Intelligent detection and positioning method and device for cable partial discharge based on multi-modal fusion

Abstract

This invention discloses a method and apparatus for intelligent detection and location of partial discharge in cables based on multimodal fusion, relating to the field of cable fault location technology. It involves synchronizing current signals, ultrasonic signals, and ultra-high frequency electromagnetic wave signals from both ends of the cable to obtain two sets of multimodal signals. These two sets of multimodal signals are then fed into a spatiotemporal hypergraph convolutional adaptive denoising network to obtain two sets of enhanced feature groups. The time difference between these two sets of enhanced feature groups is calculated to obtain a time difference value, which is used to determine the initial discharge position. Based on the initial discharge position, a preset optimization algorithm is used for dual-end positioning to obtain the target discharge position. Finally, the two sets of multimodal signals and the time difference value are substituted into a multi-feature fusion intelligent discharge recognition model to obtain the partial discharge type. The enhanced multimodal signals are used to calculate multiple sets of time difference values ​​to adaptively calibrate the equivalent ultrasonic wave propagation speed. Then, a preset optimization algorithm is used for dual-end iterative optimization to eliminate positioning errors caused by fixed wave speeds, thereby improving the accuracy of intelligent detection and location of partial discharge in cables.

Description

Technical Field

[0001] This invention belongs to the field of cable fault location technology, specifically relating to a method and device for intelligent detection and location of partial discharge in cables based on multimodal fusion. Background Technology

[0002] As a critical carrier of power transmission, the reliability of cable insulation directly affects the safe and stable operation of the power grid. Partial discharge is a major sign and manifestation of cable insulation degradation. Accurate and real-time location of partial discharge is a core technical means to achieve online monitoring of cable insulation status and prevent sudden faults. Currently, cable partial discharge location technology mainly adopts the double-end traveling wave method, which calculates the location of the discharge source based on the time difference of the partial discharge signal propagation to both ends of the cable and a preset equivalent wave velocity.

[0003] However, existing technologies mostly use single-mode signals for time difference calculation. But when a single-mode signal propagates in a cable, it is affected by dispersion, multipath reflection, and attenuation distortion, resulting in significant waveform broadening and deformation, making accurate identification of the first wave arrival time extremely difficult. For example, a cable local fault location method based on discharge detection disclosed in publication number CN118465452A only collects single current data and determines the fault location through characteristic comparison analysis. This method cannot overcome the problem of ambiguity in the first wave of a single-mode signal, leading to large deviations in time difference estimation.

[0004] Furthermore, the propagation speed of partial discharge signals is not a constant value, but changes in real time with factors such as signal frequency, cable insulation aging, and ambient temperature. Existing methods usually pre-set a fixed equivalent wave velocity and substitute it into the two-end positioning formula, ignoring the time-varying characteristics of the wave velocity. This fixed setting will introduce systematic positioning errors. More importantly, existing methods lack an adaptive calibration mechanism for the equivalent propagation speed of ultrasonic waves and cannot use the arrival time difference relationship of multi-mode signals to dynamically estimate the actual wave velocity, resulting in low positioning accuracy under different working conditions. Summary of the Invention

[0005] The purpose of this invention is to solve the problem that using a single-mode signal and relying on a fixed preset wave velocity cannot overcome the influence of the first wave ambiguity and the time-varying wave velocity, resulting in large positioning errors and low detection and positioning accuracy. Therefore, this invention proposes a method and device for intelligent detection and positioning of partial discharge in cables based on multi-mode fusion.

[0006] In a first aspect of this invention, a method for intelligent detection and localization of partial discharge in cables based on multimodal fusion is first proposed, the method comprising: Synchronously acquire current signals, ultrasonic signals, and ultra-high frequency electromagnetic wave signals from both ends of the target cable to obtain group A multimode signals and group B multimode signals; The target multimodal signal is substituted into a spatiotemporal hypergraph convolutional adaptive denoising network to obtain an enhanced feature set; the target multimodal signal consists of a group of multimodal signals A and a group of multimodal signals B. The time difference value is obtained by calculating the enhanced feature groups of group A and group B multimodal signals, and the initial discharge position is determined based on the time difference value. Based on the initial discharge position, the target discharge position is obtained by dual-end positioning using a preset optimization algorithm; The enhanced feature sets of group A and group B multimodal signals, along with the time difference value, are substituted into the multi-feature fusion intelligent discharge identification model to obtain the partial discharge type.

[0007] By synchronously acquiring current, ultrasonic, and ultra-high frequency electromagnetic wave signals from both ends of the cable, a dual-end multimodal signal group is constructed. A spatiotemporal hypergraph convolutional adaptive denoising network is used to collaboratively enhance the multimodal signals, effectively suppressing interference and sharpening the first wave features. Based on the enhanced multimodal signals, multiple sets of time difference values ​​are accurately calculated to adaptively calibrate the equivalent propagation speed of ultrasonic waves, obtaining a highly reliable initial discharge position. Then, a preset optimization algorithm is used for dual-end iterative optimization to significantly eliminate the systematic positioning error caused by the fixed wave speed. Finally, while achieving high-precision target discharge positioning, the enhanced feature group and time difference values ​​can be input into a multi-feature fusion intelligent discharge recognition model to output the partial discharge type. This solves the technical bottlenecks of low positioning accuracy, poor wave speed adaptation capability, and low detection efficiency in existing technologies, and improves the accuracy of intelligent detection and positioning of partial discharge in cables.

[0008] Optionally, the enhanced feature set obtained by substituting the target multimodal signal into a spatiotemporal hypergraph convolutional adaptive denoising network includes: Perform time-frequency transformation on the signal of each mode in the target multimodal signal to obtain the time-frequency diagram of each mode; Each time-frequency unit in the time-frequency graph of each mode is defined as a graph node, and the graph nodes of all modes are stacked to obtain the initial node feature matrix; A hypergraph structure is constructed based on the temporal proximity and intermodal association between nodes in the graph, and a learnable dynamic weight is assigned to each hyperedge in the hypergraph structure. The initial node feature matrix and hypergraph structure are input into the spatiotemporal hypergraph convolutional network. The node feature matrix output by the last spatiotemporal hypergraph convolution is temporally downsampled to obtain the enhanced signal features of each modality. An enhanced feature group is generated based on all the enhanced signal features.

[0009] Hypergraph structures can simultaneously model the temporal continuity within the same modality and the complementary correlations between different modalities, overcoming the limitation of traditional pairwise graph structures that can only describe binary relationships. Learnable dynamic weights enable the network to adaptively adjust the importance of different hyperedges according to the actual noise distribution and signal characteristics, thereby accurately preserving the time-frequency characteristics of partial discharge pulses and suppressing noise in strong interference environments. Spatiotemporal hypergraph convolution achieves synergistic enhancement of multimodal signals through multi-order neighbor information transmission, effectively sharpening the waveform characteristics at the arrival time of the first wave, providing high signal-to-noise ratio enhancement features for subsequent accurate time difference calculation and wave velocity adaptive calibration, fundamentally solving the first wave ambiguity problem and improving positioning accuracy and robustness.

[0010] Optionally, the time difference is calculated by performing time difference calculation on the enhanced feature groups of group A and group B multimodal signals to obtain the time difference value, and the initial discharge position is determined based on the time difference value, including: The first time difference is obtained by calculating the arrival time difference between the ultrasonic signal at point A and the UHF electromagnetic wave signal at point A; the second time difference is obtained by calculating the arrival time difference between the ultrasonic signal at point B and the UHF electromagnetic wave signal at point B; and the third time difference is obtained by calculating the arrival time difference between the ultrasonic signals at points A and B. The equivalent propagation velocity of the ultrasonic wave at end A is determined based on the first time difference, and the equivalent propagation velocity of the ultrasonic wave at end B is determined based on the second time difference. The target velocity is obtained by averaging the equivalent propagation velocities of the ultrasonic waves at end A and end B. Substituting the third time difference and the target velocity The initial discharge position is obtained; Where L is the distance from the power source to end A, and D is the total length of the cable from end A to end B. This represents the absolute value of the arrival time difference of the ultrasound signal between end A and end B. It is half the sum of the equivalent propagation velocities of the ultrasound at ends A and B.

[0011] By using a high-precision time reference of ultra-high frequency electromagnetic wave signals, adaptive segment-by-segment calibration of the equivalent propagation velocity of ultrasound was achieved, completely eliminating the dependence on fixed preset wave velocity in traditional methods and eliminating systematic positioning errors introduced by time-varying wave velocity. At the same time, the dual-end wave velocity averaging effectively reduced the interference of single-end noise or local distortion on wave velocity estimation. Combined with the sharpened first wave moment of the enhanced feature group, the calculation accuracy of the initial discharge position was greatly improved, providing a highly reliable initial solution for subsequent optimization algorithms, reducing positioning errors and improving detection efficiency.

[0012] Optionally, based on the initial discharge position, obtaining the target discharge position through dual-end positioning using a preset optimization algorithm includes: Based on formula The optimal value L is searched within a preset interval using a preset optimization algorithm. The specific operation of the algorithm is as follows: Initialize the differential evolution subpopulation and the particle swarm subpopulation; Differential evolutionary subpopulations are optimized iteratively through mutation, crossover, and selection; Particle swarm optimization updates particle positions using individual optimality and global optimality. If a smaller F(L) value than the current global optimum of the particle swarm is found during differential evolution, then the global guiding position of the particle swarm is replaced. The process is iterated alternately until the convergence condition is met, and the optimal L is output to obtain the target discharge position.

[0013] Differential evolution's strong global exploration capability effectively avoids particle swarm optimization from getting stuck in local optima, while the rapid convergence of particle swarm optimization improves optimization efficiency. The synergy of the two significantly enhances the optimization capability for complex non-convex positioning error models. The mechanism of dynamically replacing the global guiding position ensures the real-time fusion of the advantages of the two algorithms, enabling the target discharge position to further approximate the actual discharge source coordinates based on the initial positioning. This maximizes the elimination of residual systematic and random errors, significantly improves positioning accuracy and robustness, and meets the computational efficiency requirements of online monitoring.

[0014] Optionally, the enhanced feature sets of group A and group B multimodal signals and the time difference value are substituted into the multi-feature fusion intelligent discharge identification model to obtain the partial discharge types, including: The multi-feature fusion intelligent discharge recognition model includes a dual-end temporal feature encoder, a physical prior embedding module, a cross-modal fusion module, and a classification head; Substitute the enhanced feature sets of group A and group B multimodal signals into the dual-end temporal feature encoder to obtain the temporal feature vector; The time difference value is substituted into the physical prior embedding module and mapped into a physical feature vector; The temporal feature vector and the physical feature vector are substituted into the cross-modal fusion module to perform feature vector fusion to obtain the target fused feature; The target fusion features are substituted into the classification head to obtain the partial discharge type.

[0015] The dual-end timing feature encoder can fully utilize the spatiotemporal complementary information of multi-mode signals at both ends of the cable to capture complex discharge fingerprints that cannot be characterized by a single mode. The physical prior embedding module introduces high-precision time difference values ​​as constraint features into the model, making the classification results conform to the physical laws of discharge propagation and enhancing the interpretability and generalization ability of the model. The cross-modal fusion module realizes the synergistic enhancement of data-driven features and physical prior features, thereby improving the recognition accuracy of complex partial discharge modes such as surface discharge and multi-source discharge.

[0016] In a second aspect of this invention, a smart detection and location device for partial discharge in cables based on multimodal fusion is proposed, comprising: The signal acquisition module is used to simultaneously acquire current signals, ultrasonic signals, and ultra-high frequency electromagnetic wave signals from both ends of the target cable to obtain group A multimode signals and group B multimode signals; The feature enhancement module is used to substitute the target multimodal signal into a spatiotemporal hypergraph convolutional adaptive denoising network to obtain enhanced feature groups; the target multimodal signal consists of group A multimodal signal and group B multimodal signal; The initial discharge location determination module is used to calculate the time difference value by performing time difference calculation on the enhanced feature groups of group A multimodal signals and group B multimodal signals, and determine the initial discharge location based on the time difference value; The target discharge location determination module is used to obtain the target discharge location by performing dual-end positioning based on the initial discharge location and a preset optimization algorithm; The partial discharge type determination module is used to substitute the enhanced feature groups of group A multimodal signals and group B multimodal signals and the time difference value into the multi-feature fusion intelligent discharge recognition model to obtain the partial discharge type.

[0017] Optionally, the feature enhancement module includes: The time-frequency transformation module is used to perform time-frequency transformation on the signal of each mode in the target multimodal signal to obtain the time-frequency diagram of each mode; The initial node feature matrix determination module is used to define each time-frequency unit in the time-frequency graph of each mode as a graph node, and stack the graph nodes of all modes to obtain the initial node feature matrix; The hypergraph structure construction module is used to construct a hypergraph structure based on the temporal proximity and intermodal association between graph nodes, and to assign learnable dynamic weights to each hyperedge in the hypergraph structure. The enhanced feature group generation module is used to input the initial node feature matrix and hypergraph structure into the spatiotemporal hypergraph convolutional network, perform temporal downsampling on the node feature matrix output by the last spatiotemporal hypergraph convolution to obtain the enhanced signal features of each modality, and generate enhanced feature groups based on all enhanced signal features.

[0018] Optionally, the initial discharge location determination module includes: The time difference calculation module is used to calculate the arrival time difference between the ultrasonic signal at end A and the UHF electromagnetic wave signal at end A to obtain the first time difference; calculate the arrival time difference between the ultrasonic signal at end B and the UHF electromagnetic wave signal at end B to obtain the second time difference; and calculate the arrival time difference between the ultrasonic signals at end A and end B to obtain the third time difference. The target velocity calculation module is used to determine the equivalent propagation velocity of the ultrasonic wave at end A based on the first time difference, determine the equivalent propagation velocity of the ultrasonic wave at end B based on the second time difference, and average the equivalent propagation velocities of the ultrasonic waves at end A and end B to obtain the target velocity. The initial discharge position calculation module is used to substitute the third time difference and the target velocity into... The initial discharge position is obtained; Where L is the distance from the power source to end A, and D is the total length of the cable from end A to end B. This represents the absolute value of the arrival time difference of the ultrasound signal between end A and end B. It is half the sum of the equivalent propagation velocities of the ultrasound at ends A and B.

[0019] Optionally, the target discharge location determination module includes: The particle swarm optimization module is used for formula-based optimization. The optimal value L is searched within a preset interval using a preset optimization algorithm. The specific operation of the algorithm is as follows: Initialize the differential evolution subpopulation and the particle swarm subpopulation; Differential evolutionary subpopulations are optimized iteratively through mutation, crossover, and selection; Particle swarm optimization updates particle positions using individual optimality and global optimality. If a smaller F(L) value than the current global optimum of the particle swarm is found during differential evolution, then the global guiding position of the particle swarm is replaced. The process is iterated alternately until the convergence condition is met, and the optimal L is output to obtain the target discharge position.

[0020] Optionally, the partial discharge type determination module includes: The dual-end time-series feature encoder module is used to substitute the enhanced feature groups of group A and group B multimodal signals into the dual-end time-series feature encoder to obtain the time-series feature vector; The physical prior embedding module is used to map the time difference value into the physical prior embedding module as a physical feature vector; A cross-modal fusion module is used to substitute the temporal feature vector and the physical feature vector into the cross-modal fusion module to perform feature vector fusion to obtain the target fused feature; The classification head module is used to input the target fusion features into the classification head to obtain the partial discharge type. Attached Figure Description

[0021] The invention will now be further described with reference to the accompanying drawings.

[0022] Figure 1 A flowchart of a cable partial discharge intelligent detection and localization method based on multimodal fusion provided in an embodiment of the present invention; Figure 2 This is an execution flowchart of a spatiotemporal hypergraph convolutional adaptive denoising network provided in an embodiment of the present invention. Detailed Implementation

[0023] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0024] It should be noted that all formula calculations in the scheme are purely numerical calculations.

[0025] Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] This invention provides an intelligent detection and location method for partial discharge in cables based on multimodal fusion. See also... Figure 1 The method includes the following steps: S101, synchronously acquires current signals, ultrasonic signals and ultra-high frequency electromagnetic wave signals from both ends of the target cable to obtain group A multimode signals and group B multimode signals; S102, Substitute the target multimodal signal into the spatiotemporal hypergraph convolutional adaptive denoising network to obtain the enhanced feature group; the target multimodal signal consists of group A multimodal signal and group B multimodal signal; S103, calculate the time difference value by performing time difference calculation on the enhanced feature groups of group A multimode signal and group B multimode signal, and determine the initial discharge position based on the time difference value; S104, Based on the initial discharge position, the target discharge position is obtained by dual-end positioning through a preset optimization algorithm; S105, Substitute the enhanced feature groups and time difference values ​​of group A multimodal signals and group B multimodal signals into the multi-feature fusion intelligent discharge identification model to obtain the partial discharge type.

[0027] In one implementation, a three-in-one sensor group is installed at both end A (the beginning) and end B (the end) of the target cable (the cable to be tested, as determined in advance by the technicians). The sensor group includes a high-frequency current transformer (HFCT, bandwidth 100kHz to 50MHz), an ultrasonic sensor (AE, center frequency 40kHz to 150kHz), and an ultra-high frequency sensor (UHF, bandwidth 300MHz to 1.5GHz). All sensors are triggered synchronously, and the sampling rate is uniformly set to 100MS / s to ensure that the time difference resolution between the ultrasonic and ultra-high frequency signals reaches the 10ns level.

[0028] In one implementation, synchronous acquisition uses GPS or IEEE 1588 Precise Time Protocol (PTP) for clock synchronization, with a clock deviation of less than 20ns between the two ends. The continuous acquisition duration is 10ms, covering multiple power frequency cycles. Each acquisition is stored as raw waveform data of three modes: Group A and Group B. The raw signals undergo preliminary preprocessing: DC components are removed, power frequency interference is filtered out using a 50Hz notch filter, and out-of-band noise is filtered out using bandpass filters (current signals: 10kHz~30MHz; ultrasound: 30kHz~200kHz; ultra-high frequency: 100MHz~1GHz), resulting in preprocessed Group A multimode signals and Group B multimode signals.

[0029] In one embodiment, see Figure 2 , Figure 2 A flowchart of the execution of a spatiotemporal hypergraph convolutional adaptive denoising network is provided. The enhanced feature set obtained by substituting the target multimodal signal into the spatiotemporal hypergraph convolutional adaptive denoising network includes: S1021, Perform time-frequency transformation on the signal of each mode in the target multimodal signal to obtain the time-frequency diagram of each mode; S1022, define each time-frequency unit in the time-frequency graph of each mode as a graph node, and stack the graph nodes of all modes to obtain the initial node feature matrix; S1023, construct a hypergraph structure based on the temporal proximity relationship and intermodal association relationship between the nodes of each graph, and assign learnable dynamic weights to each hyperedge in the hypergraph structure; S1024: Input the initial node feature matrix and hypergraph structure into the spatiotemporal hypergraph convolutional network, perform temporal downsampling on the node feature matrix output by the last spatiotemporal hypergraph convolution to obtain the enhanced signal features of each modality, and generate an enhanced feature group based on all enhanced signal features.

[0030] In one implementation, the time-domain waveform is mapped to a time-frequency joint distribution, while preserving both the time and frequency information of the signal. This overcomes the limitation of single-domain analysis, which cannot simultaneously characterize the instantaneous occurrence characteristics and spectral features of partial discharge pulses. Each time-frequency unit acts as an independent node, and its node features can include multi-dimensional attributes such as amplitude, phase, and instantaneous frequency. Compared to directly using the original waveform or traditional feature extraction methods, this preserves the subtle structure of the partial discharge signal in the time-frequency plane. All time-frequency units of the current, ultrasonic, and ultra-high frequency modes are stacked in the same node feature matrix, enabling data from different modes to be jointly processed under the same graph structure, thus providing conditions for subsequent cross-modal information transmission.

[0031] In one implementation, connecting time-frequency units with continuous time windows within the same mode can capture the continuity and waveform extension characteristics of partial discharge pulses on the time axis, effectively enhancing the time-domain sharpness of the first wavefront and suppressing random spike noise; connecting time-frequency units with high correlation in different modes can uncover the physical coupling relationship between current, ultrasound, and ultra-high frequency, and use complementary information to mutually verify and collaboratively denoise.

[0032] In one implementation, a regular graph can only connect two nodes, while a hypergraph's hyperedge can connect any number of nodes, making it more suitable for describing the complex relationships of multiple time-frequency units simultaneously correlated in multimodal signals. The hypergraph structure greatly enhances the model's overall perception of partial discharge events. By using the hypergraph Laplacian operator to achieve multi-order message passing between nodes, the update of each node not only integrates information from neighboring time-frequency units in the same mode but also absorbs features from related units in other modes, achieving true multimodal collaborative enhancement.

[0033] In one implementation, current signals, ultrasonic signals, and electromagnetic wave signals are collected during the operation of the transformer. The length of each signal is L sampling points. A short-time Fourier transform is performed on the signal of each mode. Based on the transformed data, each time-frequency unit in the time-frequency graph of each mode is defined as a graph node. For a node, its initial feature vector is taken as the amplitude value of the time-frequency unit. The initial feature vectors of all nodes are stacked into a matrix in sequence.

[0034] In one implementation, for each mode m and each frequency point f, the continuous time axis is... Each frame is connected to form a hyperedge; this embodiment employs a multi-scale strategy, using... Three windows. Step size taken. The number of time superedges generated at each frequency point is For each time frame and each frequency point The same position in the three modes The three nodes are connected to form a hyperedge, which has a total of Each hyperedge has a modality; a learnable dynamic weight is assigned to each hyperedge in the hypergraph structure. Specifically, for each hyperedge... It is assigned a trainable scalar weight. The calculation method is as follows: The MLP contains one hidden layer with ReLU activation function. It is Sigmoid.

[0035] In one implementation, the initial node feature matrix and the constructed hypergraph structure are input into a spatiotemporal hypergraph convolutional network. This network consists of multiple layers of identical hypergraph convolutional modules stacked together. Each layer performs the following operations in sequence: Node-to-hyperedge aggregation: For each hyperedge, the features of all nodes within the hyperedge are aggregated into a hyperedge feature vector using an attention mechanism or weighted averaging. Hyperedge-to-node diffusion: The feature vector of each hyperedge is diffused back to the nodes it contains. The features received by each node from multiple hyperedges are weighted and summed, and then subjected to a linear transformation and a nonlinear activation function (ReLU) to obtain the intermediate features of the node. Residual connection: The features of the input nodes of the previous layer are added to the intermediate features obtained in the current layer to obtain the node features output by the current layer. Adaptive noise reduction: In each layer, a smoothing prior regularization term is introduced based on the hypergraph Laplacian matrix and the current node features. This regularization term controls the smoothing strength through an adaptively calculated coefficient, making the features of nodes connected by the same hyperedge tend to be consistent, thereby suppressing noise and highlighting effective partial discharge patterns.

[0036] In one implementation, after multiple layers of spatiotemporal hypergraph convolution, the node feature matrix output by the last layer of spatiotemporal hypergraph convolution is used as the final node feature matrix. Next, temporal downsampling is performed to recover the time series features of each mode. First, according to the mode to which the node belongs, the node feature matrix is ​​split into three sub-matrices, corresponding to current, ultrasound, and electromagnetic waves, respectively. Then, for each mode sub-matrix, average pooling is performed along the frequency dimension to average the feature vectors of all frequency points in each time frame to obtain the time feature matrix. Finally, one-dimensional convolution (kernel size 3×3, stride 2) is used to obtain the enhanced signal feature sequence of each mode. The enhanced signal feature sequences of the three modes are concatenated along the channel dimension to form the final enhanced feature group, which can be directly used by the subsequent partial discharge mode recognition network.

[0037] In one embodiment, the time difference is calculated by performing time difference calculation on the enhanced feature groups of group A and group B multimodal signals to obtain a time difference value, and the initial discharge position is determined based on the time difference value, including: The first time difference is obtained by calculating the arrival time difference between the ultrasonic signal at point A and the UHF electromagnetic wave signal at point A; the second time difference is obtained by calculating the arrival time difference between the ultrasonic signal at point B and the UHF electromagnetic wave signal at point B; and the third time difference is obtained by calculating the arrival time difference between the ultrasonic signals at points A and B. The equivalent propagation velocity of the ultrasonic wave at end A is determined based on the first time difference, and the equivalent propagation velocity of the ultrasonic wave at end B is determined based on the second time difference. The target velocity is obtained by averaging the equivalent propagation velocities of the ultrasonic waves at end A and end B. Substituting the third time difference and the target velocity The initial discharge position is obtained; Where L is the distance from the power source to end A, and D is the total length of the cable from end A to end B. This represents the absolute value of the arrival time difference of the ultrasound signal between end A and end B. It is half the sum of the equivalent propagation velocities of the ultrasound at ends A and B.

[0038] In one implementation, by simultaneously calculating three sets of time differences, consistency verification can be performed in subsequent processing to eliminate abnormal time difference values ​​caused by single-end noise or pulse false detection, thereby improving the reliability of initial positioning. Since the enhanced feature group has already undergone spatiotemporal hypergraph convolution denoising, the first wave of both the ultrasonic signal and the UHF signal waveforms is sharpened. At this time, the arrival time can be extracted using a dual-threshold or waveform matching algorithm with nanosecond-level accuracy. Compared to using only a single mode (such as pure ultrasonic dual-end positioning), introducing UHF as a time reference transforms absolute time measurement into relative time difference measurement, effectively offsetting system errors such as sensor response delay and asynchronous data acquisition.

[0039] In one implementation, estimating the global wave velocity using only the single time difference at end A or B may lead to measurement deviations due to local noise or signal attenuation at that end, severely affecting the positioning results. By averaging the wave velocities at both ends, single-end random errors can be mutually compensated, making the target velocity more robust; when the operating conditions at both ends of the cable are asymmetrical, the average velocity can still balance the overall propagation characteristics.

[0040] In one implementation, the initial discharge location does not need to be absolutely precise, but it needs to fall within a small neighborhood near the actual fault point. The initial discharge location given by multimodal time difference and adaptive wave velocity usually has high reliability, which allows the subsequent particle swarm-differential evolution hybrid optimization algorithm to converge quickly within a small search interval, avoid global blind search, and improve optimization efficiency and final positioning accuracy.

[0041] In one implementation, the first time difference is obtained by calculating the arrival time difference between the ultrasonic signal at end A and the ultra-high frequency electromagnetic wave signal at end A. The second time difference is obtained by calculating the arrival time difference between the ultrasound signal at end B and the ultra-high frequency electromagnetic wave signal at end B. ; Calculate the time difference of arrival of the ultrasound signal between end A and end B. At end A, the fixed installation distance between the ultrasonic sensor and the ultra-high frequency sensor is known. Since electromagnetic wave signals travel at the speed of light, their arrival time is considered instantaneous. If the absolute flight time of the ultrasonic signal from the discharge source to the sensor at point A is given, then the equivalent propagation speed of the ultrasonic wave at point A is... ,in This is the temperature compensation coefficient fed back by the temperature sensor at end A; similarly, the time difference at end B is used. and fixed distance Calculate the equivalent wave velocity at point B. Then through The target speed is calculated.

[0042] In one implementation, let the total length of the cable from end A to end B be D (a known constant), the distance from the discharge source to end A be L (a quantity to be determined), and the time it takes for the ultrasonic wave to travel from the discharge source to end A be... The time it takes to spread to the B end is Calculate the time difference between propagation to point A and propagation to point B. Therefore, the solution is obtained. However, in actual measurements, noise and wave speed fluctuations may cause residuals in the above analytical solutions. To improve stability, the localization problem is transformed into a nonlinear optimization problem that minimizes the objective function F(L), using the formula... The hybrid differential evolution-preset optimization algorithm searches for the optimal L within a preset interval [0, D]. The specific operations of the algorithm are as follows: Initialize the differential evolution subpopulation and the particle swarm subpopulation; The differential evolution subpopulation maintains population diversity through mutation, crossover, and selection operations to avoid getting trapped in local optima; The particle swarm subpopulation updates the particle position using individual optima and global optima; If differential evolution finds an F(L) smaller than the current global optima of the particle swarm, it replaces the global guiding position of the particle swarm with it; Iterate alternately until the convergence condition is met, and output the optimal L; The final output L is the distance of the discharge source from end A, completing the precise positioning of both ends.

[0043] In one embodiment, substituting the enhanced feature sets and time difference values ​​of group A and group B multimodal signals into the multi-feature fusion intelligent discharge identification model yields partial discharge types including: The multi-feature fusion intelligent discharge recognition model includes a dual-end temporal feature encoder, a physical prior embedding module, a cross-modal fusion module, and a classification head; Substitute the enhanced feature sets of group A and group B multimodal signals into the dual-end temporal feature encoder to obtain the temporal feature vector; Substitute the time difference value into the physical prior embedding module to map it into a physical feature vector; Substitute the temporal feature vector and physical feature vector into the cross-modal fusion module to perform feature vector fusion to obtain the target fused feature; Substituting the target fusion features into the classification head yields the partial discharge type.

[0044] In one implementation, a dual-end timing feature encoder can simultaneously process enhanced feature sequences of three modes (current, ultrasonic, and ultra-high frequency) at both ends (A and B). Compared to using only a single end or a single mode, dual-end information can reveal the time difference, waveform attenuation difference, and dispersion difference of the partial discharge pulse propagation to both ends, providing richer criteria for identifying the discharge type. For example, internal air gap discharge often produces a dual-end waveform with high symmetry, while surface discharge may result in significant asymmetry in the waveforms at both ends due to different propagation paths.

[0045] In one implementation, the partial discharge signal is essentially a transient pulse sequence. Its timing characteristics, such as pulse interval, phase distribution, and amplitude trend, are closely related to the discharge type. The timing feature encoder can automatically capture the long-range dependency between pulses through self-attention or cyclic mechanisms. It does not require manual design of statistical parameters and overcomes the shortcomings of traditional methods that rely on a few manual features, resulting in poor recognition and generalization ability.

[0046] In one implementation, different discharge types are often related to the fault location. For example, surface discharge is prone to occur at the joint, while air gap discharge inside the body is common in the middle section. The time difference value is essentially a quantitative expression of the discharge location. By embedding it into a physical feature vector, it can provide a location prior for the classification task and improve the accuracy of type recognition. The original time difference value has a limited range, and directly inputting it into the classifier does not work well. The physical prior embedding module uses a multi-layer fully connected network to map the time difference value to a high-dimensional latent space, which can extract the nonlinear interaction features between time differences.

[0047] In one implementation, the dual-end temporal feature encoder includes a multi-scale depthwise separable convolution and a bidirectional long short-term memory network. The multi-scale depthwise separable convolution is used to convert the multimodal enhancement features of groups A and B into high-dimensional temporal feature vectors, capturing the multi-scale local patterns and long-range temporal dependencies of the discharge pulses. Parallel branch convolution is performed on the multimodal enhancement features of groups A and B, with each branch using a depthwise convolution kernel of different scales to capture discharge features over different time spans. Branch 1: convolution kernel size k=3, expansion rate d=1; Branch 2: convolution kernel size... Branch 3: Kernel size k=5, dilation rate d=1; Branch 4: Kernel size k=9, dilation rate d=1; Each branch contains depthwise convolution, BatchNorm, ReLU, and pointwise convolution to obtain convolutional features. Then, the convolutional features are fed into a bidirectional long short-term memory network according to time steps to obtain semantic understanding features. The semantic understanding features of the multimodal enhancement features of groups A and B after passing through multi-scale depthwise separable convolution and bidirectional long short-term memory network are fused and averaged to obtain the temporal feature vector.

[0048] In one implementation, the time difference values ​​include the arrival time difference of the ultrasonic signal at end A relative to the UHF electromagnetic wave signal at end A; the arrival time difference of the ultrasonic signal at end B relative to the UHF electromagnetic wave signal at end B; calculating the arrival time difference of the ultrasonic signals between end A and end B; concatenating the three time differences into a time vector and substituting it into the physical prior embedding module, which is a two-layer fully connected network to obtain the physical feature vector; the activation function of the fully connected network is ReLU.

[0049] In one implementation, the cross-modal fusion module specifically projects the physical feature vector linearly onto the same dimension as the temporal feature to obtain the temporal physical feature vector, then substitutes it into the attention mechanism to obtain the attention weight, and multiplies the attention weight by the temporal physical feature vector, adds 1, and subtracts the attention weight multiplied by the physical feature vector to obtain the target fused feature.

[0050] In one implementation, the temporal feature vector originates from a high-dimensional abstraction of the actual waveform, reflecting the morphological characteristics of the discharge; the physical feature vector originates from time difference calculation, reflecting the location and propagation characteristics of the discharge. Since both are heterogeneous information, simple concatenation or addition cannot fully leverage their synergistic effect. The cross-modal fusion module can dynamically learn the correlation weights between the two types of features, enabling the model to consider both the waveform's inherent characteristics and location information when determining the discharge type.

[0051] The foregoing has described one embodiment of the present invention in detail, but this content is merely a preferred embodiment and should not be considered as limiting the scope of the present invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the scope of the claims of this invention.

Claims

1. A method for intelligent detection and localization of partial discharge in cables based on multimodal fusion, characterized in that, The method includes: Synchronously acquire current signals, ultrasonic signals, and ultra-high frequency electromagnetic wave signals from both ends of the target cable to obtain group A multimode signals and group B multimode signals; The target multimodal signal is substituted into a spatiotemporal hypergraph convolutional adaptive denoising network to obtain an enhanced feature set; the target multimodal signal consists of a group of multimodal signals A and a group of multimodal signals B. The time difference value is obtained by calculating the enhanced feature groups of group A and group B multimodal signals, and the initial discharge position is determined based on the time difference value. Based on the initial discharge position, the target discharge position is obtained by dual-end positioning using a preset optimization algorithm; The enhanced feature sets of group A and group B multimodal signals, along with the time difference value, are substituted into the multi-feature fusion intelligent discharge identification model to obtain the partial discharge type.

2. The intelligent detection and location method for partial discharge in cables based on multimodal fusion according to claim 1, characterized in that, Substituting the target multimodal signal into a spatiotemporal hypergraph convolutional adaptive denoising network yields an enhanced feature set, including: Perform time-frequency transformation on the signal of each mode in the target multimodal signal to obtain the time-frequency diagram of each mode; Each time-frequency unit in the time-frequency graph of each mode is defined as a graph node, and the graph nodes of all modes are stacked to obtain the initial node feature matrix; A hypergraph structure is constructed based on the temporal proximity and intermodal association between nodes in the graph, and a learnable dynamic weight is assigned to each hyperedge in the hypergraph structure. The initial node feature matrix and hypergraph structure are input into the spatiotemporal hypergraph convolutional network. The node feature matrix output by the last spatiotemporal hypergraph convolution is temporally downsampled to obtain the enhanced signal features of each modality. An enhanced feature group is generated based on all the enhanced signal features.

3. The intelligent detection and location method for partial discharge in cables based on multimodal fusion according to claim 1, characterized in that, The time difference value is obtained by calculating the time difference of the enhanced feature groups of group A and group B multimodal signals. The initial discharge position is determined based on the time difference value, including: The first time difference is obtained by calculating the arrival time difference between the ultrasonic signal at point A and the UHF electromagnetic wave signal at point A; the second time difference is obtained by calculating the arrival time difference between the ultrasonic signal at point B and the UHF electromagnetic wave signal at point B; and the third time difference is obtained by calculating the arrival time difference between the ultrasonic signals at points A and B. The equivalent propagation velocity of the ultrasonic wave at end A is determined based on the first time difference, and the equivalent propagation velocity of the ultrasonic wave at end B is determined based on the second time difference. The target velocity is obtained by averaging the equivalent propagation velocities of the ultrasonic waves at end A and end B. Substituting the third time difference and the target velocity The initial discharge position is obtained; Where L is the distance from the power source to end A, and D is the total length of the cable from end A to end B. This represents the absolute value of the arrival time difference of the ultrasound signal between end A and end B. It is half the sum of the equivalent propagation velocities of the ultrasound at ends A and B.

4. The intelligent detection and location method for partial discharge in cables based on multimodal fusion according to claim 3, characterized in that, Based on the initial discharge position, the target discharge position is obtained through dual-end positioning using a preset optimization algorithm, including: Based on formula The optimal value L is searched within a preset interval using a preset optimization algorithm. The specific operation of the algorithm is as follows: Initialize the differential evolution subpopulation and the particle swarm subpopulation; Differential evolutionary subpopulations are optimized iteratively through mutation, crossover, and selection; Particle swarm optimization updates particle positions using individual optimality and global optimality. If a smaller F(L) value than the current global optimum of the particle swarm is found during differential evolution, then the global guiding position of the particle swarm is replaced. The process is iterated alternately until the convergence condition is met, and the optimal L is output to obtain the target discharge position.

5. The intelligent detection and location method for partial discharge in cables based on multimodal fusion according to claim 1, characterized in that, Substituting the enhanced feature sets of group A and group B multimodal signals and the time difference value into the multi-feature fusion intelligent discharge identification model, the partial discharge types are obtained as follows: The multi-feature fusion intelligent discharge recognition model includes a dual-end temporal feature encoder, a physical prior embedding module, a cross-modal fusion module, and a classification head; Substitute the enhanced feature sets of group A and group B multimodal signals into the dual-end temporal feature encoder to obtain the temporal feature vector; The time difference value is substituted into the physical prior embedding module and mapped into a physical feature vector; The temporal feature vector and the physical feature vector are substituted into the cross-modal fusion module to perform feature vector fusion to obtain the target fused feature; The target fusion features are substituted into the classification head to obtain the partial discharge type.

6. A smart detection and location device for partial discharge in cables based on multimodal fusion, characterized in that, The device includes: The signal acquisition module is used to simultaneously acquire current signals, ultrasonic signals, and ultra-high frequency electromagnetic wave signals from both ends of the target cable to obtain group A multimode signals and group B multimode signals; The feature enhancement module is used to substitute the target multimodal signal into a spatiotemporal hypergraph convolutional adaptive denoising network to obtain enhanced feature groups; the target multimodal signal consists of group A multimodal signal and group B multimodal signal; The initial discharge location determination module is used to calculate the time difference value by performing time difference calculation on the enhanced feature groups of group A multimodal signals and group B multimodal signals, and determine the initial discharge location based on the time difference value; The target discharge location determination module is used to obtain the target discharge location by performing dual-end positioning based on the initial discharge location and a preset optimization algorithm; The partial discharge type determination module is used to substitute the enhanced feature groups of group A multimodal signals and group B multimodal signals and the time difference value into the multi-feature fusion intelligent discharge recognition model to obtain the partial discharge type.

7. The intelligent detection and positioning device for partial discharge of cables based on multimodal fusion according to claim 6, characterized in that, The feature enhancement module includes: The time-frequency transformation module is used to perform time-frequency transformation on the signal of each mode in the target multimodal signal to obtain the time-frequency diagram of each mode; The initial node feature matrix determination module is used to define each time-frequency unit in the time-frequency graph of each mode as a graph node, and stack the graph nodes of all modes to obtain the initial node feature matrix; The hypergraph structure construction module is used to construct a hypergraph structure based on the temporal proximity and intermodal association between graph nodes, and to assign learnable dynamic weights to each hyperedge in the hypergraph structure. The enhanced feature group generation module is used to input the initial node feature matrix and hypergraph structure into the spatiotemporal hypergraph convolutional network, perform temporal downsampling on the node feature matrix output by the last spatiotemporal hypergraph convolution to obtain the enhanced signal features of each modality, and generate enhanced feature groups based on all enhanced signal features.

8. The intelligent detection and positioning device for partial discharge of cables based on multimodal fusion according to claim 6, characterized in that, The initial discharge location determination module includes: The time difference calculation module is used to calculate the arrival time difference between the ultrasonic signal at end A and the UHF electromagnetic wave signal at end A to obtain the first time difference; calculate the arrival time difference between the ultrasonic signal at end B and the UHF electromagnetic wave signal at end B to obtain the second time difference; and calculate the arrival time difference between the ultrasonic signals at end A and end B to obtain the third time difference. The target velocity calculation module is used to determine the equivalent propagation velocity of the ultrasonic wave at end A based on the first time difference, determine the equivalent propagation velocity of the ultrasonic wave at end B based on the second time difference, and average the equivalent propagation velocities of the ultrasonic waves at end A and end B to obtain the target velocity. The initial discharge position calculation module is used to substitute the third time difference and the target velocity into... The initial discharge position is obtained; Where L is the distance from the power source to end A, and D is the total length of the cable from end A to end B. This represents the absolute value of the arrival time difference of the ultrasound signal between end A and end B. It is half the sum of the equivalent propagation velocities of the ultrasound at ends A and B.

9. The intelligent detection and positioning device for partial discharge of cables based on multimodal fusion according to claim 8, characterized in that, The target discharge location determination module includes: The particle swarm optimization module is used for formula-based optimization. The optimal value L is searched within a preset interval using a preset optimization algorithm. The specific operation of the algorithm is as follows: Initialize the differential evolution subpopulation and the particle swarm subpopulation; Differential evolutionary subpopulations are optimized iteratively through mutation, crossover, and selection; Particle swarm optimization updates particle positions using individual optimality and global optimality. If a smaller F(L) value than the current global optimum of the particle swarm is found during differential evolution, then the global guiding position of the particle swarm is replaced. The process is iterated alternately until the convergence condition is met, and the optimal L is output to obtain the target discharge position.

10. The intelligent detection and positioning device for partial discharge of cables based on multimodal fusion according to claim 6, characterized in that, The partial discharge type determination module includes: The dual-end time-series feature encoder module is used to substitute the enhanced feature groups of group A and group B multimodal signals into the dual-end time-series feature encoder to obtain the time-series feature vector; The physical prior embedding module is used to map the time difference value into the physical prior embedding module as a physical feature vector; A cross-modal fusion module is used to substitute the temporal feature vector and the physical feature vector into the cross-modal fusion module to perform feature vector fusion to obtain the target fused feature; The classification head module is used to input the target fusion features into the classification head to obtain the partial discharge type.

Images