A cable sheath comprehensive online monitoring method and system

Abstract

The present application relates to the technical field of electrical sensor, solve the technical problems of lacking physical constraint, the diagnosis result may violate the basic law of thermodynamics, sensitive to the distribution change of training data, and the monitoring accuracy is limited, especially relates to a kind of cable sheath comprehensive online monitoring method and system, the steps of this method are: collecting the original electrical signal of target cable, according to original electrical signal through filtering time stamp encapsulation processing, obtain synchronous data frame, the present application is based on graph neural network simultaneously using cable topological graph structure, provides the distribution input of joule heat source containing topological dependence, can greatly improve the identification accuracy, form the linkage effect of spatial topology and heat field propagation, further optimize fault location precision, avoid unnecessary excavation repair, save cost, and the diagnosis result obtained can directly guide operation and maintenance decision, greatly improve operation and maintenance work efficiency.

Description

Technical Field

[0001] This invention relates to the field of electrical sensor technology, and in particular to a method and system for comprehensive online monitoring of cable sheaths. Background Technology

[0002] Electrical sensors are devices that convert physical quantities such as cable sheath grounding current and voltage into measurable electrical signals. Current technologies for comprehensive online monitoring of cable sheaths using electrical sensors typically involve collecting sheath grounding current and reference voltage signals, then processing the signals through characteristic parameters, determining whether electrical characteristic quantities such as current are abnormal, and finally performing diagnostic monitoring of the cable based on the abnormality of the electrical characteristic quantities.

[0003] Existing technologies use purely data-driven neural network models for classification. While they can learn statistical patterns in the data, they lack physical constraints. The diagnostic results may violate the basic laws of thermodynamics and are sensitive to changes in the distribution of training data. They also have poor generalization ability and limited accuracy in new, unseen fault modes. For example, if a cable joint experiences localized overheating due to poor installation, the initial sheath current anomaly index may be only 0.2, which existing technologies may classify as normal. However, there may be slight temperature differences between adjacent monitoring points. Because existing technologies lack physical constraints, they may easily overlook abnormalities in temperature or other physical quantities, leading to unexpected cable accidents. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a comprehensive online monitoring method and system for cable sheaths, which solves the technical problems of existing technologies lacking physical constraints, diagnostic results potentially violating fundamental thermodynamic laws, being sensitive to changes in training data distribution, and having limited monitoring accuracy.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a method for comprehensive online monitoring of cable sheath, the method comprising the following steps: acquiring the original electrical signal of the target cable, and processing the original electrical signal by filtering and timestamp encapsulation to obtain a synchronous data frame; Based on the synchronous data frames, fundamental wave energy spectrum fusion processing is performed to obtain feature parameter vectors; High-frequency energy screening is performed on the feature parameter vector to obtain the dominant mode reconstruction signal. Based on the dominant mode reconstruction signal, an enhanced discharge feature sequence is generated through adaptive sequence integration processing. Based on the feature parameter vector, the core mode feature sequence is obtained through time-series dimensionality reduction. The core mode feature sequence is then subjected to dynamic time-limited output processing to obtain the sheath current anomaly index. Based on the feature parameter vector, the discharge feature enhancement sequence and the sheath current anomaly index, a fusion network coupling process is performed to obtain the physical constraint enhancement feature. The physical constraint enhancement feature is then subjected to fault probability diagnosis processing to generate a comprehensive diagnosis result and an early warning signal.

[0006] Preferably, the original electrical signal is encapsulated and processed using a filtered timestamp, including: The original electrical signal is bandpass filtered to generate a filtered and conditioned signal. Analog-to-digital conversion is performed on the filtered and conditioned signal to obtain a digital signal stream; The digital signal stream is timestamped to obtain a synchronized data frame.

[0007] Preferably, fundamental frequency energy spectrum fusion processing based on synchronous data frames includes: The fundamental amplitude and phase are obtained by processing the fundamental component of the synchronous data frame; Transient energy spectrum processing is performed based on synchronous data frames to generate frequency band energy distribution; The fundamental amplitude, phase, and frequency band energy distribution are fused to obtain a feature parameter vector.

[0008] Preferably, high-frequency energy filtering is performed on the feature parameter vector, including: Based on the feature parameter vector, a high-frequency reconstructed time-domain signal is obtained through high-frequency feature extraction. Variational mode decomposition is performed on high-frequency reconstructed time-domain signals to generate an eigenmode function set. Modal energy screening is performed on the intrinsic mode function set to obtain the dominant mode reconstruction signal.

[0009] Preferably, the adaptive sequence integration processing based on the dominant mode reconstruction signal includes: Based on the dominant mode reconstruction signal, an adaptive thresholding process is used to obtain a local adaptive threshold curve; Pulse over-threshold detection is performed based on local adaptive threshold curves and dominant mode reconstruction signals to obtain a set of candidate pulse segments; The candidate pulse segment set is processed by pulse sequence integration to generate a discharge feature enhancement sequence.

[0010] Preferably, the feature parameter vector is subjected to temporal dimensionality reduction processing, including: Based on the feature parameter vector, a current timing window matrix is ​​generated by constructing and processing the timing window. Multi-head attention feature processing is performed based on the current time-series window matrix to generate a weighted feature sequence; Principal component dimensionality reduction is performed on the weighted feature sequence to obtain the core pattern feature sequence.

[0011] Preferably, the core pattern feature sequence is subjected to dynamic time-limited output processing, including: The minimum regular distance value is obtained by processing the core pattern feature sequence through dynamic temporal similarity. Anomaly index normalization is performed based on minimum regularity distance value to generate normalized anomaly index; The normalized anomaly index is subjected to amplitude limiting output processing to obtain the sheath current anomaly index.

[0012] Preferably, the fusion network coupling process based on feature parameter vectors, discharge feature enhancement sequences, and sheath current anomaly index includes: Based on the feature parameter vector, the discharge feature enhancement sequence, and the sheath current anomaly index, a spatiotemporally aligned data tensor is generated through alignment and fusion processing. Graph convolutional network processing is performed based on spatiotemporally aligned data tensors to obtain spatial feature embedding values; Physically constrained enhanced features are obtained by coupling the spatial feature embedding values ​​with a physical neural network.

[0013] Preferably, fault probability diagnosis processing is performed on the physical constraint enhancement features, including: Based on the enhanced physical constraints, the failure probability inference process is used to obtain the failure type probability distribution; The severity of a fault is assessed based on the sheath current anomaly index, the enhanced discharge characteristic sequence, and the probability distribution of fault types, resulting in a fault severity level. The diagnostic results are processed based on the probability distribution of fault types and the severity level of faults to generate comprehensive diagnostic results and early warning signals.

[0014] This technical solution also provides a comprehensive online monitoring system for cable sheaths, which includes: The synchronization module is used to acquire the raw electrical signals of the target cable, and to obtain synchronization data frames by encapsulating and processing the raw electrical signals through filtering and timestamps. The feature parameter module is used to perform fundamental wave energy spectrum fusion processing based on synchronous data frames to obtain feature parameter vectors; The discharge feature module is used to perform high-frequency energy screening on the feature parameter vector to obtain the dominant mode reconstruction signal. Based on the dominant mode reconstruction signal, an adaptive sequence integration process is used to generate a discharge feature enhancement sequence. The sheath current module is used to obtain the core mode feature sequence by time-series dimensionality reduction based on the feature parameter vector, and to perform dynamic time-limited output processing on the core mode feature sequence to obtain the sheath current anomaly index. The comprehensive diagnostic module is used to perform fusion network coupling processing based on feature parameter vectors, discharge feature enhancement sequences and sheath current anomaly index to obtain physical constraint enhancement features, perform fault probability diagnostic processing on the physical constraint enhancement features, and generate comprehensive diagnostic results and early warning signals.

[0015] By employing the above technical solution, the present invention provides a method and system for comprehensive online monitoring of cable sheaths, which has at least the following beneficial effects: 1. This invention employs Discrete Fourier Transform to simultaneously extract the effective value and phase difference of the sheath current, providing more comprehensive steady-state characteristics at power frequency. The state discrimination of the grounding system based on phase is more accurate. Wavelet packet decomposition technology is used to decompose the signal into multiple continuous sub-bands and calculate the energy of each band, realizing refined quantization of the transient signal spectral structure. It can effectively distinguish the energy distribution characteristics of different frequency bands. The feature splicing method is used to integrate the fundamental amplitude and phase with the frequency band energy distribution into a one-dimensional feature parameter vector, providing a unified and comprehensive feature input for subsequent analysis.

[0016] 2. This invention can adaptively decompose non-stationary high-frequency reconstructed time-domain signals into multiple intrinsic mode functions. This not only achieves coarse extraction of discharge pulses but also provides a significantly improved signal-to-noise ratio input signal for subsequent adaptive threshold calculation. It solves the problem of missed or false alarms in the existing fixed threshold method. The adaptive threshold generation adopts a sliding window, which ensures that real discharge events can still be accurately identified in complex interference environments. Variational mode decomposition makes the local standard deviation more accurately reflect the noise level during subsequent adaptive threshold calculation, thus making the threshold setting more precise. Conversely, the adaptive threshold can further eliminate residual interference that was not completely filtered out by variational mode decomposition, forming a double-screening result.

[0017] 3. By calculating the correlation strength between different times, this invention not only effectively enhances the temporal features but also provides a core mode feature sequence with a higher signal-to-noise ratio. Principal component analysis is used to compress the weighted feature sequence into the core mode feature space, retaining the main changes in the data while removing redundancy and noise. Through normalization and amplitude limiting, the minimum regularization distance is converted into a sheath current anomaly index in the zero-to-one interval, which greatly improves the sensitivity to phase difference changes. This enables the accurate capture of slow and small fault features of the sheath current, providing a highly reliable anomaly quantification index for subsequent multi-physics coupled fault diagnosis.

[0018] 4. This invention, based on a graph neural network and utilizing the cable topology graph structure, provides Joule heat source distribution input containing topological dependencies, which can significantly improve identification accuracy. By embedding the thermoelectric coupling physical equation as prior knowledge into the network training process through the physical information neural network, it is not only supervised by fault labels but also needs to satisfy the residual constraints of the heat conduction equation. The spatial feature embedding provided by the graph neural network injects topological dependencies into the physical information neural network, enabling the thermal field calculation to propagate along the actual cable structure, forming a linkage effect between spatial topology and thermal field propagation. This further optimizes fault location accuracy, avoids unnecessary excavation and maintenance, saves costs, and the diagnostic results obtained can directly guide operation and maintenance decisions, greatly improving the efficiency of operation and maintenance work. Attached Figure Description

[0019] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments of this application and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of a comprehensive online monitoring method for cable sheaths according to the present invention; Figure 2 This is a structural block diagram of a comprehensive online monitoring system for cable sheaths according to the present invention. Detailed Implementation

[0020] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. This will allow for a full understanding of how the present application uses technical means to solve technical problems and achieve technical effects, and to facilitate its implementation.

[0021] Example 1: Due to the lack of physical constraints in current technologies, diagnostic results may violate fundamental laws of thermodynamics. Furthermore, these technologies are sensitive to changes in training data distribution, limiting monitoring accuracy. Please refer to [the relevant documentation / reference]. Figure 1 This embodiment provides a comprehensive online monitoring method for cable sheaths, which can comprehensively and accurately monitor cable conditions through pure data analysis combined with physical constraints, thereby improving monitoring accuracy and preventing cable accidents. The method includes the following steps: S1. Collect the original electrical signal of the target cable, and encapsulate and process the original electrical signal through filtering and timestamp to obtain a synchronous data frame. Existing technologies use simple low-pass or high-pass filtering, which is difficult to process the complex interference spectrum in cable sheath monitoring. This results in incomplete separation of effective signal and noise, lack of a unified clock reference for independent sampling, and distortion of the phase relationship of multi-channel signals, which directly affects the accuracy of fault feature extraction. Furthermore, the lack of a unified time stamp system makes it difficult to achieve synchronous fusion analysis of multi-source data. To solve the above problems, the specific implementation steps are as follows: S11. Bandpass filtering is performed on the original electrical signal to generate a filtered and conditioned signal. In this step, the input original electrical signal is first sampled in the time domain, and the instantaneous amplitude of each sampling point is used as the basic data. Then, the time domain signal is converted into a frequency domain representation through mathematical transformation. In the frequency domain, frequency components falling outside the passband range formed by the preset lower and upper frequency limits are suppressed. Specifically, the amplitude of each frequency point is multiplied by a bandpass filter operator. The operator takes a value close to 1 in the passband and a value close to 0 in the stopband, thereby achieving selective passage of signals in specific frequency bands. After completing the frequency domain filtering, the processed frequency domain data is restored to the time domain waveform through inverse transformation. In this process, each data point undergoes the above transformation and multiplication operations. Finally, all processed data points are combined to form a preliminary processed electrical signal. The original electrical signal includes the power frequency fundamental component, harmonic components, high-frequency transient components, and the instantaneous value of the three-phase voltage on the cable bus or line side.

[0022] S12. Based on the filtered and conditioned signal, perform analog-to-digital conversion to obtain a digital signal stream. In this step, the time interval is first determined according to the system's preset sampling rate. The sampling rate determines the number of times the input signal is sampled per second, and its reciprocal is the time interval between two adjacent samples. At each sampling moment, the analog-to-digital conversion operator measures the instantaneous amplitude of the conditioned signal. This measurement process compares the continuous analog amplitude with the internal reference voltage and divides the amplitude into several discrete levels according to the quantization precision. Each level corresponds to a unique digital code. In specific implementation, the sampling moment is determined by the product relationship between the sampling rate and the time variable, that is, it increases with time in fixed steps. At each calculated time point, the instantaneous amplitude of the conditioned signal is read, the amplitude is divided by the quantization step size and rounded, and then multiplied by the ratio of the reference voltage to the total number of quantization levels. Finally, a digital quantity corresponding to the amplitude at that moment is obtained. The digital quantities obtained at all sampling moments are arranged in chronological order to form a discrete digital signal stream.

[0023] S13. The digital signal stream is timestamped to obtain a synchronized data frame. In this step, the number of sampling points in each data frame is first determined according to the system's preset time window length. This number is determined by the product of the time window length and the sampling rate. That is, by multiplying the fixed time window length by the number of sampling points per unit time, the number of digital signals that a single data frame should accommodate is calculated. Then, the continuously input digital signal stream is divided into units of this number. Each time the preset number of sampling points are accumulated, a data window is formed. Within this window, all digital quantities maintain their original timing order. At the same time, the high-precision timing module assigns a unique timestamp to the current data window. The timestamp is determined by using the absolute time value corresponding to the start time of the window as a reference and associating it with the time offset of each sampling point in the window relative to the start time, thereby establishing the correspondence between each sampling point and the precise time coordinate. After the timestamp is assigned, the complete digital signal sequence in the window is combined and encapsulated with the timestamp to form a structured data frame. The timestamp is located at the beginning of the frame as an identification field, followed by a fixed-length digital signal sequence.

[0024] This invention employs bandpass filtering technology to accurately extract the characteristic frequency bands of cable sheath monitoring, effectively preserving valid signal components while more thoroughly suppressing out-of-band interference. The introduction of a synchronous sampling mechanism fundamentally eliminates time deviations between channels, ensuring the original authenticity of the phase relationships of multiple signals. By structurally combining digital signal sequences within a fixed time window with high-precision timestamps, precise alignment of multi-source data on the time axis is achieved, enabling correlation analysis of monitoring data distributed at different locations on the cable within a unified time coordinate system.

[0025] S2. Based on the synchronous data frame, the fundamental energy spectrum is fused to obtain the feature parameter vector. Existing technologies usually only calculate the effective value or peak value of the current and ignore the phase information, which makes it impossible to identify the phase change characteristics caused by multiple grounding. The energy integration of the fixed frequency band lacks the ability to finely analyze the spectrum structure, making it difficult to distinguish the frequency domain differences between partial discharge and external interference. It is impossible to form a standardized input format for subsequent algorithms, which increases the data processing complexity of the core analysis steps. To solve the above problems, the specific implementation steps are as follows: S21. Based on the synchronization data frame, the fundamental amplitude and phase are obtained through fundamental component processing. In this step, the digital signal sequence within the current time window is first extracted from the synchronization data frame. This sequence contains a fixed number of sampling points. For the calculation of the power frequency component, the frequency domain index position corresponding to the power frequency is first determined. Then, complex number operations are performed on the index position. The operation process is to multiply the value of each sampling point in the sequence with the value of the complex exponential function determined by the sampling point number and the total number of sampling points as parameters, and then sum all the product results to obtain the complex amplitude corresponding to the frequency point. To further obtain the effective value of the sheath current, the magnitude of the complex amplitude is divided by the square root of 2. That is, the sum of the squares of the real and imaginary parts of the complex number is calculated first, and then the square root of the sum is taken to obtain the magnitude. Finally, the magnitude is divided by the square root of 2 to obtain the fundamental effective value. The phase difference between the sheath current and the reference voltage is obtained by calculating the phase angle of the complex amplitude. The phase angle is determined by dividing the imaginary part of the complex number by the real part to obtain the ratio, and then performing an arctangent operation on the ratio to obtain the fundamental phase. The calculated fundamental effective value and the fundamental phase are combined to construct the fundamental amplitude phase.

[0026] S22. Based on the synchronous data frame, perform transient energy spectrum processing to generate frequency band energy distribution. In this step, the input digital signal sequence is first decomposed into multiple wavelet packet decompositions. In each decomposition, the signal is processed by low-pass and high-pass filters respectively to obtain approximate components and detail components. Each component is downsampled. This decomposition is performed layer by layer to form a complete binary tree structure. Each leaf node corresponds to a sub-band signal of a specific frequency band. For each frequency band, all wavelet packet coefficients under that frequency band are extracted. The square of each coefficient is taken, and then the square values ​​of all coefficients in the same frequency band are summed to obtain the energy value of that frequency band. This operation is specifically as follows: multiply each wavelet packet coefficient by itself to obtain a square value, and then add all the square values ​​in the frequency band in sequence. The final sum is the energy of that frequency band. Repeat the above operation for all frequency bands to calculate the energy value corresponding to each frequency band. These energy values ​​are arranged in ascending order according to the frequency band size to form the frequency band energy distribution.

[0027] S23. Perform feature fusion processing on the fundamental amplitude phase and frequency band energy distribution to obtain a feature parameter vector. In this step, firstly, extract the fundamental effective value and fundamental phase from the fundamental amplitude phase. Arrange these two values ​​as the starting part of the feature parameter vector, with the fundamental effective value placed first and the fundamental phase following immediately. Then, extract the energy values ​​of all frequency bands from the frequency band energy distribution. In order of frequency from low to high, append the energy value of each frequency band to the aforementioned two values. The specific arrangement process is as follows: take the fundamental effective value as the first element of the vector, take the fundamental phase as the second element of the vector, take the lowest frequency band energy value as the third element of the vector, take the second lowest frequency band energy value as the fourth element of the vector, and so on, until the energy values ​​of all frequency bands are placed into the vector. After all elements are arranged in the above order, a one-dimensional feature parameter vector is formed.

[0028] This invention employs Discrete Fourier Transform to simultaneously extract the effective value and phase difference of the sheath current, providing more comprehensive steady-state characteristics at power frequency. The phase-based grounding system state discrimination is more accurate. Wavelet packet decomposition technology is used to decompose the signal into multiple continuous sub-bands and calculate the energy of each band, realizing refined quantization of the transient signal spectral structure. It can effectively distinguish the energy distribution characteristics of different frequency bands. The feature splicing method is used to integrate the fundamental amplitude and phase with the frequency band energy distribution into a one-dimensional feature parameter vector, providing a unified and comprehensive feature input for subsequent analysis.

[0029] S3. High-frequency energy screening is performed on the feature parameter vector to obtain the dominant mode reconstruction signal. Based on the dominant mode reconstruction signal, an enhanced discharge feature sequence is generated through adaptive sequence integration processing. Existing technologies have difficulty distinguishing weak discharge pulses from noise fluctuations. If the threshold is set too high, the false alarm rate will remain high, and if it is set too low, the false alarm rate will surge. The denoising effect lacks adaptability and has limited ability to separate non-stationary and variable high-frequency components in cable sheath signals. Moreover, the threshold function is still globally set and cannot track changes in the local statistical characteristics of the signal, resulting in weak pulses being over-suppressed or distorted during the denoising process, and poor pulse shape preservation. For example, if the insulation of a cable line deteriorates, it will generate periodic discharge pulses with an amplitude of about 15 millivolts. At the same time, a switch cabinet in a nearby section of the same line will perform opening and closing operations, generating broadband interference pulses with an amplitude of about 20 millivolts. The spectrum of the interference pulses extends from 10 kHz to 2 MHz and completely overlaps with the discharge pulses in the time domain, forming a composite pulse waveform with a peak value of 32 millivolts. Current technology will judge the composite pulse as a discharge event, leading to a misdiagnosis of a severe discharge and triggering unnecessary power outages for maintenance. To solve the above problems, the specific steps are as follows: S31. Based on the feature parameter vector, high-frequency reconstructed time-domain signal is obtained through high-frequency feature extraction. In this step, the pre-stored high-frequency band energy data is first extracted from the feature parameter vector. These energy data correspond to the energy distribution values ​​of the original signal on multiple high-frequency sub-bands. Since the inverse Fourier transform requires frequency domain complex amplitude as input, the energy values ​​of each high-frequency band need to be converted into corresponding frequency domain complex amplitudes. The specific conversion method is to perform a square root operation on the energy value of each band to obtain the amplitude magnitude, and assign the corresponding phase value according to the preset phase information, thereby forming a complete complex amplitude at each frequency point. Subsequently, an inverse Fourier transform is performed on the complex sequence containing all high-frequency components. The transform process involves multiplying the complex amplitude at each frequency point by the value of a complex exponential function determined by the frequency point index and the total number of frequency points. The product results of all frequency points are then summed to obtain the instantaneous amplitude at a single time point in the time domain. This summation operation is repeated for each time point until the instantaneous amplitude at all time points is calculated. The calculation results of all time points are then arranged in chronological order to form a high-frequency reconstructed time-domain signal.

[0030] S32. Based on the high-frequency reconstructed time-domain signal, variational mode decomposition is performed to generate an eigenmode function set. In this step, the number of modes to be decomposed is first set, and a center frequency is initialized for each mode. Then, a variational problem is solved iteratively. The problem aims to minimize the sum of the bandwidths of all modes. In each iteration, for each mode, the time-domain signal of the mode is first multiplied by a complex exponential function to achieve spectrum modulation. The spectrum of the mode is then shifted to the baseband position. The Hilbert transform is then performed on the modulated signal to construct an analytic signal to obtain a one-sided spectrum. Then, the square of the gradient of the analytic signal is calculated as a measure of the bandwidth of the mode. Each mode function and its corresponding center frequency are continuously updated by the alternating direction multiplier method. The update method involves performing calculations on the spectrum of the current mode and the remaining components of other modes in the frequency domain. Specifically, the spectrum of the high-frequency reconstructed time-domain signal is subtracted from the sum of the spectra of all other modes, multiplied by an update coefficient in the form of a low-pass filter, and then subjected to an inverse Fourier transform to obtain the updated modal time-domain waveform. The update of the center frequency is determined by calculating the centroid position of the modal spectrum. The above update process is repeated until the center frequencies and waveforms of all modes no longer change significantly. At this point, each mode satisfies the constraint condition that the sum of all modes equals the input high-frequency reconstructed time-domain signal. The final output intrinsic mode function set contains multiple modal components, each mode corresponding to a specific frequency band.

[0031] S33. Perform modal energy screening on the intrinsic mode function set to obtain the dominant mode reconstruction signal. In this step, the energy of each modal component in the intrinsic mode function set is first calculated. Specifically, the instantaneous amplitude of the mode at each time point is squared, and then the squared values ​​at all time points are summed to obtain the total energy of the mode. Then, the total energy of the high-frequency reconstructed time domain signal is calculated. The total energy is equal to the sum of the energies of all modal components. Next, the energy ratio of each modal component is calculated, that is, the energy value of the mode is divided by the total energy value of the high-frequency reconstructed time domain signal to obtain the proportion of the mode in the total energy. The energy proportion of each mode is compared with a preset energy proportion threshold, and all mode components with an energy proportion greater than the threshold are selected. Finally, the instantaneous amplitudes of all selected mode components at the same time point are summed, that is, the time-domain waveforms of these mode components are added point by point to obtain the dominant mode reconstruction signal. Compared with the input intrinsic mode function set, the dominant mode reconstruction signal completes the dimensionality reduction screening from all components to the main energy components, eliminating noise modes and minor components with low energy proportions, and retaining the main signal components that reflect the characteristics of partial discharge pulses.

[0032] S34. Based on the dominant mode reconstruction signal, an adaptive thresholding process is used to obtain a local adaptive threshold curve. In this step, a fixed-length sliding window is first set on the dominant mode reconstruction signal. The window moves point by point along the time axis. At each window position, the mean of all signal values ​​contained in the window is calculated. Specifically, the signal values ​​at each time point in the window are summed, and the summation result is divided by the total number of signal points in the window to obtain the local mean at the center of the window. At the same time, the standard deviation is calculated. Specifically, the difference is obtained by subtracting the local mean of the window from each signal value within the window. Each difference is then squared to obtain a squared value. All squared values ​​are summed and divided by the total number of signal points within the window to obtain the variance. The square root of the variance is then taken to obtain the local standard deviation at the center of the window. The calculated local standard deviation is then multiplied by a preset adaptive coefficient to obtain the product. This product is then added to the local mean to obtain the adaptive threshold corresponding to the center of the window. After the sliding window traverses the entire signal, the adaptive thresholds calculated at all time points are arranged in chronological order to form a local adaptive threshold curve.

[0033] S35. Based on the local adaptive threshold curve and the dominant mode reconstruction signal, pulse over-threshold detection is performed to obtain a candidate pulse segment set. In this step, at each time point, the instantaneous amplitude of the dominant mode reconstruction signal is compared with the threshold of the local adaptive threshold curve at the same time point. The comparison method is to subtract the threshold of the time point from the instantaneous amplitude. If the difference is greater than or equal to zero, it is determined that the signal value at that time point exceeds the threshold, and the original signal value at that time point is retained as the component point of the candidate pulse. If the difference is less than zero, it is determined that the signal value at that time point does not exceed the threshold, and the output value at that time point is set to zero. The above comparison operation is performed one by one at all time points to form a time sequence of the same length as the input signal. The original amplitude is retained for time points exceeding the threshold, while the output is zero for time points not exceeding the threshold. Since partial discharge pulses are usually segments with multiple consecutive time points exceeding the threshold, the intervals of consecutive non-zero values ​​in this sequence constitute a complete pulse segment. Each interval of consecutive non-zero values ​​is extracted to form a candidate pulse segment set. Compared with the input dominant mode reconstruction signal and the local adaptive threshold curve, this candidate pulse segment set completes the conversion from continuous time domain waveform to discrete pulse segment, separating the pulse activity that may be caused by partial discharge from the background signal.

[0034] S36. Perform pulse sequence integration processing on the candidate pulse segment set to generate a discharge feature enhancement sequence. In this step, a complete time axis covering the time range of all candidate pulse segments is first determined. This time axis is consistent with the time length of the original signal. Then, at each time point, the instantaneous amplitude of all candidate pulse segments at that time point is summed. Specifically, the amplitude corresponding to each pulse segment at that time point is taken out. If a pulse segment has no amplitude at that time point, the value is taken as zero. All values ​​are added together to obtain the sum value at that time point. Since the candidate pulse segments do not overlap in time, at most one segment at each time point has a non-zero amplitude. Therefore, the result of the summation is the pulse amplitude at that time point. Repeat the above summation operation for all time points. Arrange the sum values ​​calculated at each time point in chronological order to form a continuous discharge feature enhancement sequence.

[0035] This invention can adaptively decompose non-stationary high-frequency reconstructed time-domain signals into multiple intrinsic mode functions. It not only avoids the empirical dependence on the selection of basis functions in wavelet transform, but also effectively separates discharge pulse components that overlap with noise frequencies. By filtering by energy proportion, only the dominant modes with higher energy proportions are retained for reconstruction. This not only achieves coarse extraction of discharge pulses, but also provides an input signal with a significantly improved signal-to-noise ratio for subsequent adaptive threshold calculation. It can solve the problem of missed or false alarms in the existing fixed threshold method. For example, a 110 kV cable line is located in the core area of ​​the city. The cable channel is laid in parallel with the subway power supply system. The subway traction current generates strong broadband electromagnetic interference, whose frequency range covers 10 kHz to 500 kHz, which highly overlaps with the frequency band of partial discharge signals. At the same time, the pulse amplitude at the intermediate joint of the cable is only about 5 millivolts, while the background noise amplitude reaches 3 to 5 millivolts, and the signal-to-noise ratio is close to zero dB. If the traditional fixed threshold method sets the threshold to 6 millivolts, it will miss all discharge pulses. If the threshold is set to 4 millivolts, it will generate thousands of false alarms per second. The adaptive threshold generation of this invention uses a sliding window to successfully preserve the entire pulse tail. The pulse sequence completely preserves the entire process of the pulse from its start, peak value to attenuation. The pulse energy calculation error is reduced to less than 8%. In contrast, the existing fixed threshold method, if the threshold is set to 30% of the peak amplitude, completely loses the pulse tail and can only extract the pulse head, resulting in a discharge energy calculation error of more than 60%. It is impossible to identify the discharge type through pulse width and attenuation characteristics. For example, in a cable line with a total length of eight kilometers, the discharge point is located at the cable terminal at 6.5 kilometers from the monitoring terminal. Due to the long distance attenuation, the high-frequency component attenuation is particularly severe. The pulse waveform widens from the initial sharp pulse to a slowly varying waveform with a rising edge of about 10 microseconds and a falling edge of about 80 microseconds. The amplitude of the pulse tail attenuates to less than 15% of the peak amplitude, and the amplitude of the tail is similar to that of the background noise. This invention utilizes the characteristic that the spectrum width of the interference pulse is much larger than that of the discharge pulse, and uses variational mode decomposition to decompose it into multiple frequency bands. Energy proportion screening filters out interference pulses due to their low energy proportion in a single frequency band, reducing the energy of the interference pulse to less than 10% of its original value. This ensures that real discharge events can still be accurately identified in complex interference environments, improving the signal-to-noise ratio by 8 to 12 dB, enhancing pulse waveform shape preservation, reducing waveform distortion rate to less than 5%, and simultaneously optimizing the false alarm rate and missed alarm rate. This provides a more reliable and complete partial discharge characteristic input for subsequent multi-physics coupling diagnosis, fundamentally improving the comprehensive diagnostic capability of online monitoring of cable sheaths. The variational mode decomposition of this invention allows the local standard deviation to more accurately reflect the noise level during subsequent adaptive threshold calculation, resulting in more precise threshold setting. Conversely, the adaptive threshold can further eliminate residual interference that was not completely filtered out by the variational mode decomposition, forming a double-screening result. For example, if the insulation of a cable line deteriorates, generating periodic discharge pulses with an amplitude of about 15 millivolts, and at the same time, a nearby switchgear on the same line segment performs opening and closing operations, generating broadband interference pulses with an amplitude of about 20 millivolts, the spectrum of which extends from 10 kHz to 2 MHz, completely overlaps with the discharge pulses in the time domain, forming a composite pulse waveform with a peak value of 32 millivolts. Existing technology judges the composite pulse as a discharge event, leading to a misdiagnosis of severe discharge and triggering unnecessary power outages for maintenance.

[0036] S4. Based on the feature parameter vector, the core mode feature sequence is obtained through time-series dimensionality reduction. The core mode feature sequence is then subjected to dynamic time-limited output processing to obtain the sheath current anomaly index. Existing technologies cannot distinguish between normal load fluctuations and current changes caused by faults. In scenarios with frequent load fluctuations, the false alarm rate is extremely high. When load changes cause the current waveform to be stretched or compressed on the time axis, even if the waveform shape is essentially normal, the Euclidean distance will produce a large deviation, causing a large number of false alarms. Furthermore, the phase difference is ignored, and the phase change characteristics caused by faults such as multi-point grounding cannot be identified. To solve the above problems, the specific steps are as follows: S41. Based on the feature parameter vectors, a current timing window matrix is ​​generated through timing window construction. In this step, a fixed window length is first preset, which determines the number of historical moments contained in each output matrix. Then, the feature parameter vectors input at each moment are continuously cached. The vector at each moment consists of two values: the effective value of the sheath current and the phase difference of the sheath current. At each current moment, the feature parameter vector at the current moment is taken as the latest data, and the historical data of the window length minus one moment is traced back. The feature parameter vectors of the consecutive window lengths including the current moment are arranged in chronological order to form a two-dimensional matrix structure. The row index of this matrix corresponds to different historical moments, and the column index corresponds to the two feature parameters contained in each moment. This arrangement process is as follows: the feature parameter vector of the earliest moment is placed in the first row of the matrix, the feature parameter vector of the second earliest moment is placed in the second row, and so on, until the feature parameter vector of the current moment is placed in the last row of the matrix, thus completing the construction of the entire current timing window matrix.

[0037] S42. Perform multi-head attention feature processing based on the current time-series window matrix to generate a weighted feature sequence. In this step, the input matrix is ​​first multiplied by three different weight matrices to obtain the query matrix, key matrix, and value matrix. Then, the transpose of the query matrix and the key matrix are multiplied to obtain the attention score matrix. Each element in this matrix represents the correlation strength between different time points. All elements in this matrix are divided by the square root of the key vector dimension to complete the scaling operation to stabilize the gradient. Then, the normalized exponent operation is performed row by row on the scaled matrix. That is, each element in each row is raised to the power of the natural constant, and all the exponents in the same row are summed. Finally, the sum of the elements in each row is divided by the exponent of each element to obtain the attention weight matrix. The sum of the elements in each row of this matrix is ​​one, reflecting the degree of attention paid to other times at each time step. Then, the attention weight matrix is ​​multiplied by the value matrix to obtain the output of a single-head attention. The above process is executed in parallel for multiple attention heads. Each head uses an independent weight matrix to perform the same operation. The outputs of all heads are concatenated along the feature dimension, that is, the corresponding elements of the output of each head are arranged in order to form a longer vector. Finally, the concatenation result is multiplied by the output weight matrix to obtain the weighted feature sequence.

[0038] S43. Perform principal component dimensionality reduction on the weighted feature sequence to obtain the core pattern feature sequence. In this step, firstly, calculate the mean of each feature dimension in the input data, that is, sum the values ​​of all time steps in each feature dimension, and then divide the sum by the total number of time steps to obtain the mean of that dimension. Then, subtract the mean of the corresponding dimension from each feature value at each time step to obtain the centered data matrix. Next, calculate the covariance matrix of the centered data matrix. Specifically, multiply the transpose of the centered data matrix with itself to obtain the covariance value between each feature dimension and other feature dimensions. Next, divide each element in the obtained matrix by the total number of time points minus one to complete the calculation of the covariance matrix. Then, solve for the eigenvalues ​​and eigenvectors of the covariance matrix. The eigenvalues ​​reflect the magnitude of the data variance in each principal component direction, and the eigenvectors correspond to the projection direction. Sort all eigenvalues ​​from largest to smallest and rearrange the eigenvectors according to the correspondence. Select the first few eigenvectors to form the projection matrix. Finally, perform matrix multiplication on the centered data matrix and the projection matrix. That is, sum the products of each original eigenvalue at each time point and the corresponding element in the projection matrix to obtain the coordinates of that time point in the new low-dimensional space, which is the core pattern feature sequence.

[0039] S44. Based on the core pattern feature sequence, the minimum normalized distance value is obtained through dynamic temporal similarity processing. In this step, the normalized distance between the current input core pattern feature sequence and each normal pattern sequence in the historical normal pattern library is calculated. For each pair of sequences, the calculation process is to construct a distance matrix, where the rows of the matrix correspond to the time points of the current sequence and the columns correspond to the time points of the normal pattern sequence. The value of each element in the matrix is ​​obtained by calculating the difference between the value of the current sequence at that time point and the value of the normal pattern sequence at the corresponding time point, squaring the difference, and finally taking the square root of the squared result. Then, a normalized path is found from the starting point to the ending point of the matrix. This path must satisfy the adjacent step size constraint and minimize the cumulative sum of all elements passed through the path. The specific way to find the minimum cumulative sum is to start from the starting point of the matrix and calculate the minimum cumulative distance to reach each point. The minimum cumulative distance of each point is equal to the distance of the point itself plus the minimum cumulative distance of the three adjacent points to its left, below, or lower left. This process is repeated until the ending point of the matrix. The cumulative distance at the ending point is the normalized distance value of the sequence. After calculating the normalized distance between the current sequence and all normal pattern sequences, the minimum value is selected from all normalized distance values ​​to obtain the minimum normalized distance value.

[0040] S45. Normalize the anomaly index based on the minimum regularization distance value to generate a normalized anomaly index. In this step, the normalization factor is first obtained. This normalization factor is predetermined by the data under historical normal operation. Specifically, the minimum regularization distance values ​​at all times during the historical normal operation period are summed, and the sum is divided by the total number of historical data to obtain the average regularization distance. This average value is used as the normalization factor. Then, the input minimum regularization distance value is divided by the normalization factor to complete the division operation and obtain the normalized anomaly index. Specifically, the division operation uses the minimum regularization distance value as the dividend and the normalization factor as the divisor to calculate the quotient. This quotient is the normalized anomaly index.

[0041] S46. Perform amplitude limiting processing on the normalized abnormal index to obtain the sheath current abnormal index. In this step, the value of the normalized abnormal index is compared with the preset upper limit value. If the normalized abnormal index is less than or equal to one, the value of the normalized abnormal index is directly used as the output value. If the normalized abnormal index is greater than one, the upper limit value is used as the output value. The specific calculation method for this comparison and value selection is as follows: first, calculate the difference between the normalized abnormal index and the value one. If the difference is negative or zero, the normalized abnormal index is selected as the result. If the difference is positive, the value one is selected as the result. After the above amplitude limiting processing, the sheath current abnormal index is output.

[0042] This invention, by calculating the correlation strength between different times, not only effectively enhances the temporal features but also provides a core mode feature sequence with a higher signal-to-noise ratio for subsequent dynamic time warping. For example, during the morning and evening peak hours, the current amplitude of a cable line increases sharply from 100 amperes to 300 amperes, causing significant waveform deformation. Existing technologies trigger false alarms due to the amplitude exceeding the threshold, while the multi-head attention mechanism focuses attention weights on the waveform change rate rather than the absolute amplitude, effectively suppressing the load fluctuation information in the weighted feature sequence and thus avoiding misjudgment of the subsequent warping distance.

[0043] This invention compresses weighted feature sequences into the core pattern feature space through principal component analysis, preserving the main changes in the data while removing redundancy and noise. Through dynamic time warping, it constructs warping paths to allow nonlinear alignment of two sequences on the time axis and calculates the minimum warping distance. For example, due to the large-scale access of electric vehicle charging piles, the load on a certain cable exhibits intermittent impact characteristics. The current waveform on the time axis exhibits both overall translation and local stretching and compression. The Euclidean distance deviates significantly due to the inability to align time points. However, dynamic time warping finds the optimal warping path to align waveform segments with similar shapes, resulting in a warping distance that is significantly smaller than the Euclidean distance. In actual tests, under load impact scenarios, the Euclidean distance deviation of normal waveforms reaches more than 40%, while the dynamic time warping distance deviation is 8%, reducing the false alarm rate by 75%.

[0044] This invention converts the minimum regularization distance into a sheath current anomaly index in the range of zero to one through normalization and amplitude limiting. In actual operation, at the initial stage of a fault, the anomaly index slowly rises from 0.1 to 0.3, and the system is judged to be in a state of alert. When the fault develops to the point where the phase difference shift exceeds five degrees, the anomaly index rises to 0.7, and the system triggers an early warning, successfully issuing an alarm before the fault worsens. This achieves early identification of gradual faults. In practice, for example, if a cable grounding wire experiences slow corrosion, the sheath current phase difference gradually increases at a rate of 0.5 degrees per day, while simultaneously experiencing daily load fluctuations. Existing technologies struggle to distinguish between phase drift and load changes. This invention reduces the false alarm rate by more than 80% under conditions of frequent load fluctuations and significantly improves the sensitivity to phase difference changes. It achieves accurate capture of slow and minute fault characteristics of the sheath current, providing a highly reliable anomaly quantification index for subsequent multi-physics coupled fault diagnosis.

[0045] S5. Based on the feature parameter vector, discharge feature enhancement sequence and sheath current anomaly index, a fusion network coupling process is performed to obtain physical constraint enhancement features. Fault probability diagnosis processing is performed on the physical constraint enhancement features to generate comprehensive diagnostic results and early warning signals. Existing technologies use pure data-driven neural network models for classification. Although they can learn the statistical laws in the data, they lack physical constraints. The diagnostic results may violate the basic laws of thermodynamics and are sensitive to changes in the distribution of training data. They also have poor generalization ability and limited accuracy in new fault modes that have not been seen before. For example, if multiple high-voltage cables are laid in a cable tunnel at the same time, the intermediate joint of one of the cables will experience local overheating due to increased contact resistance. The phase difference of the sheath current will shift by one degree, and the sheath current abnormality index will be 0.35. At the same time, the adjacent cable will generate a strong electromagnetic field due to a sudden increase in load current. Through inductive coupling, an 8 mA power frequency interference current will be introduced into the sheath of the cable. Its phase will be 180 degrees different from the normal sheath current. After superposition, the effective value distribution of the sheath current at multiple monitoring points of the cable will show an abnormal pattern of high at both ends and low in the middle. To solve the above problems, the specific implementation steps are as follows: S51. Based on the feature parameter vector, discharge feature enhancement sequence, and sheath current anomaly index, a spatiotemporally aligned data tensor is generated through alignment and fusion processing. In this step, the feature parameter vector, the instantaneous amplitude of the discharge feature enhancement sequence, and the sheath current anomaly index at the same time are aligned according to a unified high-precision timestamp to ensure that the three correspond to the same time section. Then, the three types of aligned data are spliced ​​together. Specifically, each value in the feature parameter vector, the instantaneous amplitude of the discharge feature enhancement sequence, and the sheath current anomaly index are arranged in a fixed order to form a one-dimensional data row. For multiple monitoring points arranged along the cable, the one-dimensional data rows of each monitoring point aligned and spliced ​​at the same time are stacked according to the spatial order of the monitoring points to form a two-dimensional data matrix. The rows of this matrix correspond to different monitoring points, and the columns correspond to various feature quantities. Finally, the two-dimensional data matrices of multiple consecutive times are stacked in chronological order to form a three-dimensional spatiotemporally aligned data tensor.

[0046] S52. Perform graph convolutional network processing based on spatiotemporally aligned data tensors to obtain spatial feature embedding values; obtain the cable topology graph structure. In this step, each monitoring point in the cable topology graph structure is first treated as a node, and the electrical connection relationship between nodes is treated as an edge to construct the graph structure. For each node, it is necessary to aggregate the feature information of itself and all neighboring nodes. In the specific calculation, firstly, add self-loop edges to each node, that is, let each node also establish a connection with itself, so that the node's own features can be preserved in the aggregation process. Then, calculate the adjacency matrix after adding self-loops, and calculate the degree matrix based on the matrix. Each element on the diagonal of the degree matrix is ​​equal to the number of neighbors of the corresponding node. Then, normalize the adjacency matrix. Specifically, divide the product of each element in the adjacency matrix and the square root of the corresponding row and column in the degree matrix to obtain the normalized adjacency matrix. Then, the normalized adjacency matrix is ​​multiplied by the node feature matrix of the current layer. The operation is as follows: for each node, the feature vectors of all its neighboring nodes are weighted and summed. The weights are determined by the normalized adjacency matrix to obtain the aggregated feature of the node. The aggregated feature is multiplied by the trainable weight matrix to complete the linear transformation. Then, each element in the linear transformation result is processed by a non-linear activation function. That is, a threshold judgment is performed on each input value. If it is greater than zero, the original value is retained. If it is less than zero, it is set to zero to obtain the output feature of the layer. The above process is repeated multiple times, with the output of the previous layer as the input for each layer. By stacking multiple layers of graph convolution, each node can capture information from its more distant neighbors and obtain spatial feature embedding values.

[0047] S53. Physical neural network coupling processing is performed on the spatial feature embedding values ​​to obtain physical constraint enhancement features. In this step, the effective value of the sheath current at each monitoring point is first extracted from the input spatial feature embedding. The effective value is squared and then multiplied by the sheath grounding resistance to obtain the Joule heat source term. Then, the Joule heat source term is substituted into the heat conduction equation. The heat conduction equation includes three parts: temperature gradient term, Joule heat source term, and temperature change term over time. In the specific calculation, the product of the spatial second derivative of temperature and thermal conductivity is calculated to obtain the temperature gradient divergence term. The Joule heat source term is calculated. The partial derivative of temperature with respect to time is multiplied by the product of density and specific heat capacity to obtain the temperature change term. The temperature gradient divergence term and the Joule heat source term are added together and the temperature change term is subtracted to obtain the residual value of the heat conduction equation. The residual value should be zero in an ideal case. In actual calculation, due to the deviation between the model prediction and the physical law, the residual is non-zero. The residual is squared to obtain the single-point physical constraint loss. The physical constraint losses of all monitoring points are summed to obtain the total physical loss term. At the same time, spatial features are embedded and mapped to the fault type prediction results through a fully connected layer. The prediction results are compared with the real fault labels, the difference is calculated and squared, and then the squared differences of all samples are summed and averaged to obtain the data loss term. The data loss term and the physical loss term are weighted and summed, that is, the physical loss term is multiplied by a preset weight coefficient and then added to the data loss term to obtain the total loss value. During the neural network training process, the network weights are continuously adjusted through the backpropagation algorithm to minimize the total loss value. After training is completed, the spatial features are embedded into the optimized network, and forward propagation is used to obtain the physical constraint enhancement features.

[0048] S54. Based on the enhanced physical constraint features, the fault probability inference process is used to obtain the fault type probability distribution. In this step, the enhanced physical constraint feature vector is first multiplied with the pre-trained weight matrix. That is, for each fault category, each element in the feature vector is multiplied with each element in the corresponding row of the weight matrix, and then all the product results are summed to obtain the initial score of the category. Then, the bias value of the corresponding category in the bias vector is added to the initial score to obtain the final score of the category. Repeat the above operation for all preset fault categories to obtain a score vector containing the scores of each category. Then, perform an exponential operation on each score in the score vector, that is, calculate the power value with the score as the exponent and the natural constant as the base. Then, sum the results of the exponential operation of all categories to obtain a normalized denominator. Finally, divide the exponential operation result of each category by the normalized denominator to obtain the predicted probability of each fault category. The probability values ​​of all categories constitute the fault type probability distribution, where each probability value is between 0 and 1, and the sum of the probabilities of all categories is 1.

[0049] S55. Based on the sheath current anomaly index, discharge characteristic enhancement sequence, and fault type probability distribution, the severity of the fault is assessed to obtain the fault severity level. In this step, the maximum value of the probability of all categories is first extracted from the fault type probability distribution. This maximum value represents the fault type and its probability of occurrence most likely corresponding to the current monitoring data. Then, the maximum instantaneous amplitude at all time points is extracted from the discharge characteristic enhancement sequence. This maximum amplitude is divided by the preset reference discharge amplitude to obtain the normalized discharge intensity index. Subsequently, the maximum value of the probability of all categories, the sheath current anomaly index, and the discharge intensity index are multiplied by their respective preset weight coefficients. The preset weight coefficients can be obtained by the analytic hierarchy process, that is, the maximum probability of the fault is multiplied by the first weight coefficient to obtain the first weighted value, the sheath current anomaly index is multiplied by the second weight coefficient to obtain the second weighted value, and the normalized discharge intensity index is multiplied by the third weight coefficient to obtain the third weighted value. Finally, the three weighted values ​​are summed to obtain the fault severity level.

[0050] S56. Process the diagnostic results of the fault type probability distribution and fault severity level to generate a comprehensive diagnostic result and a warning signal. In this step, firstly, extract the maximum value of the probability of all categories from the fault type probability distribution, and compare this maximum value with the preset thresholds of 0.6 and 0.8 respectively. If the maximum value is less than 0.6, the warning level is determined to be normal. If the maximum value is greater than or equal to 0.6 and less than 0.8, the warning level is determined to be attentive level, and it is recommended to strengthen monitoring. If the maximum value is greater than or equal to 0.8, the warning level is determined to be dangerous level, and a warning signal is generated. Then, combine the fault type name, fault severity level, and warning level value corresponding to the maximum probability value in the fault type probability distribution to form a comprehensive diagnostic result text description. This comprehensive diagnostic result text description specifically includes fault type, severity quantification value, and warning level information. Among them, the warning level of normal means that the operation is normal and no action is required, the attentive level means that attention is required and it is recommended to strengthen monitoring, and the dangerous level means that the dangerous state is required and it is recommended to repair immediately. Finally, the comprehensive diagnostic result and warning signal are transmitted to the cable monitoring center.

[0051] This invention utilizes a graph neural network and a cable topology graph structure. Through graph convolution operations, it not only achieves spatial enhancement of isolated features of each monitoring point, but also provides Joule heat source distribution input containing topological dependencies for subsequent physical information neural networks. This can significantly improve the accuracy of identification. In the special scenario of a multi-point grounding fault in a cable line, the fault currents of the two grounding points are superimposed, which is a problem that traditional single-point monitoring cannot distinguish the fault source.

[0052] This invention embeds the thermoelectric coupling physical equations as prior knowledge into the network training process using a physical information neural network. This is not only supervised by fault labels but also subject to the residual constraints of the heat conduction equation, namely the conservation relationship between the Joule heat source term, the temperature gradient divergence term, and the temperature change term. This ensures the diagnostic results are physically self-consistent, avoiding the physical uninterpretability that may arise from purely data-driven models. More importantly, the spatial feature embedding provided by the graph neural network injects topological dependencies into the physical information neural network, enabling the thermal field calculation to propagate along the actual cable structure, forming a linkage effect between spatial topology and thermal field propagation. In a real-world case, a cable joint experienced localized overheating due to poor installation. Initially, the sheath current anomaly index was only 0.2, which traditional methods considered normal. However, the graph neural network captured the minute temperature differences between adjacent monitoring points. Based on this, the physical information neural network solved the heat conduction equation and found that the residual between the Joule heat source term and the temperature field significantly deviated from zero, inferring the existence of an abnormal heat source and providing an early warning of the joint overheating fault two weeks in advance.

[0053] This invention integrates three aspects of information: physical constraint enhancement characteristics, sheath current anomaly index, and discharge characteristics. When the cable sheath simultaneously exhibits two fault modes—current phase drift caused by grounding wire corrosion and high-frequency pulses generated by partial discharge—traditional single-dimensional diagnostic methods cannot distinguish the primary and secondary relationship between the two. For example, in a cable tunnel where multiple high-voltage cables are laid simultaneously, the intermediate joint of one cable experiences localized overheating due to increased contact resistance, causing a one-degree shift in the sheath current phase difference and a sheath current anomaly index of 0.35. At the same time, an adjacent cable generates a strong electromagnetic field due to a sudden increase in load current, introducing an 8 mA power frequency interference current into the cable sheath through inductive coupling. Its phase differs from the normal sheath current by 180 degrees. After superposition, the effective value distribution of the sheath current at multiple monitoring points of the cable exhibits an abnormal pattern of high values ​​at both ends and low values ​​in the middle. This invention first uses the effective values ​​of sheath current, phase difference, and temperature data from eight monitoring points along the cable line as node features. Based on the cable topology, spatial convolution aggregation is performed to identify induced interference patterns as spatial high-frequency components and fault patterns as spatial low-frequency components, achieving spatial decoupling of the two physical mechanisms. A physical information neural network then performs thermoelectric coupling inference based on this. The spatial decoupling features output from the graph neural network are used to calculate the Joule heat source distribution corresponding to the two modes. The abnormal sheath current is mainly caused by external electromagnetic induction interference, rather than internal multi-point grounding faults. The real fault point requiring attention is slight overheating at the joint, and the severity level of the fault is given. This not only achieves precise separation of faults and interference but also allows the physical information neural network to verify the physical consistency of each potential fault source through spatial decoupling features, rather than attributing all electrical anomalies to the same fault source. This further optimizes fault location accuracy, avoids unnecessary excavation and maintenance, saves costs, and the diagnostic results can directly guide maintenance decisions, significantly improving maintenance efficiency.

[0054] Example 2: Due to the lack of physical constraints in current technologies, diagnostic results may violate fundamental laws of thermodynamics. Furthermore, these technologies are sensitive to changes in training data distribution, limiting monitoring accuracy. Please refer to [link to relevant documentation]. Figure 2 The diagram shown is a structural block diagram of a comprehensive online monitoring system for cable sheaths provided in this embodiment. The system includes a synchronization module, a characteristic parameter module, a discharge characteristic module, a sheath current module, and a comprehensive diagnostic module. The synchronization module is used to acquire the raw electrical signals of the target cable, and to obtain synchronization data frames by encapsulating and processing the raw electrical signals through filtering and timestamps. The feature parameter module is used to perform fundamental wave energy spectrum fusion processing based on synchronous data frames to obtain feature parameter vectors; The discharge feature module is used to perform high-frequency energy screening on the feature parameter vector to obtain the dominant mode reconstruction signal. Based on the dominant mode reconstruction signal, an adaptive sequence integration process is used to generate a discharge feature enhancement sequence. The sheath current module is used to obtain the core mode feature sequence by time-series dimensionality reduction based on the feature parameter vector, and to perform dynamic time-limited output processing on the core mode feature sequence to obtain the sheath current anomaly index. The comprehensive diagnostic module is used to perform fusion network coupling processing based on feature parameter vectors, discharge feature enhancement sequences and sheath current anomaly index to obtain physical constraint enhancement features, perform fault probability diagnostic processing on the physical constraint enhancement features, and generate comprehensive diagnostic results and early warning signals.

[0055] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, this application can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code, including but not limited to disk storage, CD-ROM, optical storage, etc.

[0056] The above embodiments provide a detailed description of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for comprehensive online monitoring of cable sheaths, characterized in that, The steps of this method are as follows: acquire the original electrical signal of the target cable, encapsulate and process the original electrical signal through filtering and timestamp to obtain a synchronization data frame; Based on the synchronous data frames, fundamental wave energy spectrum fusion processing is performed to obtain feature parameter vectors; High-frequency energy screening is performed on the feature parameter vector to obtain the dominant mode reconstruction signal. Based on the dominant mode reconstruction signal, an enhanced discharge feature sequence is generated through adaptive sequence integration processing. Based on the feature parameter vector, the core mode feature sequence is obtained through time-series dimensionality reduction. The core mode feature sequence is then subjected to dynamic time-limited output processing to obtain the sheath current anomaly index. Based on the feature parameter vector, the discharge feature enhancement sequence and the sheath current anomaly index, a fusion network coupling process is performed to obtain the physical constraint enhancement feature. The physical constraint enhancement feature is then subjected to fault probability diagnosis processing to generate a comprehensive diagnosis result and an early warning signal.

2. The method for comprehensive online monitoring of cable sheaths according to claim 1, characterized in that, The original electrical signal is processed by filtering and timestamp encapsulation, including: The original electrical signal is bandpass filtered to generate a filtered and conditioned signal. Analog-to-digital conversion is performed on the filtered and conditioned signal to obtain a digital signal stream; The digital signal stream is timestamped to obtain a synchronized data frame.

3. The method for comprehensive online monitoring of cable sheaths according to claim 1, characterized in that, Fundamental energy spectrum fusion processing based on synchronous data frames includes: The fundamental amplitude and phase are obtained by processing the fundamental component of the synchronous data frame; Transient energy spectrum processing is performed based on synchronous data frames to generate frequency band energy distribution; The fundamental amplitude, phase, and frequency band energy distribution are fused to obtain a feature parameter vector.

4. The method for comprehensive online monitoring of cable sheaths according to claim 1, characterized in that, High-frequency energy filtering of feature parameter vectors includes: Based on the feature parameter vector, a high-frequency reconstructed time-domain signal is obtained through high-frequency feature extraction. Variational mode decomposition is performed on high-frequency reconstructed time-domain signals to generate an eigenmode function set. Modal energy screening is performed on the intrinsic mode function set to obtain the dominant mode reconstruction signal.

5. The method for comprehensive online monitoring of cable sheaths according to claim 1, characterized in that, Based on the dominant mode reconstruction signal, adaptive sequence integration processing is performed, including: Based on the dominant mode reconstruction signal, an adaptive thresholding process is used to obtain a local adaptive threshold curve; Pulse over-threshold detection is performed based on local adaptive threshold curves and dominant mode reconstruction signals to obtain a candidate pulse segment set; The candidate pulse segment set is processed by pulse sequence integration to generate a discharge feature enhancement sequence.

6. The method for comprehensive online monitoring of cable sheaths according to claim 1, characterized in that, Based on the feature parameter vector, temporal dimensionality reduction processing is performed, including: Based on the feature parameter vector, a current timing window matrix is ​​generated by constructing and processing the timing window. Multi-head attention feature processing is performed based on the current time-series window matrix to generate a weighted feature sequence; Principal component dimensionality reduction is performed on the weighted feature sequence to obtain the core pattern feature sequence.

7. The method for comprehensive online monitoring of cable sheaths according to claim 1, characterized in that, Dynamic time-limited output processing is performed on the core pattern feature sequence, including: The minimum regular distance value is obtained by processing the core pattern feature sequence through dynamic temporal similarity. Anomaly index normalization is performed based on minimum regularity distance value to generate normalized anomaly index; The normalized anomaly index is subjected to amplitude limiting output processing to obtain the sheath current anomaly index.

8. The method for comprehensive online monitoring of cable sheaths according to claim 1, characterized in that, A fusion network coupling process is performed based on feature parameter vectors, discharge feature enhancement sequences, and sheath current anomaly index, including: Based on the feature parameter vector, the discharge feature enhancement sequence, and the sheath current anomaly index, a spatiotemporally aligned data tensor is generated through alignment and fusion processing. Graph convolutional network processing is performed based on spatiotemporally aligned data tensors to obtain spatial feature embedding values; Physically constrained enhanced features are obtained by coupling the spatial feature embedding values ​​with a physical neural network.

9. The method for comprehensive online monitoring of cable sheaths according to claim 1, characterized in that, Fault probability diagnosis processing is performed on the physical constraint enhancement features, including: Based on the enhanced physical constraints, the failure probability inference process is used to obtain the failure type probability distribution; The severity of a fault is assessed based on the sheath current anomaly index, the enhanced discharge characteristic sequence, and the probability distribution of fault types, resulting in a fault severity level. The diagnostic results are processed based on the probability distribution of fault types and the severity level of faults to generate comprehensive diagnostic results and early warning signals.

10. A system applied to the comprehensive online monitoring method for cable sheaths as described in any one of claims 1-9, characterized in that, The system includes: The synchronization module is used to acquire the raw electrical signals of the target cable, and to obtain synchronization data frames by encapsulating and processing the raw electrical signals through filtering and timestamps. The feature parameter module is used to perform fundamental wave energy spectrum fusion processing based on synchronous data frames to obtain feature parameter vectors; The discharge feature module is used to perform high-frequency energy screening on the feature parameter vector to obtain the dominant mode reconstruction signal. Based on the dominant mode reconstruction signal, an adaptive sequence integration process is used to generate a discharge feature enhancement sequence. The sheath current module is used to obtain the core mode feature sequence by time-series dimensionality reduction based on the feature parameter vector, and to perform dynamic time-limited output processing on the core mode feature sequence to obtain the sheath current anomaly index. The comprehensive diagnostic module is used to perform fusion network coupling processing based on feature parameter vectors, discharge feature enhancement sequences and sheath current anomaly index to obtain physical constraint enhancement features, perform fault probability diagnostic processing on the physical constraint enhancement features, and generate comprehensive diagnostic results and early warning signals.

Images