Cable partial discharge multi-mode fusion defect detection method and system

Abstract

The invention relates to the technical field of cable discharge detection, in particular to a defect detection method and system for cable partial discharge multi-mode fusion. The method comprises the following steps: firstly, synchronously collecting analog signals of HFCT and UHF sensors on a cable, converting the analog signals into digital signals, and respectively carrying out hybrid noise reduction filtering processing to obtain an HFCT de-noised signal and a UHF de-noised signal; then, based on the HFCT de-noised signal and the UHF de-noised signal, a PRPD map and a PRPS map are generated respectively, feature fusion is carried out after alignment, and a joint feature vector is obtained; next, the combined feature vector is subjected to convolutional neural network and long-short-term memory network cooperative processing, accurate capture of weak discharge signals and fine classification of defect types are achieved, finally, the UHF signal arrival time difference and a cable topological graph are combined, high-precision defect positions are obtained, the defect types are determined, and the defect positions are positioned. Identification and positioning are integrated, the pertinence and efficiency of operation and maintenance troubleshooting can be greatly improved, blind inspection is avoided, and the operation and maintenance cost is reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of cable discharge detection, in particular to a defect detection method and system for multi-modal fusion of cable partial discharge. BACKGROUND

[0002] High-voltage cross-linked polyethylene (XLPE) cables are the lifelines of modern urban power grids. The joint parts are prone to faults due to on-site manufacturing processes and concentrated electric fields. Partial discharge (PD) is an important indicator of cable insulation degradation, so it is crucial to monitor the partial discharge of cable joints online.

[0003] The current mainstream detection methods are as follows.

[0004] High-frequency current method (HFCT): detect high-frequency current pulses on the ground wire through a clamp HFCT sensor. The advantages are that it can calibrate discharge quantity, is non-intrusive, and has low cost. The disadvantages are poor anti-interference ability, susceptible to noise in complex electromagnetic environments, and low positioning accuracy.

[0005] Ultra-high frequency method (UHF): detect 300MHz-3GHz electromagnetic wave signals radiated by partial discharge through a UHF sensor. The advantages are strong anti-interference ability, high sensitivity, and precise positioning. The disadvantages are high equipment cost, inability to directly calibrate discharge quantity, and large signal propagation attenuation.

[0006] The existing technology usually uses HFCT or UHF sensors alone, which has the following inherent defects.

[0007] Single technology limitations: HFCT is susceptible to interference, UHF has difficulty quantifying discharge intensity, and cannot comprehensively evaluate insulation status.

[0008] High misjudgment rate: it is difficult to distinguish between real partial discharge signals and external interference in a strong noise background, and false positives and false negatives are likely to occur. SUMMARY

[0009] The purpose of the present application is to provide a defect detection method and system for multi-modal fusion of cable partial discharge. The technical solution of the present application is as follows: A defect detection method for multi-modal fusion of cable partial discharge, comprising the following operations: S1, synchronously collect analog signals of HFCT sensors and UHF sensors deployed on the cable, convert them into digital signals, and obtain HFCT signals and UHF signals; The HFCT signals and UHF signals are respectively filtered based on hybrid noise reduction to obtain HFCT denoised signals and UHF denoised signals; S2, extracting pulse rise time, amplitude, energy from the HFCT denoising signal, extracting spectral centroid, signal intensity, time of arrival from the UHF denoising signal, generating PRPD and PRPS spectrum respectively; Based on the propagation time difference between the HFCT signal and the UHF signal, the PRPD and PRPS spectrum are aligned and then fused to obtain the joint feature vector; S3, after the joint feature vector is processed by the convolutional neural network and the long short-term memory network, the defect recognition is performed to obtain the defect category; Based on the UHF signal arrival time difference and the topology structure diagram composed of cable joints and branch points, the defect position is obtained, and the defect category is formed to form the defect detection result.

[0010] In S1, the operation of filtering based on hybrid denoising of the HFCT signal is: wavelet packet decomposition is performed on the HFCT signal to obtain multiple subbands; the energy of each subband is obtained, and the noise dominant subband is identified; signal suppression is performed on the noise dominant subband, and the remaining subbands are reconstructed by wavelet packet to obtain a preliminary denoising signal; the pure signal in the preliminary denoising signal is extracted to obtain the HFCT denoising signal.

[0011] The operation of extracting the pure signal in the preliminary denoising signal is: the preliminary denoising signal is converted into a time domain signal, and after being reshaped into a 2D feature map, normalization processing is performed to obtain a preliminary denoising feature map; the preliminary denoising feature map is processed by convolution and residual connection to obtain a low-level detail feature map; the low-level detail feature map is processed by spatial attention to obtain a spatial enhancement feature map; the spatial enhancement feature map is processed by multiple maximum pooling downsampling and spatial attention to obtain a global dependence enhancement feature map; the global dependence enhancement feature map is processed by transposed convolution and upsampling to obtain an upsampled feature map; the upsampled feature map and the global dependence enhancement feature map are processed by feature concatenation, and then by channel attention to obtain a channel calibration feature map; the channel calibration feature map is processed by multiple transposed convolution, upsampling and channel attention to extract a multi-scale channel calibration feature map; the multi-scale channel calibration feature map is bilinearly interpolated and upsampled to a unified size to obtain a multi-scale alignment feature map; the multi-scale alignment feature map is processed by multi-scale attention to obtain a global optimal fusion feature map; the global optimal fusion feature map is processed by convolution dimension reduction and Sigmoid activation to obtain a target pulse segmentation map; the target pulse segmentation map and the preliminary denoising signal are multiplied by pixel-level mask processing to obtain a pure target pulse signal as the HFCT denoising signal.

[0012] The signal suppression method can be realized by calculating and updating the subband coefficient based on the energy entropy ratio threshold and the subband coefficient.

[0013] The method for obtaining the propagation delay difference between HFCT and UHF signals is as follows: HFCT and UHF denoised pulse signals are processed by sampling rate unification and energy adaptive window truncation to obtain HFCT and UHF pulse segments; HFCT and UHF pulse segments are processed by fast Fourier transform and phase normalization, and then multi-band weighted to obtain the cross-power spectrum; the cross-power spectrum is processed by inverse fast Fourier transform and normalized to construct a cross-correlation function; within the effective delay range of the cross-correlation function, peak values ​​that satisfy the maximum value within the neighborhood window are extracted, and the peak value corresponding to the maximum value of the cross-correlation function among all peak values ​​with a confidence level greater than the confidence threshold is selected as the propagation delay difference.

[0014] The feature fusion operation is as follows: the phase distribution, amplitude statistics, and energy in the aligned PRPD spectrum, and the time series, amplitude time series, and arrival time in the aligned PRPS spectrum are used as evidence to calculate basic probability assignments; the basic probability assignments of all evidence are merged pairwise, and the maximum support criterion is used to determine the fused feature value; all fused feature values ​​and the time-frequency domain feature set are subjected to feature standardization and dimension concatenation to obtain a high-dimensional joint feature vector.

[0015] The operation to obtain the defect location is as follows: based on the arrival time difference of multiple UHF signals and the installation coordinates of the UHF sensor, calculate the preliminary spatial coordinates of the discharge point, and abstract the cable joint, branch point and their connection relationship into a graph structure; process the preliminary spatial coordinates and the cable topology map with a graph neural network to obtain the defect location.

[0016] A defect detection system for cable partial discharge multi-mode fusion, used to implement the above-mentioned defect detection method for cable partial discharge multi-mode fusion, includes: The signal acquisition and filtering module is used to synchronously acquire analog signals from HFCT and UHF sensors deployed on the cable, convert them into digital signals, and obtain HFCT and UHF signals. The HFCT and UHF signals are then filtered based on hybrid noise reduction to obtain HFCT denoised signals and UHF denoised signals, respectively. The signal fusion module is used to extract pulse rise time, amplitude, and energy from the HFCT denoised signal, and extract spectral centroid, signal strength, and arrival time from the UHF denoised signal, generating PRPD and PRPS maps respectively. Based on the propagation delay difference between the HFCT and UHF signals, the PRPD and PRPS maps are aligned and then fused to obtain a joint feature vector. The defect identification and localization module is used to process the joint feature vectors through a convolutional neural network and a long short-term memory network to identify defects and obtain defect categories. Based on the arrival time difference of UHF signals and the topological structure diagram composed of cable joints and branch points, the defect location and defect category are obtained, forming the defect detection result.

[0017] A defect detection device for multi-modal fusion of partial discharge in cables includes a processor and a memory, wherein the processor executes a computer program stored in the memory to implement the aforementioned defect detection method for multi-modal fusion of partial discharge in cables.

[0018] A computer-readable storage medium for storing a computer program, wherein the computer program, when executed by a processor, implements the aforementioned defect detection method for multimodal fusion of partial discharge in cables.

[0019] The beneficial effects of this invention are as follows: This invention provides a defect detection method for partial discharge in cables using multimodal fusion. First, analog signals from HFCT and UHF sensors on the cable are simultaneously acquired and converted into digital signals. These signals are then subjected to hybrid noise reduction filtering to reduce noise interference while preserving the core characteristics of the discharge pulse, resulting in denoised HFCT and UHF signals. Next, based on the denoised HFCT and UHF signals, PRPD and PRPS maps are generated respectively. After alignment based on their propagation delay difference, feature fusion is performed to obtain a joint feature vector that combines the advantages of calibrated discharge quantity from HFCT and high anti-interference capability from UHF. Then, the joint feature vector is processed collaboratively using a convolutional neural network and a long short-term memory network to achieve accurate capture of weak discharge signals and fine classification of defect types, significantly reducing false alarm and false negative rates. Finally, by combining the arrival time difference of the UHF signal with the cable topology map, a high-precision defect location and a complete detection result formed by the defect category are obtained. This method clearly identifies both the defect type and location, achieving integrated identification and location, which can significantly improve the targeting and efficiency of maintenance fault diagnosis, avoid blind inspections, and reduce maintenance costs. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the exemplary embodiments of this application clearer, the technical solutions in the exemplary embodiments of this application are described clearly and completely below. Obviously, the described exemplary embodiments are only some embodiments of this application, and not all embodiments.

[0021] Based on the exemplary embodiments shown in this application, all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of this application. Furthermore, although the disclosures in this application are presented by way of one or more exemplary examples, it should be understood that each aspect of these disclosures can constitute a complete technical solution on its own.

[0022] Furthermore, the terms “comprising” and “having”, and any variations thereof, are intended to cover but not exclusively include, for example, a product or device that includes a series of components is not necessarily limited to those that are explicitly listed, but may include other components that are not explicitly listed or that are inherent to such product or device.

[0023] As used in this application, the term "module" means any known or subsequently developed hardware, software, firmware, artificial intelligence, fuzzy logic, or combination of hardware and / or software code capable of performing the functions associated with that element.

[0024] Example 1 This embodiment provides a defect detection method for partial discharge in cables using multi-mode fusion, including the following operations: S1. Synchronously acquire analog signals from HFCT and UHF sensors deployed on the cable, convert them into digital signals, and obtain HFCT and UHF signals; The HFCT signal and the UHF signal are respectively processed by filtering based on hybrid noise reduction to obtain the HFCT denoised signal and the UHF denoised signal; S2. Extract pulse rise time, amplitude, and energy from the HFCT denoised signal, and extract spectral centroid, signal strength, and arrival time from the UHF denoised signal to generate PRPD and PRPS spectra, respectively. Based on the propagation delay difference between HFCT and UHF signals, the PRPD and PRPS spectra are aligned and then fused to obtain a joint feature vector. S3. After the joint feature vector is processed by a convolutional neural network and a long short-term memory network, defect identification is performed to obtain the defect category; Based on the arrival time difference of UHF signals and the topology diagram consisting of cable joints and branch points, the defect location and defect category are obtained, forming the defect detection result.

[0025] S1. Simultaneously acquire analog signals from HFCT and UHF sensors deployed on the cable, convert them into digital signals to obtain HFCT and UHF signals; HFCT and UHF signals are respectively processed by filtering based on hybrid noise reduction to obtain HFCT denoised signals and UHF denoised signals.

[0026] Synchronously acquiring analog signals from HFCT and UHF sensors on the cable and converting them into digital signals ensures precise alignment of the time axes of the two signals, providing a time consistency basis for subsequent cross-signal analysis such as time delay difference calculation. Performing hybrid noise reduction filtering on both signals can take into account the removal effect of different types of noise, reducing noise interference while preserving the core characteristics of the discharge pulse. The resulting HFCT denoised signal and UHF denoised signal can significantly improve the accuracy and reliability of subsequent defect identification.

[0027] First, at the 500kV XLPE cable joint, clamp at least one HFCT sensor to the grounding lead, and fix at least one UHF sensor to the joint housing using magnetic attraction. The sensor spacing is optimized according to the electromagnetic wave wavelength (e.g., approximately 25cm at 300MHz).

[0028] The GPS-disciplined rubidium clock provides a unified time base for all data acquisition cards, acquiring analog signals (high-frequency current pulse signals) in the 3-80MHz frequency band at a sampling rate of 2.5GS / s, converting them into digital signals to obtain HFCT signals. Simultaneously, it acquires analog signals (ultra-high frequency electromagnetic wave signals) in the 300MHz-1.5GHz frequency band, converting them into digital signals to obtain UHF signals.

[0029] Then, the HFCT signal and the UHF signal are respectively processed by filtering based on hybrid denoising to obtain the HFCT denoised signal and the UHF denoised signal.

[0030] Taking the HFCT signal as an example, the specific steps of the HFCT signal filtering process based on hybrid noise reduction are as follows.

[0031] Step 1: Perform wavelet packet decomposition on the HFCT signal to obtain multiple sub-bands.

[0032] Step 2: Obtain the energy of each subband and identify the noise-dominant subband.

[0033] Among them, the energy entropy ratio of the noise-dominant subband is less than the energy entropy ratio threshold. The energy entropy ratio and the energy entropy ratio threshold are calculated using the following formulas: , , , , For the first j Layer k Energy entropy ratio of each subband, subband energy, energy entropy ratio threshold p For the first j Layer k coefficient of each sub-band The normalized probability distribution The standard deviation of all sub-band coefficients. For the length of the coefficient, This is an adaptive adjustment factor.

[0034] Step 3: Apply threshold processing to suppress the signal in the noise-dominant subband.

[0035] Signal suppression can be achieved by calculating and updating the subband coefficients based on the energy entropy ratio threshold and the subband coefficients (by applying a soft threshold to the coefficients). The calculation formula is as follows: , For the first j Layer k The update coefficient of each sub-band For the first j Layer k The coefficient of each sub-band.

[0036] Step 4: The remaining subbands are reconstructed using wavelet packets to obtain the preliminary denoised signal.

[0037] Step 5: Extract the clean signal from the initial denoised signal to obtain the HFCT denoised signal.

[0038] The specific steps for extracting the clean signal from the initial denoised signal are as follows.

[0039] Step 5.1: Convert the initial denoised signal into a time-domain signal, reshape it into a 2D feature map using a sliding window, and then perform normalization to eliminate dimensional differences, thus obtaining the initial denoised feature map.

[0040] Step 5.2: The initial denoised feature map is processed by convolution and residual connection to obtain a low-level detail feature map; the low-level detail feature map is processed by spatial attention to obtain a spatially enhanced feature map.

[0041] Step 5.3: The spatial augmented feature map is subjected to multiple max pooling downsampling and spatial attention processing to obtain the global dependency augmented feature map; the global dependency augmented feature map is subjected to transposed convolution and upsampling processing to obtain the upsampled feature map; the upsampled feature map and the global dependency augmented feature map are concatenated and then subjected to channel attention processing to calibrate redundant features to obtain the channel calibration feature map.

[0042] Step 5.4: The channel calibration feature map is subjected to multiple transposed convolutions, upsampling, and channel attention processing to improve the distinction between the target pulse and the background, and a multi-scale channel calibration feature map is extracted. The multi-scale channel calibration feature map is then upsampled to a uniform size through bilinear interpolation to obtain a multi-scale aligned feature map.

[0043] Step 5.5: The multi-scale aligned feature map is processed by multi-scale attention to obtain the globally optimal fused feature map; the globally optimal fused feature map is processed by convolutional dimensionality reduction and Sigmoid activation to obtain the target pulse segmentation map; the target pulse segmentation map and the preliminary denoised signal are multiplied by a pixel-level mask to obtain the pure target pulse signal, which is used as the HFCT denoised signal.

[0044] S2. Extract pulse rise time, amplitude, and energy from the HFCT denoised signal, and extract spectral centroid, signal strength, and arrival time from the UHF denoised signal to generate PRPD and PRPS maps respectively. Based on the propagation delay difference between the HFCT and UHF signals, align the PRPD and PRPS maps and perform feature fusion to obtain a joint feature vector.

[0045] Based on the HFCT denoised signal and the UHF denoised signal, PRPD and PRPS spectra are generated respectively to comprehensively capture and visualize the multi-dimensional characteristics of the discharge. The spectra are aligned based on the propagation delay difference between the two to ensure consistency in the time dimension, laying a solid foundation for cross-signal fusion. Then, by combining HFCT and UHF, the joint feature vector obtained combines the advantages of HFCT's ability to calibrate the discharge quantity and UHF's high anti-interference capability, overcoming the limitations of single technology and improving the accuracy and reliability of subsequent tasks such as defect identification and type diagnosis.

[0046] First, pulse rise time, amplitude, and energy are extracted from the HFCT denoised signal, and spectral centroid, signal strength, and arrival time are extracted from the UHF denoised signal, generating PRPD and PRPS spectra respectively.

[0047] Then, based on the propagation delay difference between the HFCT signal and the UHF signal.

[0048] The specific steps for obtaining the propagation delay difference are as follows.

[0049] Step 1: The HFCT denoised pulse signal and the UHF denoised pulse signal are processed by sampling rate unification and energy adaptive window truncation to avoid time delay calculation deviation, resulting in synchronized HFCT pulse segments and UHF pulse segments.

[0050] Sampling rate unification can be achieved by interpolating the low-sampling-rate HFCT signal (using a cubic spline interpolation function) to the UHF signal sampling rate. Energy adaptive window truncation can be achieved by truncating segments whose local energy is greater than a local energy threshold.

[0051] Step 2: After fast Fourier transform and phase normalization, the HFCT pulse segments and UHF pulse segments are weighted by multi-band weighting to enhance the phase information of the high signal-to-noise ratio band and suppress noise interference in the low signal-to-noise ratio band, thus obtaining the cross-power spectrum.

[0052] Step 3: After the cross-power spectrum undergoes inverse fast Fourier transform, it is normalized to construct the cross-correlation function, which maps the frequency domain phase information back to the time domain. The normalized cross-correlation function has a sharper peak, which can reduce spurious time delays caused by sidelobe interference.

[0053] Step 4: Within the effective time delay range of the cross-correlation function, extract the peak values ​​that satisfy the maximum value within the neighborhood window, and select the peak value corresponding to the maximum value of the cross-correlation function among all peak values ​​with a confidence level greater than the confidence threshold, as the propagation time delay difference.

[0054] Finally, after aligning the PRPD and PRPS maps, feature fusion is performed to obtain a joint feature vector.

[0055] The feature fusion operation is as follows: the phase distribution, amplitude statistics, and energy in the aligned PRPD spectrum, and the time series, amplitude time series, and arrival time in the aligned PRPS spectrum are used as evidence to calculate basic probability assignments; according to Dempster's combination rule, the basic probability assignments of all evidence are combined pairwise, and the maximum support criterion is used to determine the fused feature value; all fused feature values ​​and the time-frequency domain feature set are subjected to feature standardization and dimension concatenation to obtain a high-dimensional joint feature vector.

[0056] The aforementioned time-frequency domain feature set mainly consists of key time-frequency domain features of HFCT and UHF signals, including the rise time and / or energy of HFCT signals, and the spectral centroid of UHF signals.

[0057] S3. After the joint feature vector is processed by a convolutional neural network and a long short-term memory network, defect identification is performed to obtain the defect category. Based on the arrival time difference of the UHF signal and the topological structure diagram composed of cable joints and branch points, the defect location and defect category are obtained to form the defect detection result.

[0058] The joint feature vector is processed collaboratively by a convolutional neural network and a long short-term memory network to fully mine multi-dimensional information about the discharge and accurately identify the defect category. This solves the problem of incomplete feature extraction by a single network, enabling accurate capture of weak discharge signals and fine classification of defect types, significantly reducing false alarm and false negative rates. Furthermore, by combining the high-precision defect location obtained from the arrival time difference of the UHF signal and the cable topology map with the defect category, a complete detection result is formed, which not only clarifies the defect type but also locates the defect location, achieving integrated identification and location. This greatly improves the targeting and efficiency of troubleshooting during operation and maintenance, avoids blind inspections, and reduces operation and maintenance costs.

[0059] The joint feature vector is processed by a convolutional neural network to extract local patterns (such as edges and textures), and then processed by a long short-term memory network to capture temporal dynamics (such as the periodicity and trend of pulse sequences). After defect identification (including but not limited to classifiers or softmax layers for multi-classification), the defect category is obtained.

[0060] Meanwhile, based on the UHF signal arrival time difference and the topology graph neural network model, the defect location and defect category are obtained, forming the defect detection result.

[0061] The specific steps to obtain the defect location are as follows: Based on the Time Difference of Arrival (TDOA) of multiple UHF signals and the installation coordinates of the UHF sensors, the preliminary spatial coordinates of the discharge point are calculated. Key components such as cable joints and branch points, as well as their connection relationships, are abstracted into a graph structure (nodes and edges). The preliminary coordinates are then processed by a graph neural network (GNN) with the cable topology graph. Through a message passing mechanism, the GNN integrates geometric positioning information and topological constraints to map and optimize the preliminary coordinates to the most reasonable actual location node on the topology graph (such as a joint). Finally, a positioning result that completely corresponds to the physical structure of the cable is output, thus obtaining the defect location.

[0062] Finally, the location and type of defect together form the defect detection result.

[0063] Furthermore, a graphical software interface is constructed to display signal waveforms, PRPD / PRPS graphs, defect types, and location results in real time. When a defect is detected, the system automatically issues the defect detection result.

[0064] This embodiment also provides a defect detection system for multi-modal fusion of partial discharge in cables, used to implement the above-mentioned defect detection method for multi-modal fusion of partial discharge in cables, including: The sensing module includes at least one HFCT sensor and at least one UHF sensor; The signal acquisition and filtering module is used to synchronously acquire analog signals from HFCT and UHF sensors deployed on the cable, convert them into digital signals, and obtain HFCT and UHF signals. The HFCT and UHF signals are then filtered based on hybrid noise reduction to obtain HFCT denoised signals and UHF denoised signals, respectively. The signal fusion module is used to extract pulse rise time, amplitude, and energy from the HFCT denoised signal, and extract spectral centroid, signal strength, and arrival time from the UHF denoised signal, generating PRPD and PRPS maps respectively. Based on the propagation delay difference between the HFCT and UHF signals, the PRPD and PRPS maps are aligned and then fused to obtain a joint feature vector. The defect identification and localization module is used to identify defects and obtain defect categories after the joint feature vectors are processed by convolutional neural networks and long short-term memory networks. Based on the arrival time difference of UHF signals and the topology diagram composed of cable joints and branch points, the defect location and defect category are obtained, forming the defect detection result. The early warning module is used to build a graphical software interface to display signal waveforms, PRPD / PRPS spectra, defect types, and location results in real time; when a defect is detected, the module automatically issues the defect detection results.

[0065] This embodiment also provides a defect detection device for multi-modal fusion of partial discharge in cables, including a processor and a memory, wherein the processor executes a computer program stored in the memory to implement the above-mentioned defect detection method for multi-modal fusion of partial discharge in cables.

[0066] This embodiment also provides a computer-readable storage medium for storing a computer program, wherein the computer program, when executed by a processor, implements the above-described defect detection method for multi-modal fusion of partial discharge in cables.

[0067] This embodiment provides a defect detection method for partial discharge in cables using multimodal fusion. First, analog signals from HFCT and UHF sensors on the cable are simultaneously acquired and converted into digital signals. These signals are then subjected to hybrid noise reduction filtering to reduce noise interference while preserving the core characteristics of the discharge pulse, resulting in denoised HFCT and UHF signals. Next, based on the denoised HFCT and UHF signals, PRPD and PRPS maps are generated respectively. After alignment based on their propagation delay difference, feature fusion is performed to obtain a joint feature vector that combines the advantages of calibrated discharge quantity from HFCT and high anti-interference capability from UHF. Then, the joint feature vector is processed collaboratively by a convolutional neural network and a long short-term memory network to achieve accurate capture of weak discharge signals and fine classification of defect types, significantly reducing false alarm and false negative rates. Finally, by combining the arrival time difference of the UHF signal with the cable topology map, a high-precision defect location and a complete detection result formed by the defect category are obtained. This method clearly identifies both the defect type and location, achieving integrated identification and location, which can significantly improve the targeting and efficiency of maintenance fault diagnosis, avoid blind inspections, and reduce maintenance costs.

[0068] This embodiment provides a defect detection method for partial discharge of cables using multi-mode fusion. By combining HFCT and UHF, it combines the advantages of HFCT's ability to calibrate discharge quantity and UHF's high anti-interference capability, overcoming the limitations of single technologies, significantly improving detection sensitivity, and still working stably under strong background noise.

[0069] This embodiment provides a defect detection method for partial discharge in cables using multimodal fusion. By employing multimodal fusion and deep learning algorithms, it achieves accurate capture of weak discharge signals and fine classification of defect types, significantly reducing false alarm rate and false negative rate.

[0070] This embodiment provides a defect detection method for partial discharge in cables using multi-modal fusion. The model can be updated online through a federated learning framework, adapting to different cable structures and operating environments, and possesses good versatility and scalability.

[0071] This embodiment provides a defect detection system for partial discharge of cables with multi-modal fusion. The innovative early warning module enables continuous and quantitative assessment of the cable insulation status, supports predictive maintenance, and avoids sudden failures.

[0072] While exemplary embodiments of the invention have been described herein, many other variations or modifications conforming to the principles of the invention can be directly determined or derived from the disclosure of this invention without departing from its spirit and scope. Therefore, the scope of the invention should be understood and recognized to cover all such other variations or modifications.

Claims

1. A method for defect detection of partial discharge multi-modal fusion of a cable, characterized in that, The method comprises the following operations: S1, synchronously collecting analog signals of HFCT sensors and UHF sensors arranged on a cable, converting the analog signals into digital signals to obtain HFCT signals and UHF signals; The HFCT signals and the UHF signals are subjected to filtering processing based on hybrid noise reduction to obtain HFCT denoised signals and UHF denoised signals; S2, pulse rise time, amplitude and energy are extracted from the HFCT denoised signals, and spectral centroid, signal strength and arrival time are extracted from the UHF denoised signals to generate PRPD and PRPS graphs respectively; Based on the propagation time delay difference between the HFCT signals and the UHF signals, the PRPD graph and the PRPS graph are aligned and then subjected to feature fusion to obtain a joint feature vector; S3, the joint feature vector is subjected to convolutional neural network and long short-term memory network processing, and then subjected to defect recognition to obtain a defect category; Based on the UHF signal arrival time difference and a topological structure diagram formed by cable joints and branch points, a defect position is obtained, and the defect position and the defect category form a defect detection result.

2. The method for partial discharge multi-modal fusion based defect detection of electrical cables as claimed in claim 1 wherein, In S1, the operation of filtering processing of the HFCT signals based on hybrid noise reduction is as follows: The HFCT signals are subjected to wavelet packet decomposition to obtain a plurality of subbands; the energy of each subband is obtained, and a noise dominant subband is identified; The noise dominant subband is subjected to signal suppression, and the remaining subbands are subjected to wavelet packet reconstruction to obtain a preliminary denoised signal; a pure signal in the preliminary denoised signal is extracted to obtain the HFCT denoised signals.

3. The method for partial discharge multi-modal fusion based defect detection of electrical cables as claimed in claim 2, wherein, The operation of extracting the pure signal in the preliminary denoised signal is as follows: The preliminary denoised signal is converted into a time domain signal, and after being reshaped into a 2D feature map, normalization processing is performed to obtain a preliminary denoised feature map; the preliminary denoised feature map is subjected to convolution and residual connection processing to obtain a low-layer detail feature map; the low-layer detail feature map is subjected to spatial attention processing to obtain a spatial enhancement feature map; The spatial enhancement feature map is subjected to multiple maximum pooling downsampling and spatial attention processing to obtain a global dependence enhancement feature map; The global dependence enhancement feature map is subjected to transposed convolution and upsampling processing to obtain an upsampled feature map; The upsampled feature map and the global dependence enhancement feature map are subjected to feature splicing and then channel attention processing to obtain a channel calibration feature map; The channel calibration feature map is subjected to multiple transposed convolution, upsampling and channel attention processing to extract a multi-scale channel calibration feature map; The multi-scale channel calibration feature map is subjected to bilinear interpolation upsampling to a unified size to obtain a multi-scale alignment feature map; The multi-scale alignment feature map is subjected to multi-scale attention processing to obtain a global optimal fusion feature map; The global optimal fusion feature map is subjected to convolution dimension reduction and Sigmoid activation processing to obtain a target pulse segmentation map; the target pulse segmentation map and the preliminary denoised signal are subjected to pixel-level mask multiplication processing to obtain a pure target pulse signal as the HFCT denoised signal.

4. The method for partial discharge multi-modal fusion based defect detection of a cable as claimed in claim 2, wherein, The signal suppression method can be realized by calculating and updating the subband coefficients based on the energy entropy ratio threshold and the subband coefficients.

5. The method for partial discharge multi-modal fusion based defect detection of electrical cables as claimed in claim 1 wherein, In S2, the propagation time delay difference between the HFCT signals and the UHF signals is obtained by The HFCT de-noised pulse signal and the UHF de-noised pulse signal are subjected to sampling rate unification and energy adaptive window interception processing to obtain HFCT pulse segments and UHF pulse segments; The HFCT pulse segments and the UHF pulse segments are subjected to fast Fourier transform and phase normalization processing, and then subjected to multi-band weight weighting to obtain cross power spectrum; The cross power spectrum is subjected to inverse fast Fourier transform and normalization processing to construct a cross-correlation function; in the effective time delay range of the cross-correlation function, a peak value satisfying the maximum value in the neighborhood window is extracted, and the peak value corresponding to the maximum value of the cross-correlation function value is selected from all peak values with a confidence degree greater than a confidence degree threshold as the peak value of the propagation time delay difference.

6. The method for partial discharge multi-modal fusion based defect detection of electrical cables as claimed in claim 1 wherein, In S2, the operation of feature fusion is: The phase distribution, amplitude statistics and energy in the aligned PRPD spectrum, and the time sequence, amplitude time sequence and arrival time in the aligned PRPS spectrum are taken as evidences respectively to calculate basic probability assignments; The basic probability assignments of all evidences are merged two by two, and the fusion feature values are determined by using the maximum support criterion; the fusion feature values and the time-frequency domain feature set are subjected to feature standardization and dimension splicing processing to obtain a high-dimensional joint feature vector.

7. The method for partial discharge multi-modal fusion based defect detection of electrical cables as claimed in claim 1 wherein, In S3, the operation of obtaining the defect position is: Based on the arrival time difference of the multi-path UHF signal and the installation coordinates of the UHF sensor, the preliminary spatial coordinates of the discharge point are calculated, the cable joint, branch point and their connection relationship are abstracted into a graph structure; the preliminary spatial coordinates and the cable topology graph are subjected to graph neural network processing to obtain the defect position.

8. A system for partial discharge multi-modal fusion based defect detection of a cable for implementing the method for partial discharge multi-modal fusion based defect detection of a cable as claimed in claim 1, wherein, It comprises: A signal acquisition and filtering module is configured to synchronously acquire analog signals of HFCT sensors and UHF sensors deployed on a cable, convert the analog signals into digital signals, and obtain HFCT signals and UHF signals; the HFCT signals and the UHF signals are subjected to filtering processing based on hybrid noise reduction to obtain HFCT de-noised signals and UHF de-noised signals; A signal fusion module is configured to extract pulse rise time, amplitude and energy from the HFCT de-noised signals, extract spectral centroid, signal strength and arrival time from the UHF de-noised signals, generate PRPD spectrum and PRPS spectrum respectively, perform feature fusion on the PRPD spectrum and the PRPS spectrum after alignment based on the propagation time delay difference between the HFCT signals and the UHF signals, and obtain a joint feature vector; A defect recognition and positioning module is configured to perform defect recognition on the joint feature vector after convolutional neural network and long short-term memory network processing, obtain a defect category, obtain a defect position based on the arrival time difference of the UHF signal and a topology graph composed of cable joints and branch points, and form a defect detection result together with the defect category.

9. A partial discharge multi-modal fusion defect detection device for a cable, characterized in that, It comprises a processor and a memory, wherein the processor implements the cable partial discharge multi-modal fusion defect detection method of any one of claims 1-7 when executing a computer program stored in the memory.

10. A computer-readable storage medium, characterized in that, A computer program is stored, wherein the computer program is executed by a processor to implement the cable partial discharge multi-modal fusion defect detection method of any one of claims 1-7.