GIS ultrahigh frequency partial discharge abnormity early warning method and system

Abstract

The invention discloses a GIS ultrahigh frequency partial discharge abnormity early warning method and system, and relates to the technical field of power equipment fault diagnosis, and the method comprises the steps: collecting a pulse signal, carrying out the time domain normalization and leading edge detection, constructing a four-dimensional spatial-temporal feature matrix based on a detection result, and generating a structured spatial-temporal feature tensor through a coding rule; dividing the structured spatial-temporal characteristic tensor into a plurality of sub-matrixes according to a time window, reconstructing a six-dimensional phase space through an optimal embedding dimension and a time delay algorithm, and deeply fusing the reconstructed six-dimensional phase space with chaotic dynamic characteristics through a coupling equation to generate chaotic trajectory data; extracting kinetic parameters based on the chaotic trajectory data, and fusing the kinetic parameters through a deep learning model to generate a discharge anomaly risk score; according to the method, through phase-space reconstruction and chaos dynamics modeling, the nonlinear dynamic characteristics of partial discharge are accurately captured, and the prediction precision of the abnormal state is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power equipment fault diagnosis, and in particular to a GIS ultra-high frequency partial discharge abnormality early warning method and system. Background Art

[0002] Gas-insulated switchgear (GIS) is an indispensable key device in modern power systems, and the stability of its operating status is directly related to the safety and reliability of the power grid. Ultra-high frequency (UHF) partial discharge detection technology, due to its non-invasive nature, high sensitivity, and real-time monitoring capabilities, has become an important means of assessing the status of GIS equipment. In recent years, with the rapid development of sensor technology, signal processing algorithms, and artificial intelligence, UHF partial discharge detection technology has made significant progress. Traditional partial discharge detection methods primarily rely on time-domain and frequency-domain analysis, extracting features such as the amplitude, frequency, and phase of pulse signals to determine whether equipment anomalies exist. However, these methods are often limited to single-dimensional feature extraction and cannot fully reflect the complex dynamic characteristics of partial discharge. To address this shortcoming, researchers have begun to introduce advanced technologies such as chaos theory, phase space reconstruction, and deep learning, attempting to improve the accuracy and reliability of partial discharge anomaly detection through multi-dimensional feature fusion and complex nonlinear modeling.

[0003] While existing technologies have achieved some success in partial discharge detection, they still face several shortcomings. First, traditional methods for processing pulse signals often rely on manually set thresholds and rules, lacking adaptability and making them difficult to adapt to complex and changing field environments. Second, existing technologies for modeling the dynamic characteristics of partial discharge are often limited to linear or simple nonlinear relationships, failing to fully utilize chaotic dynamics, resulting in insufficient prediction accuracy for abnormal conditions. Furthermore, existing methods lack systematicity and dynamism in the design of risk scoring and early warning mechanisms, failing to achieve a complete closed-loop system from data collection to anomaly report generation. These shortcomings limit the comprehensiveness and real-time nature of GIS equipment status assessments, necessitating a more intelligent and efficient method for early warning of partial discharge anomalies. Summary of the Invention

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

[0005] Therefore, the present invention provides a GIS ultra-high frequency partial discharge abnormal warning method to solve the problems of the prior art lacking adaptive capability for pulse signal processing and insufficient modeling of partial discharge dynamic characteristics.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions: In the first aspect, the present invention provides a GIS ultra-high frequency partial discharge anomaly warning method, which includes collecting pulse signals and performing time domain normalization and frontier detection, constructing a four-dimensional space-time feature matrix based on the detection results, and generating a structured space-time feature tensor using coding rules; dividing the structured space-time feature tensor into multiple sub-matrices according to the time window, reconstructing the six-dimensional phase space through the optimal embedding dimension and time delay algorithm, and deeply fusing the reconstructed six-dimensional phase space with the chaotic dynamic characteristics through the coupling equation to generate chaotic trajectory data; extracting dynamic parameters based on the chaotic trajectory data, and fusing the dynamic parameters through a deep learning model to generate a discharge anomaly risk score; performing dynamic warning according to the discharge anomaly risk score, and generating an anomaly detection report.

[0007] As a preferred solution of the GIS ultra-high frequency partial discharge abnormal warning method of the present invention, wherein: the pulse signal is collected and time domain normalization and leading edge detection are performed, a dimensional spatiotemporal feature matrix is constructed based on the detection results, and a structured spatiotemporal feature tensor is generated using encoding rules. The specific steps are as follows: Pulse signals include voltage pulses, current pulses, electromagnetic pulses and light pulses; Eliminate the amplitude difference of pulse signals and unify the signal range through time domain normalization; By calculating the differential value of the pulse signal, the signal change rate area is identified, and the differential threshold T is set. When the differential value exceeds T, it is determined as the starting point of the leading edge; Perform leading edge detection on the pulse signal starting from the leading edge starting point, and record the position and time domain characteristics of all leading edge points; Based on the positions of all leading-edge points and the time domain features, the time dimension, amplitude dimension, frequency dimension and shape dimension of each pulse are extracted to generate the four-dimensional features of each pulse; Combine the four-dimensional features of each pulse into a feature vector and arrange them in time sequence to construct a four-dimensional spatiotemporal feature matrix; Embedding coding is used to convert the four-dimensional spatiotemporal feature matrix into a continuous vector and stack it in time order to generate a structured spatiotemporal feature tensor.

[0008] As a preferred solution of the GIS ultra-high frequency partial discharge abnormal warning method of the present invention, the structured spatiotemporal feature tensor is divided into multiple sub-matrices according to the time window, and the six-dimensional phase space is reconstructed by the optimal embedding dimension and time delay algorithm to generate the six-dimensional phase space trajectory. The specific steps are as follows: The spatiotemporal feature tensor is divided into multiple sub-matrices according to the time axis, and each sub-matrix serves as the spatiotemporal feature within a time window; Based on the spatiotemporal features, the mutual information method is used to calculate the time delay of each feature dimension, and the time delay with the minimum correlation between the spatiotemporal feature dimensions is selected as the optimal delay. The expression is: ; in, Indicates time The eigenvalue at represents the time delay between feature dimensions, Indicates time The eigenvalue at express and The joint probability distribution of express The marginal probability distribution of express The marginal probability distribution of represents the optimal delay; Use the false nearest neighbor method to identify the optimal embedding dimension; For the spatiotemporal features within each time window, the phase space is reconstructed according to the optimal delay and optimal embedding dimension, and a six-dimensional phase space point is generated through feature expansion; Each time point The corresponding six-dimensional phase space points are connected in time sequence to generate a six-dimensional phase space trajectory.

[0009] As a preferred solution of the GIS ultra-high frequency partial discharge abnormal warning method of the present invention, wherein: the reconstructed six-dimensional phase space is deeply integrated with the chaotic dynamic characteristics through the coupling equation to generate chaotic trajectory data. The specific steps are as follows: The six-dimensional phase space trajectory is modeled using the chaotic dynamics equation, and the six-dimensional phase space points are input into the chaotic dynamics equation as initial conditions, and the chaotic dynamics characteristics are obtained by numerical integration method. The six-dimensional phase space trajectory and chaotic dynamics characteristics are standardized; Based on the standardized six-dimensional phase space trajectory and chaotic dynamic characteristics, the generator and discriminator in GAN are trained alternately to deeply integrate the six-dimensional phase space trajectory with the chaotic dynamic characteristics to generate chaotic trajectory data. The expression is: ; in, is the normalized six-bit phase space trajectory, is the normalized hybrid dynamic characteristic, is the bias vector of the hidden layer, is the weight matrix of the generator’s hidden layer, is the weight matrix of the output layer of the generator, is the weight matrix of the generator weight adjustment layer, is the bias vector of the output layer, It is the chaotic trajectory data.

[0010] As a preferred solution of the GIS ultra-high frequency partial discharge abnormal warning method of the present invention, the steps of extracting dynamic parameters based on chaotic trajectory data are as follows: Use the Wolf algorithm to extract the Lyapunov exponent; The fractal dimension was extracted using the box counting method and the entropy value was extracted using the approximate entropy algorithm; GP algorithm is used to extract chaotic trajectory dimensional features; Extract periodic features through fast Fourier transform; The extracted kinetic parameters were arranged in chronological order to construct a kinetic parameter matrix, which was then standardized.

[0011] As a preferred solution of the GIS ultra-high frequency partial discharge anomaly early warning method of the present invention, wherein: the kinetic parameters are integrated through the deep learning model to generate the discharge anomaly risk score, the specific steps are as follows: Initialize the Transformer weights and bias parameters through Xavier; Train the initialized Transformer using forward propagation, binary cross entropy loss function, backpropagation algorithm and Adam optimizer; Based on the standardized kinetic parameter matrix, the time series characteristics of the kinetic parameters are identified through Transformer, and the nonlinear dependency between the kinetic parameters is learned by combining the multi-head attention mechanism to predict the discharge abnormality risk score. .

[0012] As a preferred solution of the GIS ultra-high frequency partial discharge anomaly early warning method of the present invention, wherein: the dynamic early warning is performed based on the discharge anomaly risk score and the anomaly detection report is generated, the specific steps are as follows: Based on the historical discharge abnormality data statistics, a low-risk threshold S1 and a high-risk threshold S2 are defined; when When ≥S2, the current GIS UHF partial discharge anomaly is considered to be at high risk and a level 1 warning is issued; When S1≤ When the value is less than S2, the current GIS UHF partial discharge anomaly is considered to be at medium risk and a level 2 warning is issued; when When the value is less than S1, the current GIS UHF partial discharge anomaly is considered to be at low risk, and a level 3 warning is issued; During the early warning process, real-time pulse data at the time of abnormal discharge is collected in real time, and an abnormality detection report is generated.

[0013] In the second aspect, the present invention provides a GIS ultra-high frequency partial discharge anomaly warning system, including a feature tensor generation module, a chaos trajectory data generation module, a risk assessment module and a report generation module; the feature tensor generation module is used to collect pulse signals and perform time domain normalization and frontier detection, construct a dimensional space-time feature matrix based on the detection results, and generate a structured space-time feature tensor using coding rules; the chaos trajectory data generation module is used to divide the structured space-time feature tensor into multiple sub-matrices according to the time window, reconstruct the six-dimensional phase space through the optimal embedding dimension and time delay algorithm, and deeply fuse the reconstructed six-dimensional phase space with the chaotic dynamic characteristics through the coupling equation to generate chaotic trajectory data; the risk assessment module is used to extract dynamic parameters based on the chaotic trajectory data, and fuse the dynamic parameters through a deep learning model to generate a discharge anomaly risk score; the report generation module is used to perform dynamic warning according to the discharge anomaly risk score and generate an anomaly detection report.

[0014] In a third aspect, the present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, any step of the GIS ultra-high frequency partial discharge abnormal warning method as described in the first aspect of the present invention is implemented.

[0015] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, any step of the GIS ultra-high frequency partial discharge abnormal warning method as described in the first aspect of the present invention is implemented.

[0016] The beneficial effects of the present invention are as follows: through phase space reconstruction and chaotic dynamics modeling, the nonlinear dynamic characteristics of partial discharge are accurately captured, and the prediction accuracy of abnormal states is significantly improved; through the Transformer model combined with the multi-head attention mechanism, the complex nonlinear dependencies between dynamic parameters are efficiently integrated, and the accuracy and reliability of risk scoring are enhanced; the risk threshold is dynamically adjusted based on historical data, realizing a complete closed loop from data collection to abnormality report generation, and improving the systematicness and dynamism of the early warning mechanism. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] 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.

[0018] Figure 1 This is a flow chart of the GIS ultra-high frequency partial discharge abnormal warning method in Example 1.

[0019] Figure 2 This is a schematic diagram of the GIS ultra-high frequency partial discharge abnormal warning system in Example 1. DETAILED DESCRIPTION

[0020] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0021] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0022] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0023] Example 1, reference Figure 1 and Figure 2 , which is the first embodiment of the present invention, provides a GIS ultra-high frequency partial discharge abnormal warning method, comprising the following steps: S1. Collect pulse signals and perform time domain normalization and leading-edge detection. Based on the detection results, construct a four-dimensional spatiotemporal feature matrix, and use encoding rules to generate a structured spatiotemporal feature tensor.

[0024] Pulse signals include voltage pulses, current pulses, electromagnetic pulses and light pulses; It should be noted that voltage pulses: High-bandwidth voltage probes or voltage dividers are typically used to capture the rapid voltage changes occurring within GIS equipment. These devices can accurately measure transient voltage fluctuations caused by partial discharge and convert them into data that can be further analyzed.

[0025] Current pulses: Rogowski coils or other types of current sensors are used to detect weak current changes caused by partial discharge within the GIS. These sensors have good frequency response characteristics and are suitable for capturing rapidly changing current signals.

[0026] Electromagnetic pulses (EMPs): These use ultra-high frequency (UHF) antennas or sensor arrays to monitor electromagnetic radiation released by partial discharges within GIS equipment. These sensors are particularly suitable for non-invasive monitoring because they can detect the electromagnetic waves emitted by internal discharge events from the outside.

[0027] Light pulses: For light pulse collection, photomultiplier tubes (PMTs) or more advanced optical sensors may be used. These devices are sensitive to the photon activity generated by partial discharge, especially in the ultraviolet and visible spectrum.

[0028] Eliminate the amplitude difference of pulse signals and unify the signal range through time domain normalization; Furthermore, the amplitude of each sampling point in each pulse signal is divided by the maximum absolute value of the pulse signal, so that the maximum amplitude of the processed pulse signal becomes 1, and the minimum amplitude is adjusted accordingly according to the situation of the original signal, but overall it ensures that all pulse signals are scaled to the same amplitude range.

[0029] By calculating the differential value of the pulse signal, the signal change rate area is identified, and the differential threshold T is set. When the differential value exceeds T, it is determined as the starting point of the leading edge; It should be noted that, first, a differential operation is performed on each collected voltage pulse, current pulse, electromagnetic pulse, and optical pulse signal. This calculation calculates the amplitude change between adjacent sampling points to determine the signal's rate of change. Next, a reasonable differential threshold, T, is set based on the actual application scenario and noise level. During the analysis process, if the differential value at a point exceeds the preset differential threshold, T, that point is considered to mark the beginning of a significant signal change and is therefore determined to be the starting point of a leading edge.

[0030] Perform leading edge detection on the pulse signal starting from the leading edge starting point, and record the position and time domain characteristics of all leading edge points; It should be noted that, first, the starting point of the leading edge in each voltage pulse, current pulse, electromagnetic pulse, and optical pulse signal is determined based on the differential threshold T set in the previous step. Once the leading edge starting point is determined, the pulse signal is scanned along the time axis from that point, identifying all points on the leading edge of the pulse. These points mark areas where the signal rapidly rises or falls. During the scanning process, the time position (relative to the timestamp of the pulse signal's start) of each point identified as a leading edge and its corresponding amplitude change are recorded as time domain features.

[0031] Based on the positions of all leading-edge points and the time domain features, the time dimension, amplitude dimension, frequency dimension and shape dimension of each pulse are extracted to generate the four-dimensional features of each pulse; It should be noted that for each voltage pulse, current pulse, electromagnetic pulse, and optical pulse signal, we first calculate its time dimension (i.e., the time interval between the starting and ending points of the leading edge), amplitude dimension (the amplitude variation range from the starting point of the leading edge to the highest or lowest point), frequency dimension (using fast Fourier transform to analyze the main frequency components of the leading edge), and shape dimension (based on the morphological characteristics of the leading edge, such as slope changes). These dimensions together constitute the four-dimensional feature set that describes the pulse characteristics.

[0032] Combine the four-dimensional features of each pulse into a feature vector and arrange them in time sequence to construct a four-dimensional spatiotemporal feature matrix; It should be noted that the time dimension, amplitude dimension, frequency dimension, and shape dimension of each pulse extracted above are combined into a feature vector, ensuring that each feature vector contains a complete four-dimensional feature description of the corresponding pulse signal. Then, these feature vectors are arranged in the chronological order of the pulse signals to form a matrix structure, namely the four-dimensional space-time feature matrix.

[0033] Embedding coding is used to convert the four-dimensional spatiotemporal feature matrix into a continuous vector and stack it in time order to generate a structured spatiotemporal feature tensor.

[0034] It should be noted that, for example, Principal Component Analysis (PCA) is used as a mapping rule to project the original high-dimensional data into a low-dimensional space through linear transformation, while maintaining the main features of the original data while reducing the data dimension, so that the characteristics of each pulse signal can be presented in a more compact and easy-to-process form. Then, according to the time sequence of the original pulse signals such as voltage pulses, current pulses, electromagnetic pulses, and light pulses, the continuous vectors obtained by the above transformation are stacked in sequence. This stacking method not only takes into account the multi-dimensional feature information within a single pulse signal, but also preserves the time series relationship between different pulse signals, thereby forming a structured spatiotemporal feature tensor.

[0035] S2. Divide the structured spatiotemporal feature tensor into multiple sub-matrices according to the time window, reconstruct the six-dimensional phase space through the optimal embedding dimension and time delay algorithm, and deeply integrate the reconstructed six-dimensional phase space with the chaotic dynamic characteristics through the coupling equation to generate chaotic trajectory data.

[0036] The spatiotemporal feature tensor is divided into multiple sub-matrices according to the time axis, and each sub-matrix serves as the spatiotemporal feature within a time window; Based on the spatiotemporal features, the mutual information method is used to calculate the time delay of each feature dimension, and the time delay with the minimum correlation between the spatiotemporal feature dimensions is selected as the optimal delay. The expression is: ; in, Indicates time The eigenvalue at represents the time delay between feature dimensions, Indicates time The eigenvalue at express and The joint probability distribution of express The marginal probability distribution of express The marginal probability distribution of represents the optimal delay; It should be noted that, first, we consider the spatiotemporal characteristics of the original signals such as voltage pulses, current pulses, electromagnetic pulses and light pulses after processing. and The eigenvalues on and ), calculate their joint probability distribution and their respective marginal probability distribution. The joint probability distribution represents the probability distribution at the same time point or delay. The probability of two eigenvalues appearing together under the same time is given by the marginal probability distribution, while the marginal probability distribution describes the probability of the eigenvalue appearing at a single time point. Next, by calculating all possible time delays Down, and The mutual information between Mutual information is defined as the logarithmic expectation of the ratio of the joint probability distribution to the marginal probability distribution. The higher the value, the stronger the correlation between the two; conversely, the greater the independence between the two. Therefore, the goal is to find a time delay , at this delay, and The mutual information between is the smallest, which means that the eigenvalues at these two time points have the lowest correlation or the highest independence. , calculate the corresponding mutual information value. Then, select the time delay that makes the mutual information reach the minimum value as the optimal delay .

[0037] Use the false nearest neighbor method to identify the optimal embedding dimension; The specific process is to determine the optimal embedding dimension for each voltage pulse, current pulse, electromagnetic pulse, and optical pulse by gradually increasing the embedding dimension and calculating the change in the nearest neighbor distance at each dimension. During this process, if the additional embedding dimension does not significantly change the nearest neighbor distance, it indicates that the embedding dimension is sufficient to correctly unfold the inherent structure of the dataset.

[0038] For the spatiotemporal features within each time window, the phase space is reconstructed according to the optimal delay and optimal embedding dimension, and a six-dimensional phase space point is generated through feature expansion; It should be noted that, first, based on the determined optimal delay and embedding dimension, corresponding data points are selected from the spatiotemporal characteristics of the voltage pulse, current pulse, electromagnetic pulse, and optical pulse within each time window. Next, these data points are used to construct high-dimensional vectors, ensuring that each vector reflects the dynamic characteristics of the original signal at the specified time delay and embedding dimension. Feature expansion operations are then performed on these high-dimensional vectors to supplement the necessary information, allowing each vector to be converted into a six-dimensional phase space point.

[0039] Each time point The corresponding six-dimensional phase space points are connected in time sequence to generate a six-dimensional phase space trajectory.

[0040] It should be noted that, first, all six-dimensional phase space points are sorted by timestamp, ensuring that they are arranged in the time sequence of their original pulse signals. Next, starting with the earliest time point, adjacent time points are connected in sequence to form a continuous trajectory. This trajectory depicts the state change path of the voltage pulse, current pulse, electromagnetic pulse, and optical pulse over time in the six-dimensional phase space.

[0041] The six-dimensional phase space trajectory is modeled using the chaotic dynamics equation, and the six-dimensional phase space points are input into the chaotic dynamics equation as initial conditions, and the chaotic dynamics characteristics are obtained by numerical integration method. It should be noted that, first, based on the six-dimensional phase space trajectory, appropriate chaotic dynamics equations are selected to describe the dynamic behavior of voltage pulses, current pulses, electromagnetic pulses, and light pulses. Next, the six-dimensional phase space point corresponding to each time point t is used as the initial condition to initiate the solution of the chaotic dynamics equations. By applying numerical integration methods, the state changes at subsequent time points are gradually calculated, thereby obtaining chaotic dynamics characteristics that reflect the complex dynamic characteristics of the original pulse signal.

[0042] The six-dimensional phase space trajectory and chaotic dynamics characteristics are standardized; Based on the standardized six-dimensional phase space trajectory and chaotic dynamic characteristics, the generator and discriminator in GAN are trained alternately to deeply integrate the six-dimensional phase space trajectory with the chaotic dynamic characteristics to generate chaotic trajectory data. The expression is: ; in, is the normalized six-bit phase space trajectory, is the normalized hybrid dynamic characteristic, is the bias vector of the hidden layer, is the weight matrix of the generator’s hidden layer, is the weight matrix of the output layer of the generator, is the weight matrix of the generator weight adjustment layer, is the bias vector of the output layer, It is the chaotic trajectory data.

[0043] It should be noted that the generator in a generative adversarial network (GAN) first performs a series of nonlinear transformations on the input normalized six-dimensional phase space trajectory and chaotic dynamic characteristics. In this process, the six-dimensional phase space trajectory, serving as the primary input feature, undergoes a series of complex operations, including but not limited to the application of activation functions to introduce nonlinear factors, thereby capturing the complex dynamic characteristics of the original pulse signal. Simultaneously, the chaotic dynamic characteristics are also incorporated into the computational framework and integrated with the information from the six-dimensional phase space trajectory. After processing through a multi-layered structure within the generator, a new representation that integrates both aspects of information is output. This new representation is then input into the discriminator along with real data samples. Through repeated iterative training, the generator parameters are optimized to ensure that the generated data closely resembles the real data distribution. Ultimately, chaotic trajectory data that accurately reflects the dynamic characteristics of voltage pulses, current pulses, electromagnetic pulses, and optical pulses is obtained.

[0044] S3. Extract dynamic parameters based on chaotic trajectory data, and fuse the dynamic parameters through a deep learning model to generate a discharge abnormality risk score.

[0045] Use the Wolf algorithm to extract the Lyapunov exponent; It should be noted that, first, based on the six-dimensional phase space trajectory, an initial point and its nearest neighbor are selected, and the distance between these two points is calculated. Next, the evolution of these two points over time is tracked along the trajectory until the distance between them exceeds a certain threshold, and the change in distance at this point is recorded. This process is repeated for multiple pairs of points along the trajectory, and the Lyapunov exponent is ultimately determined by calculating the average growth rate of the distance change between these pairs.

[0046] The fractal dimension was extracted using the box counting method and the entropy value was extracted using the approximate entropy algorithm; It should be noted that, first, the six-dimensional phase space trajectory is projected into a suitable lower-dimensional space for analysis. Then, a series of grids (boxes) of decreasing size are placed on this low-dimensional representation, and the minimum number of boxes required to cover the trajectory at each grid size is calculated. As the grid size decreases, the number of required boxes increases. By analyzing the relationship between the number of boxes and their size, the fractal dimension can be determined.

[0047] GP algorithm is used to extract chaotic trajectory dimensional features; It should be noted that, first, a set of randomly generated mathematical models is initialized based on the data of the six-dimensional phase space trajectory. Through genetic programming, selection, crossover, and mutation operations, each mathematical model is evaluated in each iteration based on its fit to the data, and the best performing model is selected for the next iteration. After multiple rounds of optimization, a highly optimized mathematical model is ultimately obtained that accurately captures and describes the complex dynamic characteristics of the six-dimensional phase space trajectory, thereby extracting the dimensional characteristics of the chaotic trajectory.

[0048] Extract periodic features through fast Fourier transform; It should be noted that, first, a fast Fourier transform is performed on the six-dimensional phase space trajectory to convert the time-domain signal into a frequency-domain representation. This is done to identify the dominant frequency components and their corresponding amplitudes within the signal. By analyzing the spectrum, it is possible to clearly observe which frequency components dominate the voltage, current, electromagnetic, and optical pulse signals, thereby extracting the periodic characteristics of the pulse signals.

[0049] The extracted kinetic parameters were arranged in chronological order to construct a kinetic parameter matrix, which was then standardized.

[0050] It should be noted that, first, the dynamic parameters extracted from voltage pulses, current pulses, electromagnetic pulses, and optical pulses, such as the Lyapunov exponent, fractal dimension, entropy, chaotic trajectory dimensionality, and periodicity, are sorted according to their corresponding time points. These ordered dynamic parameters are then combined into a matrix structure, the dynamic parameter matrix, ensuring that the position of each parameter in the matrix corresponds to its position in the original time series. Next, to eliminate possible scale differences between different parameters, each element in the dynamic parameter matrix is normalized so that all parameter values fall within the same range.

[0051] Initialize the Transformer weights and bias parameters through Xavier; Train the initialized Transformer using forward propagation, binary cross entropy loss function, backpropagation algorithm, and Adam optimizer; It should be noted that, first, the normalized kinetic parameter matrix is input into a Transformer initialized with Xavier. The forward propagation process is then performed layer by layer until the abnormal discharge risk score is output. Next, a binary cross-entropy loss function is used to compare the difference between the Transformer's predictions and the actual labels, quantifying the prediction error. Based on this error, a backpropagation algorithm is then used to calculate the gradients of the weights and biases at each layer, determining how to adjust these parameters to reduce the error. Finally, the Adam optimizer, combining the first and second moment estimates of the gradient, adaptively adjusts the learning rate of each parameter, achieving efficient parameter updates, thereby minimizing the loss function and optimizing Transformer performance. The entire process ensures consistency and accuracy from spike signal processing to Transformer training.

[0052] Based on the standardized kinetic parameter matrix, the time series characteristics of the kinetic parameters are identified through Transformer, and the nonlinear dependency between the kinetic parameters is learned by combining the multi-head attention mechanism to predict the discharge abnormality risk score. , the expression is: ; in, is the normalized kinetic parameter matrix In the The feature vector of each time step, It is a multi-head attention mechanism. is the total number of time steps, is the index of the time step, is the discharge abnormality risk score.

[0053] It should be noted that, first, the normalized dynamic parameter matrix is input into the Transformer model, and its encoder component extracts the dynamic parameter feature vector for each time step. Next, a multi-head attention mechanism is applied to calculate the importance weight of each time step feature vector, allowing the model to focus on the complex dependencies between different time steps. Each time step feature vector is then multiplied by the corresponding attention weight and processed through an activation function to enhance the representation of important features while suppressing irrelevant information. All these processed time step feature vectors are aggregated and integrated into a single output value through a nonlinear transformation (such as the tanh function), representing a comprehensive feature representation of the entire time series. Finally, this comprehensive feature representation is used to calculate the discharge anomaly risk score S(t), ensuring that the score reflects the dynamic changes and potential anomalies in the original voltage, current, electromagnetic, and optical pulse signals.

[0054] S4. Provide dynamic warning based on the discharge abnormality risk score and generate an abnormality detection report.

[0055] Based on the historical discharge abnormality data statistics, a low-risk threshold S1 and a high-risk threshold S2 are defined; when When S2 is greater than or equal to 0.8, the GIS UHF partial discharge anomaly is considered to be at high risk and a first-level warning is issued. For example, if the high-risk threshold S2 is set to 0.8, if the calculated discharge anomaly risk score is If it is 0.9, a level one warning is issued, indicating that there is a highly abnormal risk and immediate action is required.

[0056] When S1≤ When the risk score of the GIS UHF partial discharge is less than S2, the current GIS UHF partial discharge anomaly is considered to be at medium risk and a secondary warning is issued; for example, the low risk threshold S1 is set to 0.3 and the high risk threshold S2 is set to 0.8. If the calculated discharge anomaly risk score is If it is 0.6, a level 2 warning is issued, indicating that there is a moderate risk and further inspection is recommended.

[0057] when When the S1 is less than the current GIS UHF partial discharge anomaly, it is considered that the current GIS UHF partial discharge anomaly is at low risk and a level 3 warning is issued; for example, if the low risk threshold S1 is set to 0.3, if the calculated discharge anomaly risk score is If it is 0.2, a level 3 warning is issued, indicating that the current risk is low, but continuous monitoring is still needed to ensure safety.

[0058] During the early warning process, real-time pulse data at the time of abnormal discharge is collected in real time, and an abnormality detection report is generated.

[0059] It should be noted that the anomaly detection report includes a detailed analysis of GIS UHF partial discharge, including the discharge anomaly risk score. Historical trends, current risk levels (based on The report includes information on the following: (1) Level 1, 2, or 3 warnings (determined by comparison with preset thresholds S1 and S2), real-time collected voltage pulse, current pulse, electromagnetic pulse, and optical pulse data and their feature analysis, time series features identified by the Transformer model, the importance weight distribution of feature vectors at each time step using a multi-head attention mechanism, and potential fault causes and recommended maintenance measures derived from this data analysis. The report may also include predictions of possible future abnormalities, helping decision makers develop preventative strategies to ensure the safe and stable operation of power equipment.

[0060] This embodiment also provides a GIS ultra-high frequency partial discharge anomaly warning system, including: a feature tensor generation module, a chaos trajectory data generation module, a risk assessment module and a report generation module; the feature tensor generation module is used to collect pulse signals and perform time domain normalization and frontier detection, construct a dimensional space-time feature matrix based on the detection results, and generate a structured space-time feature tensor using coding rules; the chaos trajectory data generation module is used to divide the structured space-time feature tensor into multiple sub-matrices according to the time window, reconstruct the six-dimensional phase space through the optimal embedding dimension and time delay algorithm, and deeply fuse the reconstructed six-dimensional phase space with the chaotic dynamic characteristics through the coupling equation to generate chaotic trajectory data; the risk assessment module is used to extract dynamic parameters based on the chaotic trajectory data, and fuse the dynamic parameters through a deep learning model to generate a discharge anomaly risk score; the report generation module is used to perform dynamic warning according to the discharge anomaly risk score and generate an anomaly detection report.

[0061] This embodiment also provides a computer device suitable for the GIS ultra-high frequency partial discharge abnormal warning method, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the GIS ultra-high frequency partial discharge abnormal warning method proposed in the above embodiment.

[0062] The computer device may be a terminal, comprising a processor, memory, a communication interface, a display, and an input device connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores an operating system and computer programs. The internal memory provides an environment for the operating system and computer programs stored in the non-volatile storage media. The communication interface of the computer device is used to communicate with external terminals via wired or wireless communication. Wireless communication may be achieved via Wi-Fi, a carrier network, NFC (near-field communication), or other technologies. The display of the computer device may be a liquid crystal display or an electronic ink display. The input device may be a touchscreen overlay on the display, buttons, a trackball, or a touchpad on the computer device housing, or an external keyboard, touchpad, or mouse.

[0063] This embodiment also provides a storage medium having a computer program stored thereon. When the program is executed by a processor, the program implements the GIS ultra-high frequency partial discharge abnormal warning method proposed in the above embodiment; the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk.

[0064] In summary, the present invention accurately captures the nonlinear dynamic characteristics of partial discharge and significantly improves the prediction accuracy of abnormal states through: phase space reconstruction and chaotic dynamics modeling; through the Transformer model combined with the multi-head attention mechanism, it efficiently integrates the complex nonlinear dependencies between dynamic parameters and enhances the accuracy and reliability of risk scoring; dynamically adjusts the risk threshold based on historical data, realizing a complete closed loop from data collection to abnormality report generation, and improving the systematicness and dynamism of the early warning mechanism.

[0065] Example 2, referring to Table 1, is the second embodiment of the present invention. To further verify the technical solution of the present invention, experimental simulation data of the GIS ultra-high frequency partial discharge abnormal warning method are provided.

[0066] In order to verify the effectiveness of a new GIS ultra-high frequency partial discharge anomaly warning method, a comparative test was designed, using existing technologies and the method of the present invention for analysis. First, multiple sensors were installed in a laboratory environment to collect voltage pulses, current pulses, electromagnetic pulses, and optical pulse signals. These signals represent the status of GIS equipment under normal operating conditions and simulate some common fault conditions to generate abnormal data. Then, all collected pulse signals were normalized in the time domain to eliminate amplitude differences and unify the signal range. The signal change rate area was identified by calculating the differential value, and an appropriate differential threshold T was set to determine the starting point of the front. Based on the results of the front detection, the time dimension, amplitude dimension, frequency dimension, and shape dimension of each pulse were extracted to construct a four-dimensional spatiotemporal feature matrix, and further embedded coding was used to generate a structured spatiotemporal feature tensor.

[0067] Next, the structured spatiotemporal feature tensor is divided into multiple sub-matrices according to the time window, and the mutual information method is used to calculate the optimal time delay. The false nearest neighbor method is combined to determine the optimal embedding dimension and reconstruct the six-dimensional phase space. Subsequently, the chaotic dynamics equation is applied to model the six-dimensional phase space trajectory, and the chaotic dynamics characteristics are obtained by numerical integration. The generator and discriminator in GAN are used for alternating adversarial training to deeply fuse the six-dimensional phase space trajectory with the chaotic dynamics characteristics to generate chaotic trajectory data. Based on these chaotic trajectory data, the Wolf algorithm is used to extract the Lyapunov exponent, the box counting method is used to extract the fractal dimension, the approximate entropy algorithm is used to extract the entropy value, the GP algorithm is used to extract the chaotic trajectory dimensional features, and the periodic features are extracted by fast Fourier transform. Finally, these dynamic parameters are fused through a deep learning model (Transformer) to predict the discharge abnormality risk score. , and provide dynamic warnings and generate anomaly detection reports based on the scores.

[0068] The existing technologies used in this experiment specifically employ traditional partial discharge detection methods, primarily including time-domain and frequency-domain analysis techniques. These methods primarily rely on extracting single-dimensional features such as the amplitude, frequency, and phase of pulse signals, and then determine whether the equipment is abnormal by setting fixed thresholds. The specific steps include performing simple time-domain normalization processing on the collected voltage, current, electromagnetic, and optical pulse signals. These methods then use preset thresholds to detect the leading edge of the pulse signal and extract basic time and amplitude information based on these points. However, these methods lack in-depth analysis of the complex dynamic characteristics of the signals.

[0069] The details are shown in Table 1: Table 1 Comparative data of GIS partial discharge anomaly detection Through the data analysis of the above table, it can be clearly seen that the present invention has significant advantages in chaotic trajectory dimensional characteristics, fractal dimension, Lyapunov index and discharge abnormality risk score. For example, when the time window is 35.2 milliseconds, the chaotic trajectory dimension characteristic of the existing technology is 0.589, while the method of the present invention reaches 0.784; the fractal dimension is improved from 1.08 to 1.28; the Lyapunov exponent is improved from -0.123 to 0.024; the discharge abnormality risk score is significantly better than the existing technology. The accuracy of the proposed method has been improved from 0.289 to 0.684. This method uses multi-dimensional feature extraction, deep learning models, and chaotic dynamics fusion technologies to achieve a comprehensive and in-depth analysis of partial discharge signals, significantly improving detection accuracy and reliability. This allows for early detection of potential problems and provides more accurate early warning information.

[0070] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A GIS ultra-high frequency partial discharge abnormality early warning method, characterized by: include, Collect pulse signals and perform time domain normalization and leading edge detection. Based on the detection results, a four-dimensional spatiotemporal feature matrix is constructed, and a structured spatiotemporal feature tensor is generated using encoding rules. The structured spatiotemporal feature tensor is divided into multiple sub-matrices according to the time window, and the six-dimensional phase space is reconstructed through the optimal embedding dimension and time delay algorithm. The reconstructed six-dimensional phase space is deeply integrated with the chaotic dynamic characteristics through the coupling equation to generate chaotic trajectory data; Extract dynamic parameters based on chaotic trajectory data and fuse them through a deep learning model to generate a discharge abnormality risk score; Dynamic early warning is carried out based on the discharge abnormality risk score, and an abnormality detection report is generated.

2. The GIS ultra-high frequency partial discharge abnormality early warning method according to claim 1, characterized in that: The pulse signal is collected and time domain normalization and frontier detection are performed, a dimensional spatiotemporal feature matrix is constructed based on the detection results, and a structured spatiotemporal feature tensor is generated using encoding rules. The specific steps are as follows: Pulse signals include voltage pulses, current pulses, electromagnetic pulses and light pulses; Eliminate the amplitude difference of pulse signals and unify the signal range through time domain normalization; By calculating the differential value of the pulse signal, the signal change rate area is identified, and the differential threshold T is set. When the differential value exceeds T, it is determined as the starting point of the leading edge; Perform leading edge detection on the pulse signal starting from the leading edge starting point, and record the position and time domain characteristics of all leading edge points; Based on the positions of all leading-edge points and the time domain features, the time dimension, amplitude dimension, frequency dimension and shape dimension of each pulse are extracted to generate the four-dimensional features of each pulse; Combine the four-dimensional features of each pulse into a feature vector and arrange them in time sequence to construct a four-dimensional spatiotemporal feature matrix; Embedding coding is used to convert the four-dimensional spatiotemporal feature matrix into a continuous vector and stack it in time order to generate a structured spatiotemporal feature tensor.

3. The GIS ultra-high frequency partial discharge abnormality early warning method according to claim 2, characterized in that: The structured spatiotemporal feature tensor is divided into multiple sub-matrices according to the time window, and the six-dimensional phase space is reconstructed through the optimal embedding dimension and time delay algorithm to generate the six-dimensional phase space trajectory. The specific steps are as follows: The spatiotemporal feature tensor is divided into multiple sub-matrices according to the time axis, and each sub-matrix serves as the spatiotemporal feature within a time window; Based on the spatiotemporal features, the mutual information method is used to calculate the time delay of each feature dimension, and the time delay with the minimum correlation between the spatiotemporal feature dimensions is selected as the optimal delay. The expression is: ； in, Indicates time The eigenvalue at represents the time delay between feature dimensions, Indicates time The eigenvalue at express and The joint probability distribution of express The marginal probability distribution of express The marginal probability distribution of represents the optimal delay; Use the false nearest neighbor method to identify the optimal embedding dimension; For the spatiotemporal features within each time window, the phase space is reconstructed according to the optimal delay and optimal embedding dimension, and a six-dimensional phase space point is generated through feature expansion; Each time point The corresponding six-dimensional phase space points are connected in time sequence to generate a six-dimensional phase space trajectory.

4. The GIS ultra-high frequency partial discharge abnormality early warning method according to claim 3, characterized in that: The reconstructed six-dimensional phase space is deeply integrated with the chaotic dynamic characteristics through coupling equations to generate chaotic trajectory data. The specific steps are as follows: The six-dimensional phase space trajectory is modeled using the chaotic dynamics equation, and the six-dimensional phase space points are input into the chaotic dynamics equation as initial conditions, and the chaotic dynamics characteristics are obtained by numerical integration method. The six-dimensional phase space trajectory and chaotic dynamics characteristics are standardized; Based on the standardized six-dimensional phase space trajectory and chaotic dynamic characteristics, the generator and discriminator in GAN are trained alternately to deeply integrate the six-dimensional phase space trajectory with the chaotic dynamic characteristics to generate chaotic trajectory data. The expression is: ； in, is the normalized six-bit phase space trajectory, is the normalized hybrid dynamic characteristic, is the bias vector of the hidden layer, is the weight matrix of the generator’s hidden layer, is the weight matrix of the output layer of the generator, is the weight matrix of the generator weight adjustment layer, is the bias vector of the output layer, It is the chaotic trajectory data.

5. The GIS ultra-high frequency partial discharge abnormality early warning method according to claim 4, characterized in that: The specific steps of extracting dynamic parameters based on chaotic trajectory data are as follows: Use the Wolf algorithm to extract the Lyapunov exponent; The fractal dimension was extracted using the box counting method and the entropy value was extracted using the approximate entropy algorithm; The GP algorithm is used to extract the chaotic trajectory dimensional features; Extract periodic features through fast Fourier transform; The extracted kinetic parameters were arranged in chronological order to construct a kinetic parameter matrix, which was then standardized.

6. The GIS ultra-high frequency partial discharge abnormality early warning method according to claim 5, characterized in that: The kinetic parameters are integrated through the deep learning model to generate a discharge abnormality risk score. The specific steps are as follows: Initialize the Transformer weights and bias parameters through Xavier; Train the initialized Transformer using forward propagation, binary cross entropy loss function, backpropagation algorithm, and Adam optimizer; Based on the standardized kinetic parameter matrix, the time series characteristics of the kinetic parameters are identified through Transformer, and the nonlinear dependency between the kinetic parameters is learned by combining the multi-head attention mechanism to predict the discharge abnormality risk score. .

7. The GIS ultra-high frequency partial discharge abnormality early warning method according to claim 6, characterized in that: The specific steps for performing dynamic early warning based on the discharge abnormality risk score and generating an abnormality detection report are as follows: Based on the historical discharge abnormality data statistics, a low-risk threshold S1 and a high-risk threshold S2 are defined; when When ≥S2, the current GIS UHF partial discharge anomaly is considered to be at high risk and a level 1 warning is issued; When S1≤ When the value is less than S2, the current GIS UHF partial discharge anomaly is considered to be at medium risk and a level 2 warning is issued; when When the value is less than S1, the current GIS UHF partial discharge anomaly is considered to be at low risk, and a level 3 warning is issued; During the early warning process, real-time pulse data at the time of abnormal discharge is collected in real time, and an abnormality detection report is generated.

8. A GIS ultra-high frequency partial discharge abnormality early warning system, based on the GIS ultra-high frequency partial discharge abnormality early warning method according to any one of claims 1 to 7, characterized in that: Including, characteristic tensor generation module, chaos trajectory data generation module, risk assessment module and report generation module; The feature tensor generation module is used to collect pulse signals and perform time domain normalization and leading edge detection. Based on the detection results, a dimensional spatiotemporal feature matrix is constructed, and a structured spatiotemporal feature tensor is generated using encoding rules. Chaotic trajectory data generation module is used to split the structured spatiotemporal feature tensor into multiple sub-matrices according to time windows, reconstruct the six-dimensional phase space through the optimal embedding dimension and time delay algorithm, and deeply integrate the reconstructed six-dimensional phase space with the chaotic dynamic characteristics through coupling equations to generate chaotic trajectory data; The risk assessment module is used to extract dynamic parameters based on chaotic trajectory data and fuse the dynamic parameters through a deep learning model to generate a discharge abnormality risk score; The report generation module is used to provide dynamic early warning based on the discharge abnormality risk score and generate anomaly detection reports.

9. 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 GIS ultra-high frequency partial discharge abnormal warning method described in any one of claims 1 to 7 are implemented.

10. 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 GIS ultra-high frequency partial discharge abnormal warning method according to any one of claims 1 to 7 are implemented.

Images