Partial discharge detection method and device, computer device, readable storage medium and program product

Abstract

This application relates to a partial discharge detection method, apparatus, computer equipment, computer-readable storage medium, and computer program product, relating to the field of power equipment insulation testing, and can improve the accuracy of partial discharge detection. The method includes: acquiring partial discharge monitoring signals of multiple modes from the device under test, and corresponding signal quality indices for each mode of partial discharge monitoring signal; inputting the partial discharge monitoring signals of various modes into a trained fault detection model; outputting the spatiotemporal features corresponding to the partial discharge monitoring signals of various modes from the fault detection model; determining the attention weights corresponding to each mode of partial discharge monitoring signal based on each signal quality index; fusing the spatiotemporal features corresponding to the partial discharge monitoring signals of various modes using each attention weight to obtain multimodal fusion features; and determining the partial discharge detection result based on the multimodal fusion features.

Description

Technical Field

[0001] This application relates to the field of electrical equipment insulation testing technology, and in particular to a partial discharge detection method, apparatus, computer equipment, computer-readable storage medium, and computer program product. Background Technology

[0002] Partial discharge detection is a key technology for assessing the insulation condition of power equipment. However, in actual substation or transmission line operating environments, there are complex electromagnetic interferences and environmental noises.

[0003] Traditional partial discharge detection techniques often struggle to effectively separate valid discharge signals from background noise when faced with strong interference, resulting in insufficient detection sensitivity and anti-interference capabilities, and a tendency to produce false alarms or missed alarms. Furthermore, with the increasing variety of power equipment, the types of insulation faults are becoming increasingly complex. Existing fault diagnosis methods have limitations in feature mining and data processing capabilities, making it difficult to achieve high-precision identification and classification of different fault types, and thus failing to meet the power system's need for precise monitoring of equipment status. Summary of the Invention

[0004] Therefore, it is necessary to provide a partial discharge detection method, apparatus, computer equipment, computer-readable storage medium, and computer program product to address the aforementioned technical problems.

[0005] In a first aspect, this application provides a partial discharge detection method, comprising:

[0006] Acquire partial discharge monitoring signals of various modes of the device under test, and the corresponding signal quality indicators of the partial discharge monitoring signals of various modes;

[0007] The partial discharge monitoring signals of the various modalities are input into the trained fault detection model; the fault detection model is obtained by training spatial feature extraction branch and temporal feature extraction branch based on partial discharge sample data of multiple modalities and the fault labels corresponding to the partial discharge sample data.

[0008] The fault detection model outputs the spatiotemporal characteristics corresponding to the partial discharge monitoring signals of various modes;

[0009] Based on each of the aforementioned signal quality indicators, the attention weights corresponding to the partial discharge monitoring signals of the various modes are determined.

[0010] The spatiotemporal features corresponding to the partial discharge monitoring signals of various modalities are fused using the attention weights of each modality to obtain multimodal fusion features. Based on the multimodal fusion features, the partial discharge detection result is determined.

[0011] In one embodiment, the signal quality metric includes the signal-to-noise ratio; determining the attention weights corresponding to each of the multiple modal partial discharge monitoring signals based on each of the signal quality metrics includes:

[0012] The signal-to-noise ratios corresponding to the partial discharge monitoring signals of various modes are mapped to normalized numerical weights using a preset mapping function; wherein, the numerical weights increase as the signal-to-noise ratio increases.

[0013] The numerical weights are determined as the attention weights of the partial discharge monitoring signals for the corresponding modes.

[0014] In one embodiment, the spatiotemporal features corresponding to the partial discharge monitoring signals of the various modes output by the fault detection model include:

[0015] Using the spatial feature extraction branch in the fault detection model, the input partial discharge monitoring signal is convolved, and the spatial distribution features of the partial discharge monitoring signal of various modes are obtained based on the convolution operation result.

[0016] Using the time-series feature extraction branch in the fault detection model, the input partial discharge monitoring signal is processed in a time-series manner, and the time-series processing results are used to obtain the time-dependent features of the partial discharge monitoring signal of various modes.

[0017] Based on the spatial distribution characteristics and the dependency characteristics, the spatiotemporal characteristics corresponding to the partial discharge monitoring signals of the various modes are determined.

[0018] In one embodiment, acquiring partial discharge monitoring signals of multiple modes of the device under test includes:

[0019] The original analog signal for the device under test is acquired by a preset sensor group; the sensor group includes at least two of the following: ultra-high frequency sensor, ultrasonic sensor, high frequency current sensor, and transient ground voltage sensor.

[0020] Based on the pre-acquired timing signal, the original analog signals of each mode are time-aligned to obtain the aligned multimodal signals;

[0021] The aligned multimodal signal is filtered and compressed to obtain the partial discharge monitoring signal.

[0022] In one embodiment, determining the partial discharge detection result based on multimodal fusion features includes:

[0023] The multimodal fusion features are input into the fully connected classification layer of the fault detection model to calculate the probability values ​​of the device under test belonging to the normal state and various preset fault types; the preset fault types include corona discharge, suspension discharge, surface discharge, internal discharge and arc discharge.

[0024] The fault type corresponding to the highest probability value is determined as the partial discharge detection result.

[0025] In one embodiment, after determining the partial discharge detection result based on multimodal fusion features, the method further includes:

[0026] Obtain the fault type corresponding to multiple monitoring time points, and obtain the discharge intensity corresponding to each monitoring time point;

[0027] By correlating the monitoring time points, the fault types, and the discharge intensity, a historical monitoring sequence is obtained;

[0028] A trend analysis is performed on the discharge intensity corresponding to the fault type at multiple detection time points in the historical monitoring sequence, and the insulation degradation trend of the device under test is determined based on the trend analysis results.

[0029] When the insulation degradation trend meets the preset alarm conditions, a trend warning message is generated.

[0030] Secondly, this application also provides a partial discharge detection device, comprising:

[0031] The partial discharge data acquisition module is used to acquire partial discharge monitoring signals of various modes of the device under test, and the signal quality indicators corresponding to the partial discharge monitoring signals of various modes.

[0032] The model input module is used to input the partial discharge monitoring signals of various modalities into the trained fault detection model; the fault detection model is obtained by training spatial feature extraction branch and temporal feature extraction branch based on partial discharge sample data of multiple modalities and the fault labels corresponding to the partial discharge sample data.

[0033] The model analysis module is used to output the spatiotemporal characteristics corresponding to the partial discharge monitoring signals of various modes from the fault detection model;

[0034] The attention weight determination module is used to determine the attention weight corresponding to each of the various modes of partial discharge monitoring signals based on each of the signal quality indicators.

[0035] The detection result output module is used to fuse the spatiotemporal features corresponding to the partial discharge monitoring signals of various modalities using the attention weights of each module to obtain multimodal fusion features, and to determine the partial discharge detection result based on the multimodal fusion features.

[0036] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0037] Acquire partial discharge monitoring signals of various modes of the device under test, and the corresponding signal quality indicators of the partial discharge monitoring signals of various modes;

[0038] The partial discharge monitoring signals of the various modalities are input into the trained fault detection model; the fault detection model is obtained by training spatial feature extraction branch and temporal feature extraction branch based on partial discharge sample data of multiple modalities and the fault labels corresponding to the partial discharge sample data.

[0039] The fault detection model outputs the spatiotemporal characteristics corresponding to the partial discharge monitoring signals of various modes;

[0040] Based on each of the aforementioned signal quality indicators, the attention weights corresponding to the partial discharge monitoring signals of the various modes are determined.

[0041] The spatiotemporal features corresponding to the partial discharge monitoring signals of various modalities are fused using the attention weights of each modality to obtain multimodal fusion features. Based on the multimodal fusion features, the partial discharge detection result is determined.

[0042] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the following steps:

[0043] Acquire partial discharge monitoring signals of various modes of the device under test, and the corresponding signal quality indicators of the partial discharge monitoring signals of various modes;

[0044] The partial discharge monitoring signals of the various modalities are input into the trained fault detection model; the fault detection model is obtained by training spatial feature extraction branch and temporal feature extraction branch based on partial discharge sample data of multiple modalities and the fault labels corresponding to the partial discharge sample data.

[0045] The fault detection model outputs the spatiotemporal characteristics corresponding to the partial discharge monitoring signals of various modes;

[0046] Based on each of the aforementioned signal quality indicators, the attention weights corresponding to the partial discharge monitoring signals of the various modes are determined.

[0047] The spatiotemporal features corresponding to the partial discharge monitoring signals of various modalities are fused using the attention weights of each modality to obtain multimodal fusion features. Based on the multimodal fusion features, the partial discharge detection result is determined.

[0048] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, performs the following steps:

[0049] Acquire partial discharge monitoring signals of various modes of the device under test, and the corresponding signal quality indicators of the partial discharge monitoring signals of various modes;

[0050] The partial discharge monitoring signals of the various modalities are input into the trained fault detection model; the fault detection model is obtained by training spatial feature extraction branch and temporal feature extraction branch based on partial discharge sample data of multiple modalities and the fault labels corresponding to the partial discharge sample data.

[0051] The fault detection model outputs the spatiotemporal characteristics corresponding to the partial discharge monitoring signals of various modes;

[0052] Based on each of the aforementioned signal quality indicators, the attention weights corresponding to the partial discharge monitoring signals of the various modes are determined.

[0053] The spatiotemporal features corresponding to the partial discharge monitoring signals of various modalities are fused using the attention weights of each modality to obtain multimodal fusion features. Based on the multimodal fusion features, the partial discharge detection result is determined.

[0054] The aforementioned partial discharge detection method, apparatus, computer equipment, computer-readable storage medium, and computer program product acquire partial discharge monitoring signals of multiple modalities of the device under test, and the corresponding signal quality indicators of each modal partial discharge monitoring signal; input the partial discharge monitoring signals of each modality into a trained fault detection model; the fault detection model is obtained by training spatial feature extraction branches and temporal feature extraction branches based on partial discharge sample data of multiple modalities and fault labels corresponding to the partial discharge sample data; the fault detection model outputs the spatiotemporal features corresponding to the partial discharge monitoring signals of each modality; based on each signal quality indicator, the attention weights corresponding to each partial discharge monitoring signal of each modality are determined; the spatiotemporal features corresponding to the partial discharge monitoring signals of each modality are fused using each attention weight to obtain multimodal fusion features; and based on the multimodal fusion features, the partial discharge detection result is determined. In this application, by acquiring partial discharge monitoring signals of multiple modes and their corresponding signal quality indices, the complementarity of different physical quantities in characterizing insulation defects is fully utilized. A pre-trained fault detection model is used to output the spatiotemporal features of each mode in parallel, ensuring the complete preservation of signal details and temporal evolution patterns in the frequency domain. More importantly, by introducing an attention weight mechanism based on signal quality indices, dynamic optimal fusion is achieved in the multi-modal feature fusion stage. That is, based on real-time signal quality, the feature weights of severely interfered (low quality indices) modes are automatically suppressed, while the feature contribution of high-quality modes is enhanced. This effectively solves the problem of decreased overall recognition rate due to noise in some sensors under complex electromagnetic environments, significantly improving the accuracy of partial discharge detection and diagnosis of power equipment. Attached Figure Description

[0055] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1 This is a flowchart illustrating a partial discharge detection method in one embodiment;

[0057] Figure 2 This is a flowchart illustrating a partial discharge detection method in another embodiment;

[0058] Figure 3 This is a structural block diagram of a partial discharge detection device in one embodiment;

[0059] Figure 4 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0061] The partial discharge detection method provided in this application is applicable to insulation condition monitoring and fault diagnosis scenarios for various high-voltage electrical equipment. The equipment under test includes, but is not limited to, oil-immersed transformers, gas-insulated switchgear (GIS), high-voltage cables, high-voltage switchgear, instrument transformers, and insulator strings. This application scenario widely covers key power nodes such as substations, wind farms, photovoltaic power stations, and rail transit traction power supply systems, aiming to prevent power accidents caused by insulation failure through real-time capture and analysis of partial discharge signals.

[0062] In one embodiment, such as Figure 1 As shown, a partial discharge detection method is provided. This embodiment illustrates the application of this method to a terminal. It is understood that this method can also be applied to a server, and further to a system including both a terminal and a server, and implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0063] Step S101: Acquire partial discharge monitoring signals of various modes of the device under test, and the corresponding signal quality indicators of the partial discharge monitoring signals of various modes.

[0064] Among them, the partial discharge monitoring signal can be a sequence of physical quantity data reflecting the non-penetrating discharge activity of the internal insulation defects of the equipment under the action of an electric field, which is used to characterize the current insulation status and potential fault characteristics of the equipment. The signal is usually collected by different types of sensors deployed on or around the equipment, and can be preprocessed by filtering, noise reduction, compression or time alignment to form a standardized digital signal suitable for model input.

[0065] Signal quality indicators can be numerical parameters used to quantitatively evaluate the effectiveness, purity, or interference level of a monitoring signal, aiming to provide confidence in different data sources for subsequent multimodal feature fusion. These indicators can be calculated based on the ratio of the effective pulse amplitude to the background noise level in the monitoring signal, or determined based on statistical characteristics such as waveform entropy, smoothness, or peak factor of the signal.

[0066] Specifically, the terminal maintains a connection with the sensor array via a wired or wireless communication interface, receiving raw data from multiple different modes of the device under test in parallel. The terminal can be configured to perform preliminary cleaning and formatting of this raw data to acquire partial discharge monitoring signals for multiple modes. Simultaneously, the terminal can perform real-time quality assessment of each monitoring signal, calculate the corresponding signal quality index, and associate and store or synchronously transmit this index with the corresponding monitoring signal, so that subsequent processing steps can perceive the reliability of each data stream.

[0067] Step S102: Input the partial discharge monitoring signals of various modes into the trained fault detection model; the fault detection model is obtained by training the spatial feature extraction branch and the temporal feature extraction branch based on partial discharge sample data of multiple modes and the fault labels corresponding to the partial discharge sample data.

[0068] The fault detection model can be an algorithmic architecture based on deep learning, used to map the input raw monitoring signal into a high-dimensional hidden layer feature representation. This model is obtained through supervised training on a large amount of historically accumulated partial discharge sample data of various modalities and their corresponding fault labels. In order to simultaneously understand the distribution pattern of partial discharge phenomena in physical space and its evolution trend in the time dimension, the model is configured in network structure design to include at least two parallel processing paths: a spatial feature extraction branch for capturing data morphological structure information, and a temporal feature extraction branch for capturing the dynamic change pattern of data.

[0069] Specifically, the terminal invokes a pre-stored or cloud-loaded trained fault detection model, transmitting the acquired partial discharge monitoring signals of various modes as input data to the model's input layer. The terminal utilizes the model's inference engine to guide the data flow into the model's internal structure, causing it to enter the spatial feature extraction branch and the temporal feature extraction branch for parallel or serial computation processing. During this process, the model's internal weight parameters (i.e., the parameters determined during training) can perform mathematical operations with the input signals to extract deeper physical features of the signals.

[0070] Optionally, during the training phase, the parameters within the spatial and temporal branches are continuously adjusted using backpropagation or gradient descent optimization strategies until the error between the model's predicted result and the fault label corresponding to the sample converges to a preset range. For example, the spatial feature extraction branch can be configured to focus on the clustering features of the signal in the phase-amplitude plane (such as Phase Resolved Partial Discharge (PRPD) spectral features), while the temporal feature extraction branch can be configured to focus on the fluctuations or repetition frequency characteristics of the signal over time.

[0071] Step S103: The fault detection model outputs the spatiotemporal characteristics corresponding to the partial discharge monitoring signals of various modes.

[0072] Among them, spatiotemporal features can be a high-dimensional set of numerical values ​​or feature vectors obtained by deep encoding the original signal through a neural network, and are the mapping expression of the original physical signal in the abstract feature space.

[0073] Specifically, after inputting partial discharge monitoring signals of multiple modes into the fault detection model, the terminal can trigger forward propagation calculations within the model. The terminal obtains the computation results for each mode signal, i.e., spatiotemporal features, from the model's output layer or a specific intermediate layer. For each input mode signal, the fault detection model can output a set of spatial distribution vectors in parallel through its internal spatial feature extraction branch, and output a set of temporal dependency vectors through its temporal feature extraction branch.

[0074] Optionally, the spatiotemporal features acquired by the terminal can be a combination of the aforementioned spatial distribution vector and temporal dependency vector. In some implementations, the model can directly output the concatenated long vector as the spatiotemporal features of that modality; in other implementations, the model can also output two independent feature matrices, which are temporarily stored by the terminal outside the model. It should be noted that, for different input modalities (such as ultrasound and ground voltage), although the spatiotemporal features output by the model have different physical meanings, they can be configured to have the same dimension in terms of data structure to facilitate subsequent standardization processing.

[0075] Step S104: Based on each signal quality index, determine the attention weights corresponding to the partial discharge monitoring signals of various modes.

[0076] Among them, attention weights can be numerical coefficients used to characterize the importance or contribution ratio of different modal data in the final decision. Mathematically, they can be represented as a set of normalized scalar values, which are used to scale or weight the corresponding feature vectors during the feature fusion stage.

[0077] Specifically, after acquiring the signal quality indices corresponding to the partial discharge monitoring signals of various modes, the terminal can initiate the weight calculation logic. The terminal is configured to establish a mapping relationship, using each signal quality index as an input variable, and calculates the corresponding numerical result as the attention weight for that mode through a specific algorithm or lookup table method. During this process, the terminal can comprehensively consider the quality status of all input modes, ensuring that the generated weight value not only reflects the absolute quality of a single mode but also its relative advantage within the overall multi-modal group.

[0078] Step S105: The spatiotemporal features corresponding to the partial discharge monitoring signals of various modes are fused using the attention weights to obtain multimodal fusion features. Based on the multimodal fusion features, the partial discharge detection results are determined.

[0079] Among them, multimodal fusion features can be a unified feature vector that can characterize the overall insulation state of the device under test, used to eliminate the one-sidedness of single-modal features and highlight the key fault information carried by high-confidence signals.

[0080] Partial discharge detection results can be the final judgment information reflecting the current insulation health level or specific fault attributes of the equipment under test, which can be used to guide maintenance personnel to perform equipment maintenance or trigger automated protection logic. The result can be a binary normal / abnormal status identifier, or a detailed diagnostic report containing specific fault types (such as corona, floating discharge, etc.) and their probability of occurrence.

[0081] Specifically, the terminal is configured to adaptively adjust the amplitude or assign importance to the corresponding spatiotemporal features using attention weights. Through this process, the terminal can enhance the expressive power of high-quality, high-weight modal features in the final fused features, while suppressing interference from low-quality modal features. Subsequently, the terminal integrates the adjusted modal features into a multimodal fused feature, and based on this fused feature, uses pre-defined decision logic (such as classifiers, threshold discrimination, or cluster analysis) to perform inference, ultimately determining and outputting the partial discharge detection result.

[0082] Optionally, when determining the detection result, the terminal can input the multimodal fusion features into a mapping layer or a classification network. This mapping layer is configured to find the correspondence between the feature space and the fault category space. For example, the terminal can calculate the probability distribution of the fusion features belonging to multiple preset fault types, thereby determining the partial discharge detection result.

[0083] In this embodiment, by acquiring partial discharge monitoring signals of multiple modes and their corresponding signal quality indices, the complementarity of different physical quantities in characterizing insulation defects is fully utilized. A pre-trained fault detection model is used to output the spatiotemporal features of each mode in parallel, ensuring the complete preservation of signal details and temporal evolution patterns in the frequency domain. More importantly, by introducing an attention weight mechanism based on signal quality indices, dynamic optimal fusion is achieved in the multi-modal feature fusion stage. That is, based on real-time signal quality, the feature weights of severely interfered (low quality indices) modes are automatically suppressed, while the feature contribution of high-quality modes is enhanced. This effectively solves the problem of decreased overall recognition rate due to noise in some sensors under complex electromagnetic environments, significantly improving the accuracy of partial discharge detection and diagnosis of power equipment.

[0084] In one embodiment, the signal quality metrics include the signal-to-noise ratio; based on each signal quality metric, the attention weights corresponding to the various modalities of partial discharge monitoring signals are determined, including:

[0085] The signal-to-noise ratios (SNRs) of partial discharge monitoring signals of various modes are mapped to normalized numerical weights using a preset mapping function. The numerical weights increase with the increase of the SNR. The numerical weights are then determined as the attention weights of the partial discharge monitoring signals of the corresponding modes.

[0086] The signal-to-noise ratio (SNR) can be used to quantitatively evaluate the intensity of the effective partial discharge pulse component in the monitoring signal relative to the background noise intensity. It can be calculated based on the energy spectrum distribution of the signal during the preprocessing stage.

[0087] The mapping function can be a mathematical logic used to convert the signal-to-noise ratio index of the physical layer into the weight coefficient of the algorithm layer. It not only defines the monotonic correspondence between the two, but also can amplify the weight of high-quality signals and suppress the weight of low-quality signals through specific normalization or nonlinear transformation methods.

[0088] Specifically, the terminal first calculates the signal-to-noise ratio (SNR) of each modality (e.g., UHF, ultrasound, high-frequency current, and transient ground voltage). Then, the terminal processes these SNR values ​​using a preset mapping function. During this process, modal signals with higher SNR values ​​are assigned larger weights, while those with lower SNR values ​​are assigned smaller weights. Finally, the terminal locks the calculated normalized weights as the attention weights for the corresponding modalities in subsequent feature fusion steps.

[0089] Optionally, in order to accurately quantify this mapping relationship and prevent the influence of extreme values, the terminal can adopt a specific calculation strategy of first normalizing and then exponentially weighting. The first step can be SNR normalization, as shown in formula (1):

[0090] (1)

[0091] In some embodiments, The upper limit threshold for signal-to-noise ratio can be selected as: , The lower limit threshold of the signal-to-noise ratio can be selected as: The specific value can be determined by actual measurement and statistics at the substation.

[0092] After normalization, the attention weights can be determined using the normalized signal-to-noise ratio, as shown in the formula.

[0093] (2)

[0094] in, For the first Modal weights, satisfying .

[0095] In this embodiment, by establishing a signal-to-noise ratio-oriented weight mapping mechanism, the attention allocation of each mode can be dynamically adjusted to ensure the fusion of features dominated by high-confidence modes, effectively suppressing single-mode interference in complex electromagnetic environments and improving the robustness of partial discharge detection.

[0096] In one embodiment, the spatiotemporal characteristics corresponding to various modes of partial discharge monitoring signals output by the fault detection model include:

[0097] The spatial feature extraction branch of the fault detection model is used to perform convolution operations on the input partial discharge monitoring signals, and the spatial distribution features of the partial discharge monitoring signals of various modes are obtained based on the convolution operation results. The temporal feature extraction branch of the fault detection model is used to perform temporal processing on the input partial discharge monitoring signals, and the time-dependent features of the partial discharge monitoring signals of various modes are obtained based on the time-dependent processing results. Based on the above spatial distribution features and dependent features, the spatiotemporal features corresponding to the partial discharge monitoring signals of various modes are determined.

[0098] Convolution operation can be a mathematical processing procedure that performs weighted summation on the input data matrix based on a sliding window mechanism. It is mainly used to capture spatial structure information such as texture, edge or phase distribution of local data, thereby extracting spatial distribution features that reflect the static physical form of the signal.

[0099] Time series processing is a computational process that performs memory and forgetting operations on the time dimension of a data sequence. Its purpose is to uncover the causal relationships or dynamic evolution patterns between moments before and after a signal, and then generate dependent features (or time series features) that reflect the fluctuations of the discharge process over time.

[0100] Specifically, the terminal sends the pre-processed partial discharge monitoring signal (e.g., data represented in matrix or tensor form) to two branches of the fault detection model simultaneously or sequentially.

[0101] In the spatial feature extraction branch, the terminal is configured to perform convolutional operations using a convolutional neural network (CNN) structure. This branch performs a sliding scan on the signal data using a convolutional kernel of a preset size, calculates the feature response of local regions, and may use pooling layers for dimensionality reduction, ultimately outputting a set of spatial distribution feature vectors that can characterize, for example, the shape of a PRPD map cluster or the contour of a waveform.

[0102] In the temporal feature extraction branch, the terminal is configured to perform temporal processing using a recurrent neural network (RNN) and its variants (such as a long short-term memory network (LSTM) or a gated recurrent unit (GRU)). This branch treats the signal as a time series, controls the transmission of information and state updates through an internal gating mechanism, captures the long-term dependencies of discharge pulses over time, and ultimately outputs a set of dependency feature vectors that characterize the discharge growth trend or pulse repetition pattern. Finally, based on the above two sets of features, the terminal determines the spatiotemporal features corresponding to the modal signal by splicing or stacking them.

[0103] Optionally, in order to balance the feature extraction capability and computational complexity of the model, the terminal can adopt the following specific network structure configuration.

[0104] For the spatial feature extraction branch, the terminal can construct a structure containing 3 convolutional layers and 1 max pooling layer. The size of the convolutional kernel can be configured as 3×3, with a stride of 1, and edge padding operations are used to maintain the feature map size; the kernel size of the pooling layer can be set to 2×2.

[0105] For the temporal feature extraction branch, the terminal can construct a structure containing a two-layer LSTM network. The number of hidden units in each LSTM layer can be set to 64, and the dropout rate can be set to 0.2 to prevent overfitting. This structure can effectively handle sequence data with long time spans and accurately capture the gradual process of insulation degradation.

[0106] Optionally, to ensure the accuracy of feature extraction, the fault detection model can be trained for both spatial and temporal feature extraction branches. In one embodiment, the model can be trained based on fault data collected from 100 different types of power equipment (such as transformers, GIS, etc.), with a dataset size of up to 2000 faults per type, divided into training and test sets in an 8:2 ratio. The Adam optimizer can be used during training, with a learning rate of... It can be set to 0.001 to achieve rapid convergence of parameters. The loss function used in the training process can be the cross-entropy function, as shown in formula (3). After training, the model achieves a fault identification accuracy of 98.5% and a false alarm rate of 0.8% on the test set, which is 9.3 percentage points higher than the traditional single-modal model (accuracy of 89.2%).

[0107] (3)

[0108] Where M is the number of samples, For real labels (one-hot encoded). To predict probabilities.

[0109] In this embodiment, a parallel architecture with two branches for spatial and temporal feature extraction is constructed to achieve spatiotemporal global feature deconstruction of partial discharge signals. The spatial branch extracts morphological features of the discharge spectrum to accurately identify discharge patterns, while the temporal branch extracts dynamic evolution features to compensate for the lack of static analysis. The fusion of features from both branches significantly improves the model's ability to identify signals with similar or ambiguous morphologies, enhancing the accuracy of partial discharge detection in complex environments.

[0110] In one embodiment, acquiring partial discharge monitoring signals of multiple modes of the device under test includes:

[0111] The original analog signals for the device under test are acquired by a pre-set sensor group; the sensor group includes at least two of the following: ultra-high frequency sensor, ultrasonic sensor, high frequency current sensor and transient ground voltage sensor; the original analog signals of each mode are time-aligned based on the pre-acquired timing signal to obtain the aligned multi-mode signal; the aligned multi-mode signal is filtered and compressed to obtain the partial discharge monitoring signal.

[0112] The sensor group can be a hardware set deployed on or around the device under test, used to convert invisible physical field fluctuations inside the device into electrical signals that can be recognized by electronic devices. It is usually composed of multiple sensors covering different frequency bands or physical dimensions, and includes at least two of the following: Ultra High Frequency (UHF) sensors, Acoustic Emission (AE) sensors, High Frequency Current Transformer (HFCT) sensors, and Transient Earth Voltage (TEV) sensors.

[0113] Specifically, after receiving multiple raw signals, the terminal can parse the timing signals contained within or received synchronously, and use these timing signals as a reference to calibrate the start time of each modal data. By interpolating or aligning timestamps, it generates aligned multimodal signals that are strictly synchronized on the time axis. In some embodiments, GPS timing signals can be used, with an accuracy controlled to ≤10ns, to ensure the timing consistency of multimodal data and avoid feature misalignment.

[0114] Subsequently, the terminal uses an adaptive filtering algorithm to filter out power frequency harmonics and fixed frequency interference, and uses a transform domain algorithm to compress and encode the data. In some embodiments, the compression process can use a wavelet transform compression algorithm (db4 wavelet basis, 5-level decomposition) to compress the original data volume to 1 / 5, reducing the pressure of subsequent transmission and storage, while retaining the key features of partial discharge pulses. Next, the signal is subjected to DC component removal, pulse peak detection, and feature standardization. Finally, differentiated features are extracted for the signal characteristics of different modes to construct a partial discharge monitoring signal containing multi-dimensional information (i.e., the original feature matrix).

[0115] Specifically, a segmented programmable filter (low end 10-80kHz, high end 100-400kHz) is used, combined with the Least Mean Squares (LMS) algorithm to suppress power frequency harmonics and fixed frequency interference. The filter coefficient iteration is shown in formula (4):

[0116] (4)

[0117] in, Let n be the filter coefficient vector. =0.005 (step size factor), error signal , The reference signal is free from interference. After filtering, the power frequency interference amplitude of the output signal is reduced by 92%, and the output signal is shown in formula (5):

[0118] (5)

[0119] Specifically, at the signal optimization and feature construction level, the terminal can further perform the following processing steps to generate the final feature matrix:

[0120] DC component removal is performed by subtracting the mean from the DC components of each modal signal, as shown in formula (6):

[0121] (6)

[0122] Where x(n) is the original signal, N is the number of sampling points, and x′(n) is the signal after DC removal.

[0123] Peak pulse detection uses the 3σ criterion (σ being the signal standard deviation) to identify the timing of partial discharge pulses, extracting time-domain features such as pulse amplitude, rise time, and duration, and eliminating interference pulses with amplitudes less than 3σ. The extracted features are normalized to the [0,1] interval to eliminate dimensional influence, and the feature standardization is shown in formula (7).

[0124] (7)

[0125] in, These are the original eigenvalues. , These are the maximum and minimum values ​​of the characteristic, respectively.

[0126] Optionally, the sensor group connected to the terminal may include the following specific configuration parameters: an ultra-high frequency (UHF) sensor deployed at the opening of the equipment casing (such as a transformer tank or GIS busbar), with a bandwidth covering 300MHz-3GHz, gain ≥3.6dBi, and sensitivity ≤-70dBm, to capture highly interference-resistant electromagnetic pulses; an ultrasonic (AE) sensor attached to the metal casing of the equipment (such as the side wall of a transformer tank), with a frequency band covering 20-200kHz, sensitivity ≤50μV / Pa, and resolution 1μV, used to collect vibration signals to assist in spatial positioning; a high frequency current (HFCT) sensor sleeved on the equipment grounding wire, with an aperture of 50-100mm, a bandwidth of 10kHz-100MHz, and a measurement range of 10nA-100mA, used to capture ground current pulses; and a transient ground voltage (TEV) sensor adsorbed on the metal casing (such as a switch cabinet door), with a measurement range of 0-600V, resolution 1V, and response time ≤1μs, used to monitor changes in surface transient voltage.

[0127] Optionally, the terminal extracts differentiated features for different modes. For UHF signals, phase distribution features (skewness and kurtosis of the PRPD spectrum) and pulse amplitude distribution (maximum amplitude, mean, and standard deviation) are extracted; for AE signals, time-domain features (pulse rise time, duration, and peak factor) and frequency-domain features (center frequency, bandwidth, and spectral entropy) are extracted; for HFCT / TEV signals, statistical features (pulse count, mean / maximum discharge amount) and trend features (slope of discharge amount change within 1 hour) are extracted. Finally, the terminal constructs an original feature matrix with dimensions of 4 (modes) × 12 (number of features) × N (number of sampling points) as the partial discharge monitoring signal.

[0128] Optionally, in a low-cost scenario, the implementation can be simplified to dual-mode acquisition using both ultra-high frequency and ultrasonic sensors, eliminating the high-frequency current and transient ground voltage sensors. This reduces hardware costs by 40% and lowers the fault identification accuracy to 96% (still higher than existing technologies), making it suitable for low-cost scenarios such as distribution network transformer substations. In a high-precision scenario, an infrared thermal imaging mode (temperature range -20~150℃, resolution 640×512) can be added to acquire the local temperature rise signal accompanying partial discharge, improving the detection rate of internal latent faults by 8%, and making it suitable for enclosed equipment such as GIS busbars and cable joints.

[0129] Optionally, the edge preprocessing unit can use adaptive filtering algorithms to filter out power frequency harmonics and fixed frequency interference, use transform domain algorithms to compress and encode the data, and perform DC component removal, pulse peak detection, and feature standardization on the signal. In some embodiments, the connected edge preprocessing unit can integrate an FPGA chip (such as a Xilinx Zynq-7000 model) and an embedded processor (such as an ARM Cortex-A9).

[0130] In this embodiment, the detection blind zone is eliminated by integrating multimodal sensors and utilizing the complementary effects of acoustic and electrical properties; the timing alignment is used to ensure the consistency of multi-source data time sequence, laying the foundation for spatiotemporal feature correlation analysis; signal enhancement and data dimensionality reduction are achieved through filtering and compression, improving the ability to identify weak faults, thereby reducing the system load while meeting the real-time requirements of high-frequency continuous monitoring in industrial sites.

[0131] In one embodiment, the partial discharge detection result is determined based on multimodal fusion features, including:

[0132] The multimodal fusion features are input into the fully connected classification layer of the fault detection model to calculate the probability values ​​of the device under test belonging to the normal state and various preset fault types. The preset fault types include corona discharge, suspension discharge, surface discharge, internal discharge and arc discharge. The fault type corresponding to the highest probability value is determined as the partial discharge detection result.

[0133] The probability value is the numerical result output by the classification layer, which represents the likelihood that the device under test is currently in a certain state (such as normal or a certain type of fault).

[0134] The preset fault types are a set of classification labels defined by the system based on the common insulation defect mechanisms of power equipment. They cover a variety of typical physical failure modes, from air gap discharge to solid insulation breakdown, including but not limited to corona discharge, floating discharge, surface discharge, internal discharge and arc discharge.

[0135] Specifically, the terminal uses the multimodal fusion features as input data and sends them to the fully connected classification layer of the fault detection model. The terminal uses pre-trained weight parameters within this classification layer and the input multimodal feature vector to perform inference calculations, thereby calculating the raw log-probability corresponding to each preset state. Subsequently, the terminal normalizes these log-probability values ​​using an activation function, generating a probability distribution vector. Each element in this vector corresponds to the probability value of "normal state" and a specific fault type among "corona discharge, suspension discharge, surface discharge, internal discharge, and arc discharge." Finally, the terminal iterates through this probability distribution vector, compares the magnitudes of each probability value, and determines the state label (e.g., internal discharge) corresponding to the highest probability value as the final partial discharge detection result.

[0136] Optionally, in order to achieve accurate probability calculation, the terminal can use the Softmax function as the activation function in the fully connected layer.

[0137] Optionally, the input multimodal fusion features can be determined based on the weighted summation formula (8).

[0138] (8)

[0139] in, For the first Modal eigenvectors This is a multimodal fusion feature.

[0140] In this embodiment, a fully connected classification layer is used to analyze the multimodal fusion features, thereby achieving refined qualitative and quantitative diagnosis of insulation status and improving the objectivity and accuracy of the diagnostic results.

[0141] In one embodiment, after determining the partial discharge detection result based on multimodal fusion features, the method further includes:

[0142] The system acquires the fault types corresponding to multiple monitoring time points and the discharge intensity corresponding to each monitoring time point; it then correlates the monitoring time points, fault types, and discharge intensities to obtain a historical monitoring sequence; it performs trend analysis on the discharge intensities corresponding to the fault types at multiple detection time points in the historical monitoring sequence, and determines the insulation degradation trend of the equipment under test based on the trend analysis results; and it generates trend warning information when the insulation degradation trend meets the preset alarm conditions.

[0143] Among them, the historical monitoring sequence can be a structured collection of multi-dimensional status data of the device under test arranged in chronological order. It not only records the discharge intensity (such as discharge quantity value) of the device at various times in the past, but also associates the fault type judgment result and monitoring timestamp at that time.

[0144] Trend warning information can be an alarm signal automatically generated by the system when the predicted degree of degradation reaches the safety threshold, which is used to prompt maintenance personnel to take intervention measures before the failure actually occurs.

[0145] Specifically, after determining the partial discharge detection result at the current moment, the terminal can retrieve the historical records of the device under test from its internal memory or cloud database at multiple past monitoring points. The terminal obtains the fault type (such as a confirmed internal discharge) and discharge intensity corresponding to these time points, and maps them one-to-one with the monitoring time points to construct a historical monitoring sequence arranged in chronological order. Subsequently, the terminal uses statistical algorithms or time series deep learning models to extrapolate and predict the above historical monitoring sequence, uncovering the nonlinear law of insulation performance degradation over time. This allows it to determine the insulation degradation trend reflecting the future health status of the equipment, and calculate the degradation curve of the insulation performance of the device under test over time or the predicted value for future moments.

[0146] When the terminal determines that the insulation degradation trend meets the preset alarm conditions (e.g., the predicted value exceeds the safety threshold), it can generate a trend warning message and push it to relevant personnel through the display interface or network message. In one embodiment, the preset alarm conditions can be configured as a three-level warning mechanism: if the predicted discharge exceeds the equipment threshold (e.g., the transformer threshold of 500pC) after 3 months, a yellow warning is generated (reminding for periodic retesting); if the predicted discharge exceeds the threshold after 2 months, an orange warning is generated (reminding for special inspection); if the predicted discharge exceeds the threshold after 1 month, a red warning is generated (reminding for emergency shutdown and maintenance).

[0147] Optionally, to improve the accuracy of trend prediction, the historical monitoring sequence constructed by the terminal can also include environmental data. For example, the terminal selects "monthly average discharge amount, fault type probability, and ambient temperature and humidity (e.g., temperature 0-40℃, humidity 20%-80%)" as joint input features, and sets the time window to 6 months (including 12 sampling points).

[0148] Optionally, during trend analysis, the terminal can invoke a time-series prediction model based on LSTM (Long Short-Term Memory). This model is configured to include one input layer, two LSTM layers (with 128 hidden units), and one fully connected output layer. The terminal uses this model to predict future (e.g., the next 6 months) discharge values. During model training, the terminal can use Mean Squared Error (MSE) as the loss function, as shown in formula (9), until the loss difference is reached. 10 -6 Training is stopped when the final prediction error is ≤5%.

[0149] (9)

[0150] in This represents the actual historical discharge volume. The value is the LSTM prediction, and T=12 (number of historical samples).

[0151] In this embodiment, by performing correlation modeling on long-term historical data, the system can accurately identify the evolution trend of insulation defects and capture potential high-speed degradation hazards. Combined with a graded early warning mechanism, it can reserve sufficient maintenance window period for operation and maintenance decisions, effectively avoid unplanned power outages, and significantly reduce equipment operation and maintenance costs and grid operation risks.

[0152] In one exemplary embodiment, this application provides a more detailed method for partial discharge detection, such as... Figure 2 As shown, the method specifically includes:

[0153] Step S201: Coordinated acquisition and refined preprocessing of partial discharge multi-mode signals.

[0154] Specifically, the terminal connects to a sensor array deployed around the device under test (such as a transformer or GIS) via a high-speed communication interface. To achieve full-band coverage, the sensor array is configured to include: an ultra-high frequency sensor with a bandwidth covering 300MHz to 3GHz (for capturing electromagnetic pulses), an ultrasonic sensor with a frequency band covering 20kHz to 200kHz (for locating sound sources), a high-frequency current sensor with a bandwidth covering 10kHz to 100MHz (for detecting grounding pulses), and a transient ground voltage sensor with a measurement range of 0 to 600V. While receiving these four raw analog signals, the terminal analyzes the synchronously acquired GPS timing signal and marks and aligns each signal with nanosecond-level time accuracy, generating a strictly time-synchronized alignment signal. Subsequently, the terminal uses the least mean square adaptive filtering algorithm to purify the signal. This algorithm filters out power frequency interference by calculating the error signal in real time and iteratively updating the filtering coefficients with a small step size factor. Next, the wavelet transform algorithm is used to decompose and reconstruct the signal in multiple layers, retaining key coefficients and compressing the data volume to about one-fifth of the original size. Finally, the terminal performs DC removal operation and pulse extraction operation based on statistical standard deviation, and constructs an original feature matrix containing multi-dimensional information such as phase, amplitude, and frequency as a partial discharge monitoring signal.

[0155] Step S202: Adaptive calculation of attention weights based on signal quality.

[0156] Specifically, after acquiring the aforementioned multimodal monitoring signals, the terminal immediately calculates the signal-to-noise ratio (SNR) of each signal. To convert these physical indicators into algorithm weights, the terminal maps the SNR of each mode to a unified dimensionless range, for example, by linearly normalizing it based on preset upper and lower limits (e.g., 35 dB) and a lower limit (e.g., 5 dB). Subsequently, the terminal uses an exponential function to non-linearly amplify the normalized value and determines the proportion of this value in the sum of all modal exponential values ​​as the attention weight of that mode. Through this calculation, the terminal can dynamically adjust the influence of each mode. If the ultrasonic signal is subjected to strong vibration interference at a certain moment, causing a sudden drop in SNR, its weight will be automatically reduced to near zero by the algorithm, while the weight of clear ultra-high frequency signals will be significantly increased.

[0157] Step S203: Parallel extraction and weighted fusion of the spatiotemporal features of the partial discharge signal.

[0158] Specifically, the terminal inputs the preprocessed monitoring signal into a trained fault detection model. During the training phase, the fault detection model is optimized using an efficient optimizer (e.g., Adam, with a learning rate of 0.001) and a cross-entropy loss function. Internally, the data stream is distributed along two paths: one path uses a CNN branch containing multiple convolutional layers (e.g., 3 layers) and pooling layers to extract spatial distribution features reflecting the texture of the discharge spectrum using convolutional kernels; the other path uses a temporal branch containing multiple LSTM units (e.g., 2 layers, each containing 64 hidden units) to extract dependency features reflecting the fluctuation patterns of the pulse sequence using a gating mechanism. Subsequently, the terminal uses the attention weights calculated in the previous step to perform a weighted summation of the spatiotemporal features output by each modality. This involves multiplying the feature vector of each modality by its corresponding weight coefficient and then summing the results to generate high-dimensional multimodal fusion features.

[0159] Step S204: Classification and diagnosis of partial discharge fault types.

[0160] Specifically, the terminal inputs the generated multimodal fusion features into the fully connected classification layer of the model, calculating the numerical values ​​of the input features belonging to various fault types. Subsequently, the terminal uses the Softmax function to transform these values ​​into a probability vector that sums to 1. The terminal compares these probability values ​​and determines the fault type (such as internal discharge) corresponding to the highest probability (e.g., exceeding 98%) as the current detection result.

[0161] Step S205: Prediction and graded early warning of insulation degradation trend.

[0162] Specifically, after completing a single diagnosis, the terminal stores the fault type and discharge intensity in the database and constructs a monitoring sequence by combining it with historical data from a past period (e.g., 6 months). The terminal calls a dedicated LSTM trend prediction model (e.g., containing 2 layers of LSTM and 128 hidden units) to perform time-series analysis on the sequence, using mean squared error as the loss target, to predict the discharge volume for the next few months (e.g., 1 to 3 months). If the prediction shows that the discharge volume will exceed the equipment safety threshold in the next 3 months, the terminal generates a yellow warning; if it predicts exceeding the threshold within 2 months, an orange warning is generated; and if it predicts exceeding the threshold within 1 month, a red warning is generated.

[0163] To verify the technical effectiveness of the partial discharge detection method in the above embodiments in actual engineering, this application also provides an engineering embodiment as a verification illustration of the aforementioned general method in a specific high-voltage equipment scenario, demonstrating the deployment details and application effectiveness of this application in actual engineering.

[0164] Specifically, the aforementioned partial discharge detection method can be applied to specific engineering applications and measured results on a typical high-voltage power equipment such as a 110kV oil-immersed transformer. In this scenario, the transformer, as a power grid hub, has a complex internal insulation structure (including oil, paper, iron core, etc.), which places extremely high demands on the spatial layout of multi-modal sensors and signal acquisition strategies.

[0165] Specifically, to achieve full coverage monitoring of the transformer's insulation status, the terminal is connected to a customized sensor array via an edge computing unit: two ultra-high frequency sensors are configured and installed at the openings on both sides of the transformer tank, with the installation position strictly controlled to be 1.5m away from the internal iron core, transmitting signals through a 50Ω coaxial cable to ensure optimal capture of deep electromagnetic pulses; four ultrasonic sensors are configured and uniformly attached to the sidewall of the tank using magnetic adsorption, and are distributed vertically at gradients of 0.5m, 1.5m, 2.5m, and 3.5m from the top to construct a three-dimensional acoustic positioning array and avoid mechanical vibration interference; one high-frequency current sensor is configured and fitted onto the 80mm diameter transformer neutral grounding wire to ensure close contact with the grounding wire to capture ground leakage current; two transient ground voltage sensors are configured and adsorbed onto the metal flange at the root of the high-voltage bushing, and located on the shaded side to avoid direct sunlight affecting measurement accuracy. The aforementioned sensors converge to an edge unit installed in the control cabinet next to the transformer. This unit is powered by 220VAC and maintains a real-time connection with the remote monitoring platform via a 4G communication module.

[0166] At the signal acquisition and processing level, the sampling rate was set to 1MHz, with a single acquisition duration of 10 minutes. Acquisition was initiated at 2:00 AM and 12:00 PM daily to avoid background electromagnetic interference caused by peak electricity consumption in the morning and evening. After adaptive processing of the acquired raw data, the terminal achieved significant signal optimization results: the signal-to-noise ratio of the UHF signal increased from 15dB to 32dB, the environmental interference pulse rejection rate in the ultrasonic signal reached 92%, and after compression algorithm processing, the amount of data collected in a single round decreased significantly from 1.2GB to 240MB, effectively reducing the transmission load on the 4G network.

[0167] At the AI ​​diagnostic and trend warning level, the terminal, based on multimodal fusion features, output a diagnostic result indicating that the transformer had "internal discharge," with a confidence probability as high as 99.2%. Maintenance personnel conducted a cover inspection based on this result and ultimately found a 5cm² tear in the core insulation paper, completely consistent with the system's diagnostic conclusion. Simultaneously, based on accumulated historical discharge data from the past six months (monthly averages of 120, 150, 180, 220, 260, and 310 pC respectively), the terminal used an LSTM model to predict that the discharge amounts for the 7th to 9th months would reach 380 pC, 450 pC, and 520 pC respectively. Since the predicted value for the 9th month exceeded the preset 500 pC alarm threshold, the terminal immediately triggered a yellow warning signal. Based on this warning, maintenance personnel arranged for a shutdown and maintenance two months in advance, successfully preventing equipment burnout due to further insulation deterioration. This fully demonstrates the technical value of this application's shift from "reactive post-event repair" to "predictive maintenance."

[0168] In this embodiment, through multimodal sensing collaboration, adaptive filtering, and attention fusion mechanisms, a signal-to-noise ratio improvement of ≥17dB and efficient rejection of interference pulses of ≥90% are achieved, reducing the false alarm rate to below 0.8%. Combined with an improved CNN-LSTM spatiotemporal feature fusion model, the system shortens the diagnostic cycle from hours to minutes while maintaining a fault identification accuracy of ≥98.5%. Furthermore, the system constructs an insulation degradation early warning system lasting 3-6 months based on LSTM time-series prediction, helping to reduce unplanned power outages by 43% and significantly lowering maintenance costs. This solution supports flexible configuration and cost control for different equipment such as GIS and transformers, as well as diverse scenarios such as substations and new energy sources, demonstrating strong engineering adaptability.

[0169] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps. It is understood that the steps in different embodiments can be freely combined as needed, and all non-contradictory solutions formed by such combinations are within the scope of protection of this application.

[0170] Based on the same inventive concept, this application also provides a partial discharge detection device for implementing the partial discharge detection method described above. The solution provided by this device is similar to the solution described in the above method; therefore, the specific limitations in one or more embodiments of the partial discharge detection device provided below can be found in the limitations of the partial discharge detection method described above, and will not be repeated here.

[0171] In one exemplary embodiment, such as Figure 3 As shown, a partial discharge detection device is provided, comprising: a partial discharge data acquisition module 310, a model input module 320, a model analysis module 330, an attention weight determination module 340, and a detection result output module 350, wherein:

[0172] The partial discharge data acquisition module 310 is used to acquire partial discharge monitoring signals of various modes of the device under test, and the signal quality indicators corresponding to the partial discharge monitoring signals of various modes.

[0173] The model input module 320 is used to input the partial discharge monitoring signals of various modalities into the trained fault detection model; the fault detection model is obtained by training a spatial feature extraction branch and a temporal feature extraction branch based on partial discharge sample data of various modalities and the fault labels corresponding to the partial discharge sample data.

[0174] Model analysis module 330 is used to output the spatiotemporal characteristics corresponding to the partial discharge monitoring signals of various modes from the fault detection model;

[0175] Attention weight determination module 340 is used to determine the attention weight corresponding to each of the various modes of partial discharge monitoring signals based on each of the signal quality indicators.

[0176] The detection result output module 350 is used to fuse the spatiotemporal features corresponding to the partial discharge monitoring signals of various modes using the attention weights of each mode to obtain multimodal fusion features, and to determine the partial discharge detection result based on the multimodal fusion features.

[0177] In one embodiment, the signal quality metric includes the signal-to-noise ratio; the attention weight determination module 340 is further configured to:

[0178] The signal-to-noise ratios corresponding to the partial discharge monitoring signals of various modes are mapped to normalized numerical weights using a preset mapping function; wherein, the numerical weights increase as the signal-to-noise ratio increases.

[0179] The numerical weights are determined as the attention weights of the partial discharge monitoring signals for the corresponding modes.

[0180] In one embodiment, the model analysis module 330 is further configured to:

[0181] Using the spatial feature extraction branch in the fault detection model, the input partial discharge monitoring signal is convolved, and the spatial distribution features of the partial discharge monitoring signal of various modes are obtained based on the convolution operation result.

[0182] Using the time-series feature extraction branch in the fault detection model, the input partial discharge monitoring signal is processed in a time-series manner, and the time-series processing results are used to obtain the time-dependent features of the partial discharge monitoring signal of various modes.

[0183] Based on the spatial distribution characteristics and the dependency characteristics, the spatiotemporal characteristics corresponding to the partial discharge monitoring signals of the various modes are determined.

[0184] In one embodiment, the partial discharge data acquisition module 310 is configured to:

[0185] The original analog signal for the device under test is acquired by a preset sensor group; the sensor group includes at least two of the following: ultra-high frequency sensor, ultrasonic sensor, high frequency current sensor, and transient ground voltage sensor.

[0186] Based on the pre-acquired timing signal, the original analog signals of each mode are time-aligned to obtain the aligned multimodal signals;

[0187] The aligned multimodal signal is filtered and compressed to obtain the partial discharge monitoring signal.

[0188] In one embodiment, the detection result output module 350 is further configured to:

[0189] The multimodal fusion features are input into the fully connected classification layer of the fault detection model to calculate the probability values ​​of the device under test belonging to the normal state and various preset fault types; the preset fault types include corona discharge, suspension discharge, surface discharge, internal discharge and arc discharge.

[0190] The fault type corresponding to the highest probability value is determined as the partial discharge detection result.

[0191] In one embodiment, the detection result output module 350 is further configured to:

[0192] Obtain the fault type corresponding to multiple monitoring time points, and obtain the discharge intensity corresponding to each monitoring time point;

[0193] By correlating the monitoring time points, the fault types, and the discharge intensity, a historical monitoring sequence is obtained;

[0194] A trend analysis is performed on the discharge intensity corresponding to the fault type at multiple detection time points in the historical monitoring sequence, and the insulation degradation trend of the device under test is determined based on the trend analysis results.

[0195] When the insulation degradation trend meets the preset alarm conditions, a trend warning message is generated.

[0196] Each module in the aforementioned partial discharge detection device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in the processor of a computer device in hardware form or independent of it, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0197] In one exemplary embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 4As shown, the computer device includes a processor, memory, input / output interface, communication interface, display unit, and input device. The processor, memory, and input / output interface are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interface. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The input / output interface is used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, Near Field Communication (NFC), or other technologies. When executed by the processor, the computer program implements a partial discharge detection method.

[0198] Those skilled in the art will understand that Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0199] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0200] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0201] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0202] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0203] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0204] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0205] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A partial discharge detection method characterized by, The method includes: Acquire partial discharge monitoring signals of various modes of the device under test, and the corresponding signal quality indicators of the partial discharge monitoring signals of various modes; The partial discharge monitoring signals of the various modalities are input into the trained fault detection model; the fault detection model is obtained by training spatial feature extraction branch and temporal feature extraction branch based on partial discharge sample data of multiple modalities and the fault labels corresponding to the partial discharge sample data. The fault detection model outputs the spatiotemporal characteristics corresponding to the partial discharge monitoring signals of various modes; Based on each of the aforementioned signal quality indicators, the attention weights corresponding to the partial discharge monitoring signals of the various modes are determined. The spatiotemporal features corresponding to the partial discharge monitoring signals of various modalities are fused using the attention weights of each modality to obtain multimodal fusion features. Based on the multimodal fusion features, the partial discharge detection result is determined.

2. The method of claim 1, wherein, The signal quality metrics include the signal-to-noise ratio; determining the attention weights corresponding to each of the multiple modal partial discharge monitoring signals based on each of the signal quality metrics includes: The signal-to-noise ratios corresponding to the partial discharge monitoring signals of various modes are mapped to normalized numerical weights using a preset mapping function; wherein, the numerical weights increase as the signal-to-noise ratio increases. The numerical weights are determined as the attention weights of the partial discharge monitoring signals for the corresponding modes.

3. The method of claim 1, wherein, The spatiotemporal features corresponding to the partial discharge monitoring signals of various modes output by the fault detection model include: Using the spatial feature extraction branch in the fault detection model, the input partial discharge monitoring signal is convolved, and the spatial distribution features of the partial discharge monitoring signal of various modes are obtained based on the convolution operation result. Using the time-series feature extraction branch in the fault detection model, the input partial discharge monitoring signal is processed in a time-series manner, and the time-series processing results are used to obtain the time-dependent features of the partial discharge monitoring signal of various modes. Based on the spatial distribution characteristics and the dependency characteristics, the spatiotemporal characteristics corresponding to the partial discharge monitoring signals of the various modes are determined.

4. The partial discharge detection method of claim 1, wherein, The acquisition of partial discharge monitoring signals of multiple modes of the device under test includes: The original analog signal for the device under test is acquired by a preset sensor group; the sensor group includes at least two of the following: ultra-high frequency sensor, ultrasonic sensor, high frequency current sensor, and transient ground voltage sensor. Based on the pre-acquired timing signal, the original analog signals of each mode are time-aligned to obtain the aligned multimodal signals; The aligned multimodal signal is filtered and compressed to obtain the partial discharge monitoring signal.

5. The method of claim 1, wherein, The determination of partial discharge detection results based on multimodal fusion features includes: The multimodal fusion features are input into the fully connected classification layer of the fault detection model to calculate the probability values ​​of the device under test belonging to the normal state and various preset fault types; the preset fault types include corona discharge, suspension discharge, surface discharge, internal discharge and arc discharge. The fault type corresponding to the highest probability value is determined as the partial discharge detection result.

6. The method of claim 5, wherein, After determining the partial discharge detection result based on multimodal fusion features, the process also includes: Obtain the fault type corresponding to multiple monitoring time points, and obtain the discharge intensity corresponding to each monitoring time point; By correlating the monitoring time points, the fault types, and the discharge intensity, a historical monitoring sequence is obtained; A trend analysis is performed on the discharge intensity corresponding to the fault type at multiple detection time points in the historical monitoring sequence, and the insulation degradation trend of the device under test is determined based on the trend analysis results. When the insulation degradation trend meets the preset alarm conditions, a trend warning message is generated.

7. A partial discharge detection device, characterized by, The device includes: The partial discharge data acquisition module is used to acquire partial discharge monitoring signals of various modes of the device under test, and the signal quality indicators corresponding to the partial discharge monitoring signals of various modes. The model input module is used to input the partial discharge monitoring signals of various modalities into the trained fault detection model; the fault detection model is obtained by training spatial feature extraction branch and temporal feature extraction branch based on partial discharge sample data of multiple modalities and the fault labels corresponding to the partial discharge sample data. The model analysis module is used to output the spatiotemporal characteristics corresponding to the partial discharge monitoring signals of various modes from the fault detection model; The attention weight determination module is used to determine the attention weight corresponding to each of the various modes of partial discharge monitoring signals based on each of the signal quality indicators. The detection result output module is used to fuse the spatiotemporal features corresponding to the partial discharge monitoring signals of various modalities using the attention weights of each module to obtain multimodal fusion features, and to determine the partial discharge detection result based on the multimodal fusion features. 8.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-7. When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

10. A computer program product comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images