A method and system for identifying partial discharge of valve-side casing based on artificial intelligence

Abstract

The present invention discloses an artificial intelligence-based method and system for identifying partial discharge in valve-side bushings, which relates to the technical field of power equipment status monitoring and fault diagnosis. The method comprises the following steps: synchronously collecting multi-source signals; performing pre-processing and feature fusion based on the multi-source signals to obtain a fused feature matrix for partial discharge identification; and identifying and classifying partial discharge types using an artificial intelligence hybrid model based on the fused feature matrix for partial discharge identification. The present invention improves the model's robustness in identifying multiple types of discharge behaviors in complex scenarios, provides subsequent artificial intelligence models with high-quality input with more complete semantic expression and more reasonable structure, achieves accurate identification and stable classification of partial discharge types, effectively reduces the probability of false alarms and missed alarms, and possesses online learning and model evolution capabilities, enabling the system to maintain a high recognition rate during long-term operation, thus meeting the dual requirements of online monitoring accuracy and adaptability for smart grids.

Description

Technical Field

[0001] The present invention relates to the technical field of power equipment status monitoring and fault diagnosis, and in particular to an artificial intelligence-based valve-side bushing partial discharge identification method and system. Background Art

[0002] As power systems continue to expand and voltage levels continue to rise, the insulation reliability of large transformers and their key components has become a core technical issue for ensuring the safe operation of power grids. Transformer bushings, as critical insulating components connecting the main transformer to the external power grid, carry the burden of transmitting high voltage and high current. Their insulation condition is directly related to the operational stability of power equipment. Bushings on the DC converter valve side, in particular, are vulnerable areas prone to partial discharge due to their unique structure and operating conditions. Partial discharge, a typical precursor to insulation degradation in electrical equipment, is crucial for early detection and accurate identification of partial discharges, ensuring fault warning and condition-based maintenance. Currently, methods for detecting partial discharges have evolved from traditional single-physical quantity sensing methods, such as pulse current, ultrasonic testing, and electromagnetic radiation detection, to a complex detection system that integrates multimodal collaborative sensing and intelligent identification. Specifically, ultra-high frequency (UHF) technology enables sensitive capture of electromagnetic radiation signals and is suitable for identifying far-field high-frequency signals. Ultrasonic sensing technology has excellent spatial positioning characteristics and can be used to determine the location of partial discharge sources. Transient earth voltage (TEV) signals reflect the characteristics of partial discharge current propagation along the equipment grounding path and are particularly sensitive to near-field partial discharges. Although various detection methods have achieved certain application results in laboratories and engineering sites, the increasingly complex operating environment of power equipment, the diverse types of partial discharges, and the non-stationary signal forms make a single detection signal have a low signal-to-noise ratio in complex interference backgrounds, which can easily lead to incomplete information and unstable identification results.

[0003] To improve the accuracy and robustness of partial discharge identification, researchers have recently introduced multi-source signal fusion and artificial intelligence modeling techniques, striving to overcome the limitations of traditional methods in terms of information dimensionality and spatiotemporal resolution. Typical research approaches include: joint judgment methods based on feature fusion, multi-channel time-domain synchronization mechanisms, direction-of-arrival (DOA) positioning algorithms, and end-to-end classification models based on deep learning. In particular, in the area of ​​intelligent identification, convolutional neural networks (CNNs), long short-term memory networks (LSTMs), and their variants have gradually become the mainstream algorithmic frameworks in fault diagnosis. Although some academic research has attempted to jointly analyze UHF, TEV, and ultrasonic signals and apply deep learning models for identification, the following common problems still exist: Signal fusion methods often rely on simple concatenation or weighted averaging, failing to consider the differences in the importance of spatiotemporal feature expression across modal signals, resulting in limited value for the fused feature information; inadequate pre-processing design and insufficient noise removal lead to unstable model training performance and weak generalization; deep models mostly use a single network structure, making it difficult to simultaneously account for local transient features and long-term trends, making it difficult to meet the requirements for stable identification in multiple types of partial discharge scenarios; and most existing methods lack dynamic model updates and online adaptability, making them unable to cope with changes in the sensing environment, electrical load, equipment aging, and other factors in actual operating conditions. Therefore, how to construct a partial discharge identification method that combines multi-source information collaborative modeling, accurate feature extraction, multimodal fusion expression, and intelligent, evolvable recognition capabilities has become a research hotspot and a technical bottleneck in the field of high-voltage equipment state perception. Summary of the Invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by the present invention is: the existing partial discharge online monitoring method has the problems of single signal perception dimension, limited feature extraction quality, single intelligent recognition model structure and lack of adaptive ability, and how to realize a partial discharge type intelligent identification solution based on synchronous acquisition, differentiated feature extraction and dynamic fusion of multimodal high-frequency signals, construct an artificial intelligence hybrid model that integrates local and global features, and has self-learning and model optimization mechanisms.

[0006] To solve the above technical problems, the present invention provides the following technical solutions: a method for identifying partial discharge of a valve-side casing based on artificial intelligence, comprising synchronously collecting multi-source signals; performing pre-processing and feature fusion based on the multi-source signals to obtain a fused feature matrix for partial discharge identification; identifying and classifying partial discharge types based on the fused feature matrix of partial discharge identification through an artificial intelligence hybrid model; the fused feature matrix of partial discharge identification comprises adjusting the attention weight coefficient of the fused feature matrix, dynamically evaluating the contribution of modal features to discharge type identification through the attention mechanism, and obtaining a weight-enhanced fused feature matrix; the artificial intelligence hybrid model comprises extracting local features from the fused feature matrix using a one-dimensional convolutional neural network, extracting local spatial and short-term temporal information features from the fused feature matrix using a one-dimensional convolution kernel to obtain a local feature representation, extracting long-term dependent global temporal features from the local feature representation to form a global temporal feature representation, fusing the local feature representation and the global temporal feature representation through a jump connection to form a mixed feature vector containing both short-term local and long-term global information, inputting the mixed feature vector into a classification layer, and outputting the partial discharge type identification and classification results through a Softmax function.

[0007] As a preferred solution of the artificial intelligence-based valve-side casing partial discharge identification method described in the present invention, the synchronous acquisition of multi-source signals includes acquiring electromagnetic wave signals through a UHF sensor, acquiring ultrasonic signals through an ultrasonic sensor, and acquiring transient ground voltage signals through a transient ground voltage sensor. During the sampling process, a consistent time tag is assigned to each set of data, and the data is acquired synchronously.

[0008] As a preferred solution of the artificial intelligence-based valve-side casing partial discharge identification method described in the present invention, the pre-processing includes: denoising the electromagnetic wave signal using a wavelet packet decomposition method, extracting the time-frequency domain characteristic parameters of the electromagnetic wave signal, spatially locating the ultrasonic signal based on a direction of arrival algorithm, and extracting the pulse amplitude and change rate characteristics of the transient ground voltage signal.

[0009] Feature fusion includes aligning the time-frequency domain characteristic parameters of the electromagnetic wave signal and the pulse amplitude and change rate characteristics of the transient ground voltage signal to obtain a feature vector.

[0010] As a preferred solution of the artificial intelligence-based valve-side casing partial discharge identification method described in the present invention, the fusion feature matrix of the partial discharge identification includes calculating the correlation weight coefficients between feature vectors based on the attention mechanism, performing weighted fusion on the feature vectors, and generating a weighted fusion feature matrix.

[0011] As a preferred solution of the artificial intelligence-based valve side casing partial discharge identification method described in the present invention, the artificial intelligence hybrid model includes: using a residual network to extract spatial features of the fused feature matrix to obtain a multi-scale spatial feature map, using a Transformer encoder to extract temporal features of the fused feature matrix to obtain a long-term dependent temporal feature representation, and performing feature cascade on the spatial feature map and the temporal feature representation to form a comprehensive feature vector that integrates spatial and temporal information.

[0012] As a preferred solution of the artificial intelligence-based valve-side casing partial discharge identification method described in the present invention, the method includes dynamically adjusting the alarm threshold according to the discharge type and the current load rate, obtaining the real-time operating load rate information of the current equipment, calling the preset alarm threshold based on the discharge type, matching the adjustment factor according to the real-time operating load rate information, using the adjustment factor to perform weighted correction on the reference threshold in the alarm threshold, dynamically adjusting the alarm threshold, extracting the fusion feature parameters in the fusion feature vector, comparing the fusion feature parameters with the alarm threshold, and determining whether to trigger an alarm.

[0013] As a preferred solution of the artificial intelligence-based valve-side casing partial discharge identification method described in the present invention, the adaptive optimization of the artificial intelligence hybrid model includes real-time collection of fusion feature matrices and actual discharge type labels, updating historical data sets, retraining the updated historical data sets, and using the gradient descent method to optimize the network weights and bias parameters in the artificial intelligence hybrid model. Based on the latest training results, the model weight parameters are dynamically updated, and the probability threshold of the Softmax classifier is adjusted in real time. Through continuous iterative updates, the parameters of the artificial intelligence hybrid model are adaptively optimized.

[0014] Another object of the present invention is to provide an artificial intelligence-based valve-side casing partial discharge identification system, which can realize the synchronous acquisition, feature fusion and intelligent identification and early warning of multi-modal high-frequency signals, and solve the problems in the existing technology of difficult to accurately identify the type of partial discharge signal, strong noise interference and insufficient spatial positioning accuracy.

[0015] As a preferred solution of the artificial intelligence-based valve-side casing partial discharge identification system described in the present invention, it includes: a signal acquisition module, a feature processing module, and an intelligent identification module; the signal acquisition module includes a multi-source acquisition unit and a time domain synchronization unit, the multi-source acquisition unit is used to synchronously acquire ultra-high frequency, ultrasonic, and transient ground voltage signals, and the time domain synchronization unit is used to add a unified timestamp to the acquired signals to achieve time domain alignment of multi-source signals; the feature processing module includes a pre-noise reduction unit and a feature fusion unit, the pre-noise reduction unit is used to remove noise from the multi-modal original signal and extract effective features, and the feature fusion unit is used to weightedly fuse the multi-modal features through an attention mechanism to form a fusion feature matrix; the intelligent identification module includes a hybrid modeling unit and an adaptive optimization unit, the hybrid modeling unit is used to use an artificial intelligence hybrid model to extract and classify features of the fusion feature matrix, and the adaptive optimization unit is used to optimize model parameters in real time according to field feedback data.

[0016] A computer device includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement a method for identifying partial discharge of a valve-side casing based on artificial intelligence.

[0017] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of a method for identifying partial discharge of a valve-side casing based on artificial intelligence.

[0018] The beneficial effects of the present invention are as follows: by setting a unified time reference and synchronously collecting multi-source signals, the parallel acquisition of electromagnetic wave signals, ultrasonic signals and transient ground voltage signals in a unified time window is realized, thereby establishing a multi-modal, highly time-consistent original data foundation; by accurately extracting local discharge features and constructing a fusion feature matrix in a weighted fusion manner, the model's robustness in identifying multiple types of discharge behaviors in complex scenarios is improved, and high-quality input with more complete semantic expression and more reasonable structure is provided for subsequent artificial intelligence models, thereby enhancing the classifier's ability to recognize discharge patterns under weak, multi-source interference conditions; by constructing an artificial intelligence hybrid architecture with spatiotemporal collaborative modeling capabilities, accurate identification and stable classification of local discharge types are achieved, effectively reducing the probability of false alarms and missed alarms; and at the same time, with online learning and model evolution capabilities, the system can maintain a high recognition rate during long-term operation, meeting the dual requirements of smart grids for online monitoring accuracy and adaptability. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0020] Figure 1 This is an overall flow chart of an artificial intelligence-based method for identifying partial discharge of a valve-side casing according to the first embodiment of the present invention. DETAILED DESCRIPTION

[0021] To make the above-mentioned objects, features, and advantages of the present invention more clearly understood, the following detailed description of the specific embodiments of the present invention is given in conjunction with the accompanying drawings. It is obvious that the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in this field without creative work should fall within the scope of protection of the present invention.

[0022] Example 1, with reference to Figure 1 , as one embodiment of the present invention, provides a valve-side bushing partial discharge identification method based on artificial intelligence, comprising:

[0023] S1: Synchronously collect multi-source signals.

[0024] Furthermore, synchronously collecting multi-source signals includes collecting electromagnetic wave signals through ultra-high frequency sensors, collecting ultrasonic signals through ultrasonic sensors, and collecting transient voltage signals through transient voltage sensors. During the sampling process, a consistent time tag is assigned to each set of data, and data is collected synchronously.

[0025] It should be noted that electromagnetic radiation signals generated in the 300MHz to 3GHz frequency range during partial discharge in the valve-side casing are collected using ultra-high frequency sensors to reflect the far-field high-frequency electromagnetic energy distribution characteristics during the discharge process. Ultrasonic sensors collect acoustic emission signals in the 20kHz to 200kHz frequency range during partial discharge to capture the spatial propagation path of the discharge source within the casing structure and changes in acoustic pressure amplitude. Transient ground voltage sensors collect transient current pulse signals generated in the casing ground wire accompanying partial discharge to identify the coupling effect and pulse response morphology of the discharge along the ground loop. During the sampling process, to ensure the consistency of multimodal signal alignment in the time dimension, a unified clock reference is configured for all data sampling channels, and a consistent time tag is assigned to each set of collected data. This enables synchronous acquisition and time-domain registration of multi-source signals. Different types of signal sources have their own advantages in response mechanisms, propagation paths, signal characteristics, and anti-interference capabilities. UHF (ultra-high frequency) signals can capture spatial radiation energy variations with high sensitivity, ultrasonic signals have strong discharge source localization capabilities, and TEV (transient earth voltage) signals can reveal the conduction trajectory of discharge currents within metal grounding structures. This multi-source collaborative acquisition mechanism captures multi-dimensional, multi-angle response information during the same discharge event, building a highly complementary, time-consistent raw data input. This lays a solid data foundation for subsequent feature extraction, fusion, and artificial intelligence modeling, effectively enhancing the system's robustness in identifying partial discharge signals and the accuracy of discharge type discrimination under complex operating conditions.

[0026] S2: Perform pre-processing and feature fusion based on multi-source signals to obtain a fusion feature matrix for partial discharge identification.

[0027] Furthermore, the pre-processing includes denoising the electromagnetic wave signal using the wavelet packet decomposition method, extracting the time-frequency domain characteristic parameters of the electromagnetic wave signal, spatially locating the ultrasonic signal based on the direction of arrival algorithm, and extracting the pulse amplitude and change rate characteristics of the transient ground voltage signal.

[0028] Feature fusion includes aligning the time-frequency domain characteristic parameters of the electromagnetic wave signal and the pulse amplitude and change rate characteristics of the transient ground voltage signal to obtain a feature vector.

[0029] It should be noted that the collected electromagnetic wave signals were denoised using a wavelet packet decomposition method. The sym6 wavelet basis function, which has a good response to sudden changes, was selected as the decomposition basis. The decomposition layer was set to 6. A hierarchical frequency band expansion method was used to extract the high-frequency transient features contained in the partial discharge signal, retaining the effective high-frequency pulse components and suppressing background low-frequency interference, thereby improving the characteristic fidelity of the electromagnetic wave signal in complex interference environments. Furthermore, the time-frequency domain characteristic parameters of the electromagnetic wave signal, including pulse peak value, rise time, and spectral centroid, were extracted from the effective frequency band after wavelet decomposition. These parameters are used as input for subsequent discharge type pattern recognition. The collected ultrasonic signal is spatially localized using a direction of arrival algorithm. The pulse amplitude and rate of change characteristics of the transient ground voltage signal are extracted to represent the absolute intensity of the energy released by the partial discharge event and the steepness of the signal edge, respectively, thereby enhancing the perception of low-amplitude and rapidly changing discharge behavior.

[0030] Feature fusion involves dimensional alignment and feature matching of key feature vectors extracted from multi-source sensor signals after pre-processing. This involves pairing the time-frequency domain statistical characteristic parameters of the electromagnetic wave signal (such as peak value, rise time, and spectral centroid) with the pulse amplitude and rate of change characteristics of the transient ground voltage signal dimensionally to construct a fused feature vector group. An attention weighting mechanism is introduced to adaptively adjust the weight of each feature during the fusion process to account for information redundancy and feature importance differences between different modal signals. Ultimately, a fused feature vector with complete expression, unified dimensions, and high discriminative power for partial discharge (PD) types is formed. This fusion vector serves as input for the subsequent artificial intelligence hybrid model. The time-frequency characteristics of the electromagnetic wave signal reflect the energy release pattern of PD behavior in the high-frequency electromagnetic domain and are suitable for capturing corona and far-field discharge characteristics. The pulse amplitude and edge change rate of the transient ground voltage signal are more representative of the transient impact characteristics of the near-field discharge current. By fusing this multi-modal heterogeneous information and jointly modeling it at the feature level, the system achieves higher information integrity and robustness, effectively improving the accuracy and adaptability of PD type identification in complex interference environments.

[0031] The direction of arrival algorithm includes determining the delay time for the ultrasonic wave to reach different sensor elements based on the time difference between the sensor elements receiving the ultrasonic signal, calculating the phase difference of the received signal of each sensor element based on the delay time, constructing a spatial steering vector, and using the spatial spectrum estimation method to search for the spatial angle at which the steering vector and the measured data are most matched. The direction of arrival of the ultrasonic signal is determined. The delay time for the ultrasonic wave to reach different sensor elements is expressed as:

[0032] ;

[0033] in, Indicates the The delay time for each ultrasonic sensor element to receive the partial discharge acoustic emission signal is Indicates the sensor number in the sensor array, represents the center distance between adjacent sensor array elements, represents the incident angle of the acoustic signal relative to the normal direction of the sensor array, Indicates the propagation speed of ultrasonic waves in the medium.

[0034] The phase difference of the received signal of each sensor array element is calculated based on the delay time and expressed as:

[0035] ;

[0036] in, Indicates the The phase difference between the signal received by the first sensor element and the signal received by the first sensor element, Indicates the frequency of the ultrasonic signal generated by partial discharge.

[0037] The constructed spatial steering vector is represented as:

[0038] ;

[0039] in, represents the steering vector, Indicates the The phase delay term of each sensor element, Indicates the sensor array The phase response of each array element to the signal from the direction of the sound source.

[0040] The direction of arrival of the ultrasonic signal is determined as:

[0041] ;

[0042] in, represents the direction of arrival angle, Indicates the response intensity of the array to different incident directions.

[0043] The steering vector sum and spatial spectrum maximum search constitute the core of the direction of arrival estimation algorithm. By linking with the data of multi-channel ultrasonic sensors, it can achieve high-precision inversion of the spatial location of the local discharge source, providing spatial constraint input for subsequent fusion feature construction and artificial intelligence models, greatly improving the system's interpretability and discrimination performance for local discharge types and source points.

[0044] Furthermore, the fusion feature matrix of partial discharge identification includes calculating the correlation weight coefficients between feature vectors based on the attention mechanism, performing weighted fusion on the feature vectors, and generating a weighted fusion feature matrix.

[0045] It should also be noted that the association weight coefficient between feature vectors calculated based on the attention mechanism is expressed as:

[0046] Constructing electromagnetic wave eigenvectors , the modal features are uniformly transformed to form the input feature matrix , in the attention mechanism:

[0047] ;

[0048] in, The eigenvector matrix representing the UHF electromagnetic wave signal, The eigenvector matrix representing the ultrasonic signal, The eigenvector matrix representing the transient ground voltage signal, is the characteristic number, Feature dimension, 、 、 Represent query, key, and value vector representations respectively, 、 、 is the trainable projection matrix.

[0049] The calculation of the attention weight matrix is ​​expressed as:

[0050] ;

[0051] in, represents the attention weight matrix, Indicates normalizing each row in the matrix to convert it into a probability distribution, which is used to control the weight distribution of each feature dimension. represents the transpose of the key matrix, Indicates the dimension of the key vector.

[0052] The fusion feature matrix after generating weighted fusion is expressed as:

[0053] ;

[0054] in, Represents the fused feature matrix, which serves as the input of the subsequent artificial intelligence hybrid model. Represents the value matrix, that is, the eigenvector matrix before fusion.

[0055] S3: Based on the fusion feature matrix of partial discharge identification, partial discharge type identification and classification are performed through an artificial intelligence hybrid model.

[0056] Furthermore, the artificial intelligence hybrid model includes using a residual network to extract spatial features from the fused feature matrix to obtain a multi-scale spatial feature map, using a Transformer encoder to extract temporal features from the fused feature matrix to obtain a temporal feature representation of long-term dependencies, and performing feature cascade on the spatial feature map and the temporal feature representation to form a comprehensive feature vector that integrates spatial and temporal information.

[0057] It should be noted that the artificial intelligence hybrid model is expressed as:

[0058] ;

[0059] in, Represents the comprehensive feature vector after integrating spatial and temporal information, is the spatial feature map representation, is the time series feature representation, It means concatenating (cascading) according to the feature dimension to obtain the joint feature vector of spatial and temporal information.

[0060] A joint modeling architecture is constructed based on the fused feature matrix, and a residual neural network (Res Net) is used to extract multi-scale spatial feature representations from the fused feature matrix to obtain a spatial feature map that characterizes the distribution relationship of different modal partial discharge features in the spatial dimension. A Transformer encoder is used to globally model the fused feature matrix, and a multi-head attention mechanism is used to extract the dependencies of each feature vector in the time series to construct a long-term dependent temporal feature representation that represents the changing trend of partial discharge behavior. The spatial feature map and the temporal feature representation are concatenated through a skip connection structure to form a joint expression vector that contains both local spatial features and global temporal structure. The artificial intelligence hybrid model combines the deep convolutional expression capabilities of Res Net in spatial modeling with the long-distance dependency capture capability of Transformer in sequence modeling. It can effectively solve the modeling balance problem between local sensitivity and global discriminability of multimodal fusion features, and improve the stability and accuracy of partial discharge type identification under abnormal working conditions such as complex background noise and missing modal features.

[0061] Furthermore, the identification and classification of partial discharge types include calculating the probability distribution of each discharge type corresponding to the comprehensive feature vector through the Softmax classifier based on the comprehensive feature vector, sorting the probabilities of each discharge type using the probability threshold discrimination method, determining the partial discharge type and outputting the discharge type label, performing confidence analysis on the output discharge type label, and when the confidence is lower than the preset threshold, calling historical feature data to re-calibrate and adjust the artificial intelligence hybrid model parameters, and adaptively optimizing the artificial intelligence hybrid model.

[0062] It should be noted that the probability distribution of the calculated comprehensive eigenvector corresponding to each type of discharge is expressed as:

[0063] ;

[0064] in, Represents the output discharge type classification probability vector, Represents a multi-class normalization function that outputs the probability distribution of discharge types. express The weight matrix of the classifier, Represents the bias vector of the classifier.

[0065] The fused comprehensive feature vector is input into a Softmax classifier, which calculates the probability distribution for various partial discharge types (such as corona discharge, surface discharge, and internal discharge) and outputs a confidence probability value corresponding to each discharge type. The Softmax function maps the fused feature vector in a multi-class space, ensuring that the output value satisfies normalization conditions, forming a probability prediction result for all predefined discharge types. Based on the probability output, the maximum probability criterion is used to distinguish and rank the various partial discharge types, determine the discharge type most likely corresponding to the current sample, and then output the discharge type label. To improve model reliability and recognition accuracy, the system further performs confidence analysis on the output discharge type label. When the maximum probability value output by the classifier falls below the set confidence threshold, the system triggers an adaptive feedback mechanism, automatically calling the historical feature vector library to perform similarity matching and label re-estimation between the current sample and existing samples. The model's decision boundary is corrected through a resampling mechanism. The revalidated samples and historical samples are incorporated into the joint training set to fine-tune the parameters of the artificial intelligence hybrid model and update the network weight matrix, thus achieving adaptive optimization and dynamic evolution of the model structure in the actual operating environment. This mechanism significantly improves the robustness and adaptability of the model in complex electromagnetic environments, multi-source modal loss, and rare discharge type conditions, ensuring that the local discharge type classification results have long-term stability and high credibility.

[0066] Furthermore, the adaptive optimization of the artificial intelligence hybrid model includes real-time collection of fusion feature matrices and actual discharge type labels, updating of historical data sets, retraining of the updated historical data sets, and optimization of network weights and bias parameters in the artificial intelligence hybrid model using the gradient descent method. Based on the latest training results, the model weight parameters are dynamically updated, and the probability threshold of the Softmax classifier is adjusted in real time. Through continuous iterative updates, the parameters of the artificial intelligence hybrid model are adaptively optimized.

[0067] It should be noted that during operation, the fusion feature matrix and the actual output partial discharge type labels are collected in real time, a new sample data set is continuously constructed, and the updated samples are merged with the existing historical data to form a dynamic historical sample library. For the historical sample library, the system triggers incremental training of the artificial intelligence hybrid model and uses a gradient descent-based optimization algorithm to retrain the neural network structure in the model, including updating the weight parameters and bias parameters in the convolution layer and Transformer encoder layer, so that the model can continuously adapt to the changes in the discharge feature distribution under the latest working conditions. During the training process, the system dynamically monitors the training process based on the loss function and classification performance indicators (accuracy), judges the convergence state of the model in real time, and adjusts the network parameters such as learning rate and regularization factor accordingly to improve training efficiency and stability. At the same time, the Softmax classifier at the output end is updated online, and the probability threshold distribution of each category is re-estimated for fuzzy classification boundaries or new types of partial discharge appearing in new samples to ensure the reliability and discrimination of the model classification output. Through continuous iterative model fine-tuning and structural enhancement, the artificial intelligence hybrid model has the ability to self-evolve and adapt, and can cope with feature migration problems caused by factors such as equipment aging, noise interference, and environmental disturbances in long-term operation, thereby maintaining the high stability and high generalization performance of the partial discharge identification system, and effectively meeting the needs of continuous optimization of the performance of the online monitoring system in complex power grid operation environments.

[0068] Example 2 is an embodiment of the present invention, which provides an artificial intelligence-based valve-side casing partial discharge identification system, including a signal acquisition module, a feature processing module, and an intelligent identification module.

[0069] Among them: the signal acquisition module includes a multi-source acquisition unit and a time domain synchronization unit. The multi-source acquisition unit is used to synchronously acquire ultra-high frequency, ultrasonic, and transient ground voltage signals. The time domain synchronization unit is used to add a unified timestamp to the acquired signals to achieve time domain alignment of multi-source signals.

[0070] It should also be noted that the multi-source acquisition unit transmits the acquired multi-modal signals to the time domain synchronization unit, and the synchronization signal output by the time domain synchronization unit is transmitted to the pre-noise reduction unit.

[0071] The feature processing module includes a pre-denoising unit and a feature fusion unit. The pre-denoising unit is used to remove noise from the multimodal original signal and extract effective features. The feature fusion unit is used to weightedly fuse the multimodal features through the attention mechanism to form a fusion feature matrix.

[0072] It should also be noted that the denoising features output by the pre-denoising unit are transmitted to the feature fusion unit, and the fused feature matrix output by the feature fusion unit is transmitted to the hybrid modeling unit.

[0073] The intelligent identification module includes a hybrid modeling unit and an adaptive optimization unit. The hybrid modeling unit is used to extract and classify features of the fused feature matrix using an artificial intelligence hybrid model. The adaptive optimization unit is used to optimize model parameters in real time based on field feedback data.

[0074] It should also be noted that the classification results and feedback information output by the hybrid modeling unit are transmitted to the adaptive optimization unit.

Claims

1. A method for identifying partial discharge of valve-side casing based on artificial intelligence, characterized in that: include: Synchronous acquisition of multi-source signals; Based on multi-source signals, pre-processing and feature fusion are performed to obtain a fusion feature matrix for partial discharge identification; Based on the fusion feature matrix of partial discharge identification, partial discharge type identification and classification are performed through artificial intelligence hybrid model; Pre-processing includes denoising the electromagnetic wave signal using wavelet packet decomposition, extracting the time-frequency domain characteristic parameters of the electromagnetic wave signal, spatially locating the ultrasonic signal based on the direction of arrival algorithm, and extracting the pulse amplitude and change rate characteristics of the transient ground voltage signal; Feature fusion includes aligning the time-frequency domain characteristic parameters of the electromagnetic wave signal and the pulse amplitude and change rate characteristics of the transient ground voltage signal to obtain a feature vector; The fusion feature matrix for partial discharge identification includes adjusting the attention weight coefficient of the fusion feature matrix. The contribution of modal features to discharge type recognition is dynamically evaluated through the attention mechanism to obtain the weight-enhanced fusion feature matrix. The correlation weight coefficient between feature vectors is calculated based on the attention mechanism and is expressed as: Construct electromagnetic wave feature vector F UHF ∈R n×d ,F US ∈R n×d ,F TEV ∈R n×d , the modal features are uniformly transformed to form the input feature matrix X∈R n×d , in the attention mechanism: Q=XW Q ,K=XW K ,V=XW V Among them, F UHF The eigenvector matrix representing the UHF electromagnetic wave signal, F US The eigenvector matrix representing the ultrasonic signal, F TEV The eigenvector matrix represents the transient ground voltage signal, n is the number of features, d is the feature dimension, Q, K, V represent the query, key, and value vector representation respectively, W Q 、W K 、W V is a trainable projection matrix; The calculation of the attention weight matrix is ​​expressed as: Among them, A represents the attention weight matrix, softmax(.) represents normalizing each row in the matrix to convert it into a probability distribution, which is used to control the weight distribution of each feature dimension, and K T represents the transpose of the key matrix, d k represents the dimension of the key vector; The fusion feature matrix after generating weighted fusion is expressed as: F fused =OFF Among them, F fused represents the fused feature matrix, and V represents the value matrix; The artificial intelligence hybrid model includes using a one-dimensional convolutional neural network to extract local features from the fused feature matrix, extracting local spatial and short-term temporal information features from the fused feature matrix through a one-dimensional convolution kernel to obtain a local feature representation, extracting long-term dependent global temporal features from the local feature representation to form a global temporal feature representation, fusing the local feature representation and the global temporal feature representation through a jump connection to form a mixed feature vector that contains both short-term local and long-term global information, inputting the mixed feature vector into the classification layer, and outputting the local discharge type recognition and classification results through the Softmax function.

2. The artificial intelligence-based valve-side casing partial discharge identification method according to claim 1, characterized in that: The synchronous acquisition of multi-source signals includes acquiring electromagnetic wave signals through a UHF sensor, acquiring ultrasonic signals through an ultrasonic sensor, and acquiring transient voltage signals through a transient voltage sensor. During the sampling process, a consistent time tag is assigned to each set of data, and data is acquired synchronously.

3. The artificial intelligence-based valve-side casing partial discharge identification method according to claim 2, characterized in that: The fusion feature matrix of partial discharge identification includes calculating the correlation weight coefficients between feature vectors based on the attention mechanism, performing weighted fusion on the feature vectors, and generating a weighted fusion feature matrix.

4. The artificial intelligence-based valve-side casing partial discharge identification method according to claim 3, characterized in that: The artificial intelligence hybrid model includes using a residual network to extract spatial features from the fused feature matrix to obtain a multi-scale spatial feature map, using a Transformer encoder to extract temporal features from the fused feature matrix to obtain a temporal feature representation of long-term dependencies, and performing feature cascading on the spatial feature map and the temporal feature representation to form a comprehensive feature vector that integrates spatial and temporal information.

5. The artificial intelligence-based valve-side casing partial discharge identification method according to claim 4, characterized in that: The partial discharge type identification and classification includes calculating the probability distribution of each discharge type corresponding to the comprehensive feature vector through a Softmax classifier based on the comprehensive feature vector, sorting the probabilities of each discharge type using a probability threshold discrimination method, determining the partial discharge type and outputting a discharge type label, performing a confidence analysis on the output discharge type label, and when the confidence is lower than a preset threshold, calling historical feature data to re-verify and adjust the parameters of the artificial intelligence hybrid model, and adaptively optimizing the artificial intelligence hybrid model.

6. The artificial intelligence-based valve-side casing partial discharge identification method according to claim 5, characterized in that: The adaptive optimization of the artificial intelligence hybrid model includes real-time collection of fusion feature matrices and actual discharge type labels, updating historical data sets, retraining the updated historical data sets, and optimizing network weights and bias parameters in the artificial intelligence hybrid model using the gradient descent method. Based on the latest training results, the model weight parameters are dynamically updated, and the probability threshold of the Softmax classifier is adjusted in real time. Through continuous iterative updates, the parameters of the artificial intelligence hybrid model are adaptively optimized.

7. A system using the artificial intelligence-based valve-side casing partial discharge identification method according to any one of claims 1 to 6, characterized in that: Including signal acquisition module, feature processing module, and intelligent identification module; The signal acquisition module includes a multi-source acquisition unit and a time domain synchronization unit. The multi-source acquisition unit is used to synchronously acquire ultra-high frequency, ultrasonic, and transient ground voltage signals. The time domain synchronization unit is used to add a unified time stamp to the acquired signals to achieve time domain alignment of multi-source signals. The feature processing module includes a pre-noise reduction unit and a feature fusion unit. The pre-noise reduction unit is used to remove noise from the multimodal original signal and extract effective features. The feature fusion unit is used to weightedly fuse the multimodal features through an attention mechanism to form a fusion feature matrix. The intelligent identification module includes a hybrid modeling unit and an adaptive optimization unit. The hybrid modeling unit is used to extract and classify features of the fusion feature matrix using an artificial intelligence hybrid model, and the adaptive optimization unit is used to optimize model parameters in real time based on field feedback data.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the valve-side bushing partial discharge identification method based on artificial intelligence according to any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the valve-side bushing partial discharge identification method based on artificial intelligence according to any one of claims 1 to 6 are implemented.

Images