviation threshold is set to 10%, the positioning deviation threshold is set to 10cm, and the severity level deviation does not exceed level 1.

[0045] When the deviation exceeds the preset threshold, a small-sample online fine-tuning method is adopted. On-site deviation samples are selected as incremental training data to adjust the feature weights and judgment thresholds of the fusion diagnostic model. There is no need to retrain the entire model, ensuring calibration efficiency. If the deviation is within the allowable range, the detection results (including partial discharge type, severity, location information, and detection time) are stored in the system database and a detection report is generated to provide data support for the operation and maintenance of power equipment. At the same time, effective samples are added to the training dataset to further optimize model performance.

[0046] In one embodiment of this application, in step S1, synchronous acquisition is achieved through GPS timing or a high-precision clock synchronization module, and the time synchronization error between the transient ground voltage signal and the multi-band ultrasonic signal does not exceed 1μs.

[0047] Specifically, the GPS timing module receives precise time signals from satellites and calibrates the clock of the acquisition device in real time, ensuring complete consistency of the time base across different acquisition channels. This is suitable for monitoring equipment in open outdoor environments. The high-precision clock synchronization module uses the PTP precision time protocol, achieving sub-millisecond time synchronization through the device's internal clock synchronization circuit. This is more suitable for enclosed indoor environments or environments with weak satellite signals. The time synchronization error is controlled within 1μs because the duration of partial discharge signals is typically in the microsecond to millisecond range. Extremely small time deviations can cause discrepancies in the timing correlation between transient voltage signals and ultrasonic signals, thus affecting the accuracy of subsequent feature extraction. This makes it impossible to accurately correspond the electrical and acoustic signals of the same partial discharge event, reducing the reliability of diagnostic results.

[0048] In one embodiment of this application, in step S2, the amplitude statistics include statistical analysis of the peak value, effective value, kurtosis, and waveform factor of the transient ground voltage signal, and the pulse phase analysis adopts the phase-resolved partial discharge spectrum analysis method to extract the pulse phase distribution characteristics as the first feature parameter.

[0049] It should be noted that the peak value reflects the maximum instantaneous intensity of the transient ground voltage signal and is directly related to the peak energy release of partial discharge; the effective value is used to characterize the average energy level of the signal, avoiding the random influence of a single peak value; kurtosis is used to describe the steepness of the signal waveform, which can effectively identify the difference between partial discharge pulses and background noise. The higher the kurtosis value, the more concentrated the pulse signal and the stronger the distinguishability; the waveform factor reflects the regularity of the signal waveform through the ratio of the peak value to the effective value. Different types of partial discharge will exhibit different waveform factor characteristics.

[0050] The phase-resolved partial discharge spectrum used in pulse phase analysis uses the grid voltage phase as the horizontal axis and the discharge pulse amplitude as the vertical axis. By statistically analyzing the number and amplitude distribution of discharge pulses in different phase intervals, features such as phase concentration intervals, phase symmetry, and pulse amplitude dispersion are extracted as the first feature parameters. These features can effectively distinguish different types of partial discharges and provide core basis for subsequent diagnosis.

[0051] In one embodiment of this application, in step S2, the time-frequency analysis uses wavelet transform or short-time Fourier transform to decompose the ultrasonic signals of each frequency band in the time-frequency domain. The second set of feature parameters includes the peak value, frequency centroid, number of spectral peaks, spectral peak width, and energy ratio of each frequency band signal.

[0052] It should be noted that wavelet transform has good time-frequency localization characteristics, which can effectively decompose non-stationary ultrasonic signals into wavelet components of different frequency bands, accurately capturing the instantaneous frequency changes and amplitude characteristics of partial discharge ultrasonic signals. It is especially suitable for the analysis of short-duration, high-frequency partial discharge acoustic signals. Short-time Fourier transform, on the other hand, performs piecewise Fourier transform on the signal through a sliding time window, which can clearly present the frequency distribution law of ultrasonic signals in different time periods, making it convenient to analyze the duration and frequency evolution of partial discharge.

[0053] The second set of characteristic parameters is as follows: the peak value of each frequency band reflects the strongest intensity of the ultrasonic signal in that band and is related to the energy level of the partial discharge; the frequency centroid characterizes the frequency position where the signal energy is concentrated, and there are significant differences in the ultrasonic frequency centroids of different types of partial discharge; the number and width of spectral peaks reflect the complexity of the signal frequency components and can be used to distinguish partial discharge from interference signals; the energy proportion reflects the contribution ratio of each frequency band signal to the total energy, which can reflect the energy distribution characteristics of partial discharge and provide an important reference for the identification of partial discharge type and the assessment of severity.

[0054] In one embodiment of this application, before constructing the multi-source feature vector in step S3, the first feature parameter and each second feature parameter set are normalized. The normalization process adopts the min-max normalization or z-score normalization method to eliminate the influence of different dimensions on the feature vector. After the normalization process, principal component analysis is used to screen features, remove redundant features, and retain the top k-dimensional features with the highest correlation with the type and severity of partial discharge, thereby reducing the computational complexity of the model.

[0055] It should be noted that, due to the significant difference in dimensions between the first and second feature parameters—for example, the pulse phase feature ranges from 0 to 360°, while the peak value of an ultrasonic signal may be tens to hundreds of millivolts—direct fusion can lead to the larger-dimensional feature dominating and masking the smaller-dimensional but important discriminative feature. Min-max normalization maps all feature values ​​to the [0,1] interval, suitable for scenarios with relatively uniform feature distribution. Z-score normalization standardizes features based on their mean and standard deviation, making them conform to a standard normal distribution, suitable for scenarios with outliers.

[0056] After normalization, principal component analysis is used for feature selection. By calculating the covariance matrix of the features, the top k principal components with the largest variance contribution rate are extracted. These principal components can retain the key information of the original features to the maximum extent, while eliminating redundant correlations between different features and reducing the dimension of the feature vector. This not only reduces the computational complexity of the fusion diagnostic model and shortens the training and inference time, but also avoids the model overfitting problem caused by redundant features and improves the model's generalization ability.

[0057] In one embodiment of this application, in step S4, the fusion diagnostic model is a deep learning-based fusion model, including a feature fusion layer, a feature extraction layer, and a classification and regression layer. The feature fusion layer uses an attention mechanism to perform weighted fusion of the first feature parameter and each set of second feature parameters. The feature extraction layer is connected to the feature fusion layer, receives the fused features after attention-weighted fusion, and performs multi-layer nonlinear transformation, deep feature abstraction, and discriminative feature enhancement on them. The classification and regression layer is connected to the feature extraction layer, receives high-order deep features, and performs joint diagnostic output.

[0058] Specifically, the attention mechanism can adaptively allocate the weights of each feature. Based on the correlation between different features and the partial discharge diagnosis task, it assigns higher weights to important features (such as phase features strongly correlated with discharge type and energy percentage features strongly correlated with severity) and lower weights to minor or interfering features. This solves the problem of low feature utilization caused by fixed weights in traditional feature fusion and achieves efficient fusion of multi-source features.

[0059] The feature extraction layer adopts a convolutional neural network or recurrent neural network structure. Through multi-layer convolution, pooling or iterative operations, it performs in-depth processing on the fused features, strips away redundant surface information, and abstracts high-order deep features with strong discriminative power, thereby enhancing the feature differences between different types of partial discharges and different degrees of severity.

[0060] The classification and regression layer uses a fully connected layer combined with a softmax classification function and a linear regression function. The classification part is responsible for outputting the specific type of partial discharge (such as corona discharge, surface discharge, internal discharge, etc.), while the regression part is responsible for outputting the severity level of partial discharge (such as mild, moderate, severe). This enables joint diagnosis of partial discharge type identification and severity assessment, improving the comprehensiveness and accuracy of the diagnostic results.

[0061] In one embodiment of this application, in step S4, the positioning information is calculated by the time delay difference positioning algorithm of the ultrasonic sensor array. Specifically, based on the time difference of the same partial discharge signal collected by each ultrasonic sensor and combined with the arrangement coordinates of the sensor array, the specific location of the partial discharge is calculated by the triangulation positioning method.

[0062] It should be noted that when partial discharge occurs, an ultrasonic signal is generated. This signal propagates outwards at a certain speed, and ultrasonic sensors at different locations receive the signal sequentially. By calculating the time difference between the signals received by each sensor using the cross-correlation function method, the order and time interval of the signal arrival at each sensor are determined. Based on the triangulation method, a circle is drawn with two sensors as centers and the product of the signal propagation speed and the time difference as the radius. The intersection of the two circles is the approximate location of the partial discharge. Combining this with the time difference from a third sensor can further determine the unique discharge location, achieving precise localization of the partial discharge with centimeter-level accuracy. This provides clear location guidance for subsequent equipment maintenance, improving maintenance efficiency.

[0063] In one embodiment of this application, in step S4, the joint judgment logic further includes a hierarchical decision-making mechanism based on fuzzy reasoning or decision tree, which determines whether there is partial discharge activity based on the activity level of the transient ground voltage signal. If the activity level is lower than the background noise threshold, a no-discharge state is output. If there is discharge, the discharge type is further determined based on the center frequency offset and spectral characteristics of the ultrasonic signal, and the severity level is jointly assessed by combining the multi-band energy ratio and the transient ground voltage pulse phase characteristics.

[0064] Specifically, the hierarchical decision-making mechanism adopts a logic of first determining presence or absence, then identifying the type, and finally evaluating the level, thereby reducing diagnostic errors and improving the rationality of decisions. The activity level of the transient ground voltage signal is calculated using parameters such as the number of pulses per unit time and the total pulse amplitude. The background noise threshold is determined through long-term signal monitoring statistics under no-discharge conditions, ensuring the accuracy of no-discharge condition judgment and avoiding misjudging background noise as partial discharge.

[0065] When a discharge is detected, the center frequency offset of the ultrasonic signal is used to distinguish different types of partial discharge. For example, the center frequency of corona discharge is usually between 100-300kHz, while the center frequency of surface discharge is between 50-150kHz. The degree of center frequency offset can reflect the stability of the discharge.

[0066] Spectral features further refine the discharge type through the shape of the time-frequency spectrum and the distribution of spectral peaks. In severity assessment, a higher proportion of multi-band energy indicates greater partial discharge energy, and a more concentrated transient ground voltage pulse phase indicates more frequent and stable discharges. Combining these two factors allows for precise classification into mild, moderate, and severe levels, providing a scientific basis for equipment operation and maintenance. The application of fuzzy reasoning or decision trees can effectively handle uncertainties in the diagnostic process (such as fluctuations in characteristic parameters and the influence of interference signals).

[0067] In one embodiment of this application, in step S5, the preset threshold is 5%-15%, and the online fine-tuning of model parameters adopts the small sample transfer learning method, which only fine-tunes the output layer and feature fusion layer parameters of the fusion diagnostic model to avoid model overfitting. During the online fine-tuning process, an incremental learning mechanism is introduced, and the calibrated data samples are added to the training cache pool. The model is periodically retrained asynchronously to continuously improve the model's adaptability to specific devices or environments.

[0068] Specifically, a preset threshold of 5%-15% is used to determine the degree of deviation between the model's diagnostic results and the actual calibration results. When the deviation exceeds this threshold, the model is triggered to fine-tune online to ensure that the model's diagnostic accuracy always meets the actual monitoring requirements. The specific value of the threshold can be adaptively adjusted according to the equipment type and operating environment. For example, outdoor equipment is more susceptible to environmental interference, so the threshold can be set to 10%-15%, while indoor equipment is less susceptible to interference, so it can be set to 5%-10%.

[0069] The few-shot transfer learning method is based on the parameters of a pre-trained model. It does not require retraining the entire model, but only fine-tunes the parameters of the output layer and the feature fusion layer. It can adapt to the characteristics of small-sample calibration data in the field, avoid model overfitting caused by insufficient training data, and significantly shorten the fine-tuning time, ensuring that the model can quickly adapt to the actual field situation.

[0070] The incremental learning mechanism adds valid data samples (manually verified partial discharge data) after each calibration to the training cache pool and performs asynchronous retraining on the model periodically (e.g., weekly, monthly), gradually supplementing the model's training data. This enables the model to continuously learn the partial discharge characteristics under different operating conditions and environments, continuously improving the model's adaptability to specific equipment or environments. It solves the feature drift problem caused by equipment aging and environmental changes during long-term operation, ensuring the long-term reliability and accuracy of the diagnostic model.

[0071] It should be noted that, in this document, the terms “comprising,” “including,” or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0072] The present application and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present application. The actual structure is not limited to this. In conclusion, if a person skilled in the art is inspired by this description and designs a similar structure and embodiment without departing from the spirit of the present application, such design should fall within the protection scope of the present application.

Claims

1. A portable partial discharge detection method integrating TEV and ultrasonic multi-band signals, characterized in that, Includes the following steps: S1. Using a transient ground voltage sensor and an ultrasonic sensor array, the transient ground voltage signal and multi-band ultrasonic signal at the power equipment under test are collected simultaneously. S2. Perform amplitude statistics and pulse phase analysis on the acquired transient ground voltage signal to extract the first feature parameter; Time-frequency analysis and feature extraction are performed on the collected ultrasonic signals of each frequency band to obtain the second feature parameter set corresponding to each frequency band; S3. Construct a multi-source feature vector based on the first feature parameter and each of the second feature parameter sets; S4. Input the multi-source feature vector into the pre-trained fusion diagnostic model. The fusion diagnostic model outputs the partial discharge type identification result, severity level and location information according to the built-in joint judgment logic. The joint judgment logic includes at least the association rules based on the activity of transient ground voltage signal and the center frequency and spectral features of ultrasonic signal. S5. Compare and verify the results output by the fusion diagnostic model with the actual on-site detection. If the deviation between the identification result and the actual situation exceeds the preset threshold, calibrate by fine-tuning the model parameters online. If the deviation is within the allowable range, the detection result is confirmed to be valid.

2. The portable partial discharge detection method integrating TEV and ultrasonic multi-band signals according to claim 1, characterized in that, In step S1, the synchronous acquisition is achieved through GPS timing or a high-precision clock synchronization module, and the time synchronization error between the transient ground voltage signal and the multi-band ultrasonic signal does not exceed 1μs.

3. The portable partial discharge detection method integrating TEV and ultrasonic multi-band signals according to claim 1, characterized in that, In step S2, the amplitude statistics include statistical analysis of the peak value, effective value, kurtosis, and waveform factor of the transient ground voltage signal. The pulse phase analysis uses the phase-resolved partial discharge spectrum analysis method to extract the pulse phase distribution characteristics as the first feature parameter.

4. The portable partial discharge detection method integrating TEV and ultrasonic multi-band signals according to claim 1, characterized in that, In step S2, the time-frequency analysis uses wavelet transform or short-time Fourier transform to decompose the ultrasonic signals of each frequency band in the time-frequency domain. The second set of feature parameters includes the peak value, frequency centroid, number of spectral peaks, spectral peak width, and energy ratio of each frequency band signal.

5. The portable partial discharge detection method integrating TEV and ultrasonic multi-band signals according to claim 1, characterized in that, In step S3, before constructing the multi-source feature vector, the first feature parameter and each second feature parameter set are normalized. The normalization process uses min-max normalization or z-score normalization to eliminate the influence of different dimensions on the feature vector. After normalization, principal component analysis is used to screen features, remove redundant features, and retain the top k features with the highest correlation to the type and severity of partial discharge, thereby reducing the computational complexity of the model.

6. The portable partial discharge detection method integrating TEV and ultrasonic multi-band signals according to claim 1, characterized in that, In step S4, the fusion diagnostic model is a deep learning-based fusion model, including a feature fusion layer, a feature extraction layer, and a classification and regression layer. The feature fusion layer uses an attention mechanism to perform weighted fusion of the first feature parameter and each set of second feature parameters. The feature extraction layer is connected to the feature fusion layer, receives the fused features after attention-weighted fusion, and performs multi-layer nonlinear transformation, deep feature abstraction, and discriminative feature enhancement on them. The classification and regression layer is connected to the feature extraction layer, receives high-order deep features, and performs joint diagnostic output.

7. The portable partial discharge detection method integrating TEV and ultrasonic multi-band signals according to claim 1, characterized in that, In step S4, the positioning information is calculated using the time delay difference positioning algorithm of the ultrasonic sensor array. Specifically, based on the time difference between the acquisition of the same partial discharge signal by each ultrasonic sensor and the arrangement coordinates of the sensor array, the specific location of the partial discharge is calculated using the triangulation method.

8. The portable partial discharge detection method integrating TEV and ultrasonic multi-band signals according to claim 1, characterized in that, In step S4, the joint judgment logic also includes a hierarchical decision-making mechanism based on fuzzy reasoning or decision tree. It judges whether there is partial discharge activity based on the activity of transient ground voltage signal. If the activity is lower than the background noise threshold, it outputs a no-discharge state. If there is discharge, it further judges the discharge type based on the center frequency offset and spectrum characteristics of ultrasonic signal, and jointly evaluates the severity level by combining the multi-band energy ratio and transient ground voltage pulse phase characteristics.

9. A portable partial discharge detection method integrating TEV and ultrasonic multi-band signals according to claim 1, characterized in that, In step S5, the preset threshold is 5%-15%, and the online fine-tuning of model parameters adopts the small sample transfer learning method, which only fine-tunes the output layer and feature fusion layer parameters of the fusion diagnostic model to avoid model overfitting. In the online fine-tuning process, an incremental learning mechanism is introduced to add the calibrated data samples to the training cache pool and periodically retrain the model asynchronously to continuously improve the model's adaptability to specific devices or environments.

Images