An artificial intelligence grounding fault detection method for a submarine single-core power transmission cable

Abstract

The application discloses an artificial intelligence grounding fault detection method for a submarine single-core power transmission cable, which comprises the following steps: coupling a capacitor clamp to a sheath and / or armor of a submarine single-core cable, collecting background signals and performing frequency spectrum analysis, selecting and injecting high-frequency signals; noise cancellation is performed through an improved adaptive filter, and adaptive filtering state features are extracted; the amplitudes and phases of high-frequency voltage and current are obtained, grounding impedance and grounding resistance are calculated, and sheath and armor grounding current distribution ratios are calculated based on a line core-sheath-armor three-conductor pi type equivalent circuit model; the above-mentioned physical parameters, adaptive filtering state features and time domain, frequency domain and time-frequency domain statistical features are used to construct a multi-channel feature tensor, which is input into a trained convolution-attention deep learning model for comprehensive intelligent diagnosis, and grounding condition classification results and probabilities are output. The application realizes non-invasive online detection and has high precision and high robustness under strong electromagnetic interference.

Description

Technical Field

[0001] This invention relates to the field of online monitoring and fault diagnosis technology for high-voltage cables, and in particular to an artificial intelligence-based grounding fault detection method for submarine single-core power transmission cables. Background Technology

[0002] With the continuous expansion of power systems and the large-scale integration of new energy sources such as offshore wind power, the application of submarine high-voltage single-core cables in long-distance power transmission is becoming increasingly widespread. Offshore wind farms typically use 110kV and above voltage-level single-core submarine high-voltage cables, which have become the main channel for onshore-sea power transmission. Their operational reliability directly affects the power generation utilization hours of offshore wind farms and grid security. Unlike terrestrial cables, due to limitations in submarine laying conditions, single-core submarine cables cannot use conventional measures such as cross-interconnection to suppress induced voltage and circulating current in the metal sheath. In engineering, a simplified method of directly grounding the sheath and armor at both ends is commonly used. This method has inherent defects: when the grounding structure experiences increased contact resistance due to corrosion or loosening, it will cause abnormal distribution of grounding current between the sheath and armor, easily leading to continuous high-temperature overheating at poor contact points, ultimately inducing serious accidents such as main insulation breakdown. This results in extremely high maintenance costs and can cause prolonged shutdowns of the wind farm.

[0003] High-voltage cable grounding faults are complex in nature, including metallic grounding with extremely low resistance, transitional grounding with variable resistance, and intermittent grounding with intermittent continuity. Their transient and steady-state characteristics differ significantly, further increasing the difficulty of accurate diagnosis.

[0004] Traditional diagnostic methods have significant shortcomings in addressing these challenges. Traditional methods for diagnosing submarine cable grounding faults primarily rely on manual inspections and offline testing, or simple current and voltage threshold protection. Due to the long length of submarine cables, harsh environments, and severe near-field electromagnetic interference, these methods generally suffer from problems such as high non-invasive measurement intensity, low signal-to-noise ratios caused by strong electromagnetic interference, and insufficient intelligence.

[0005] In recent years, although some studies have proposed using wavelet transform and convolutional neural networks for feature extraction and intelligent diagnosis of grounding faults in terrestrial high-voltage cables, achieving high fault location accuracy and diagnosis speed in 10–35kV systems, these methods are usually based on single-point measurements of phase voltage and phase current, lacking modeling of the grounding current distribution mechanism of submarine cable sheaths and armor, and cannot be directly applied to the complex three-conductor coupling scenario of single-core submarine cables.

[0006] In summary, it is necessary to propose a non-invasive online detection method that simultaneously considers the current distribution characteristics of the sheath / armor of submarine single-core cables and the intelligent diagnostic capabilities of deep learning, so as to achieve robust and high-precision identification of grounding loop status and grounding faults in environments with strong electromagnetic interference. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention enables non-invasive online detection without altering the grounding structure. By combining a core-sheath-armor three-conductor π-type equivalent circuit model with a deep learning model, it improves the sensitivity and interpretability of fault diagnosis. In environments with strong electromagnetic interference, it can achieve robust and high-precision identification of the grounding circuit status and various types of grounding faults in submarine single-core transmission cables.

[0008] To achieve the above objectives, this invention provides an artificial intelligence-based grounding fault detection method for submarine single-core power transmission cables, comprising the following steps: Step S1: Arrange coupling capacitor clips on the outside of the sheath and / or armor of the submarine single-core cable. When no high-frequency injection signal is injected, collect the background voltage signal and / or background current signal at each measuring point through the coupling capacitor clips; perform spectrum analysis on the background voltage signal and / or background current signal, select the frequency band far away from the power frequency and integer multiples of the power frequency as the candidate injection frequency band, and determine the frequency of the high-frequency injection signal within the candidate injection frequency band; Step S2: A high-frequency injection signal is injected into the sheath and / or armor on which the coupling capacitor clips are arranged through a high-frequency signal source. The main input signals of loop voltage and loop current, as well as the reference noise signal, are sampled using Kelvin connection method. An improved adaptive filter is constructed to cancel the noise of the main input signal to obtain the enhanced signal component. At the same time, the error signal and adaptive step size factor of the improved adaptive filter at each moment during the noise cancellation process are recorded. The sequence of changes of the error signal and the adaptive step size factor over time is used as the adaptive filter state feature. Step S3: In the digital domain, the enhanced signal components are digitally quadrature demodulated to obtain the amplitude and phase of the high-frequency voltage and the amplitude and phase of the high-frequency current at each measurement point; the equivalent impedance and grounding resistance of the grounding loop are calculated based on the amplitude and phase of the high-frequency voltage and the amplitude and phase of the high-frequency current; simultaneously, based on the π-type equivalent circuit model of the core-sheath-armor three conductors of the submarine single-core cable, the sheath grounding current distribution ratio and the armor grounding current distribution ratio are calculated. Step S4: Combine the grounding resistance, the sheath grounding current distribution ratio, the armor grounding current distribution ratio, the amplitude and phase of the high-frequency voltage and the amplitude and phase of the high-frequency current, and the adaptive filtering state features with the time-domain statistical features, frequency-domain energy features, and time-frequency domain statistical features within the same analysis time window to construct a multi-channel feature tensor; input the multi-channel feature tensor into the trained convolutional-attention deep learning model to output the classification result and classification probability of the grounding condition; Step S5: Based on the classification probability output by the trained convolutional-attention deep learning model, the grounding resistance, the sheath grounding current distribution ratio, and the deviation of the armor grounding current distribution ratio from the normal reference value, a comprehensive judgment is made. When it is determined that there is a grounding fault or a grounding structure defect, an alarm message is output.

[0009] Furthermore, the improved adaptive filter in step S2 is an improved normalized least mean square adaptive filter, and the weight update formula of the improved normalized least mean square adaptive filter is: ; in, For the first The weight vector of the improved normalized minimum mean square adaptive filter at time step. For the first The weight vector of the improved normalized minimum mean square adaptive filter at time step. For the first The adaptive step size factor at time, For the first The error signal between the main input signal at time t and the output of the improved normalized minimum mean square adaptive filter. For the first The reference noise vector at time step, For the first The squared energy of the reference noise vector at time step [time]. This is the regularization parameter.

[0010] Furthermore, the series impedance matrix per unit length of the π-type equivalent circuit model of the core-sheath-armor three conductors of the submarine single-core cable described in step S3... and parallel admittance matrix Represented as: ; in, For series impedance matrix, This is the impedance per unit length between wire cores. This is the series impedance per unit length from the conductor to the sheath. The series impedance per unit length from the conductor to the armor is given. This is the series impedance per unit length from the sheath to the conductor. The series impedance per unit length from sheath to sheath. The series impedance per unit length from the sheath to the armor. The series impedance per unit length from the armor to the conductor is given. This is the series impedance per unit length from armor to sheath. The series impedance per unit length from armor to armor is given. For parallel admittance matrix, The parallel admittance per unit length of the wire core is... For the parallel admittance per unit length of the sheath, The parallel admittance per unit length of the armor is given; the series impedance matrix and the parallel admittance matrix are obtained by electromagnetic transient simulation calculation based on the conductor and insulation material parameters and structural dimensions of the submarine single-core cable.

[0011] Furthermore, the convolutional-attention deep learning model described in step S4 sequentially includes a one-dimensional convolutional layer, a pooling layer, an attention layer, and a fully connected layer.

[0012] Furthermore, the attention layer in the convolutional-attention deep learning model described in step S4 employs a multi-head self-attention mechanism.

[0013] Furthermore, the spectrum analysis in step S1 is achieved by performing a fast Fourier transform on the background voltage signal and / or background current signal.

[0014] Furthermore, the classification results of the grounding conditions in step S4 include: normal operating conditions, metallic grounding faults, transitional grounding faults, intermittent grounding faults, and abnormal grounding resistance.

[0015] Furthermore, the time-domain statistical features mentioned in step S4 include: the changes in the sheath grounding current distribution ratio and the armor grounding current distribution ratio over time, as well as the mean and variance of the grounding resistance; The frequency domain energy characteristics include: the frequency domain energy distribution and harmonic content obtained by spectral analysis of the enhanced signal components; The time-frequency domain statistical features include: the time-frequency local energy distribution obtained by wavelet packet decomposition of the enhanced signal components.

[0016] Furthermore, in step S5, when the comprehensive discrimination determines that a grounding structure defect exists, the following defect location determination is further performed: The deviation of the sheath grounding current distribution ratio from its normal reference value is recorded as the sheath deviation, and the deviation of the armor grounding current distribution ratio from its normal reference value is recorded as the armor deviation. The sheath deviation and the armor deviation are compared: when the ratio of the sheath deviation to the armor deviation exceeds a preset first ratio threshold, the grounding structure defect is determined to be located on the sheath side; when the ratio of the armor deviation to the sheath deviation exceeds a preset second ratio threshold, the grounding structure defect is determined to be located on the armor side; otherwise, the degree of the grounding structure defect on the sheath side and the armor side is determined to be equivalent. The alarm information includes information on the location of the grounding structure defect.

[0017] Furthermore, the adaptive filtering state features include: Within the analysis time window, the instantaneous amplitude, variance, and peak-to-peak value of the error signal are calculated to reflect the intensity of non-steady-state interference in the main input signal. Within the analysis time window, the mean and number of jumps of the adaptive step size factor are calculated to reflect the convergence state changes of the improved normalized minimum mean square adaptive filter. When the classification result of the grounding condition is an intermittent grounding fault, the error signal exhibits periodic pulses and the adaptive step size factor undergoes synchronous jumps; the adaptive filtering state features capture the periodic pulses and synchronous jumps to provide fault transient discrimination basis for the convolution-attention deep learning model.

[0018] Compared with the prior art, the present invention has at least the following beneficial effects: (1) Non-invasive online detection: High-frequency signals are injected through capacitive coupling and measured using Kelvin connection. Without disconnecting the original grounding lead of the cable, the high-frequency response signal of the grounding circuit can be obtained online, realizing real-time monitoring of the grounding status of submarine single-core cables.

[0019] (2) High-precision parameter estimation under strong interference: By using the improved normalized minimum mean square adaptive filter and digital quadrature demodulation technology, strong electromagnetic interference such as power frequency and its higher harmonics can be effectively suppressed, thereby accurately estimating the equivalent impedance and grounding resistance of the grounding loop, and improving the signal-to-noise ratio and result stability of the measurement.

[0020] (3) Integration of physical model and artificial intelligence: The sheath and armor grounding current distribution ratio calculated by the three conductor π-type equivalent circuit model of the core-sheath-armor of submarine single-core cable is combined with the convolution-attention deep learning model to achieve highly robust identification of grounding resistance anomalies and multiple types of grounding faults, and the diagnostic results are more physically interpretable.

[0021] (4) Applicable to long-distance submarine power transmission projects: The algorithm has a moderate amount of computation and can be deployed on edge computing devices at offshore booster stations or onshore landing stations. It is suitable for engineering applications and promotion of 110kV and above long-distance single-core submarine cables. Attached Figure Description

[0022] Figure 1 This is a flowchart of an artificial intelligence-based grounding fault detection method for submarine single-core power transmission cables, provided as an embodiment of the present invention.

[0023] Figure 2 This is a schematic diagram of the π-type equivalent circuit model of a three-conductor system consisting of a core, sheath, and armor. Detailed Implementation

[0024] The embodiments of the present invention can be combined with Figure 1 The method flowchart shown is provided for understanding. The detection system in this embodiment includes: a high-frequency signal source, a coupling capacitor clamp, a sampling resistor, a data acquisition and control unit, a signal processing and deep learning diagnostic unit, and an alarm and human-computer interaction module. The signal processing and deep learning diagnostic unit includes a signal processing module, a feature fusion module, a model training module, and an online diagnostic module. Specific implementation details are as follows: Step S1, multi-point non-invasive measurement and background spectrum analysis includes: Step S1.1: Several coupling capacitor clips are arranged outside the sheath and / or armor of the submarine single-core cable on both the offshore and onshore substation sides. The coupling capacitor clips are isolated from the outer sheath of the submarine single-core cable by insulating material, forming a non-intrusive high-frequency injection signal coupling channel.

[0025] Step S1.2: Without injecting a high-frequency signal, the data acquisition and control unit synchronously samples the background voltage signal and / or background current signal at each measurement point to obtain a background noise sequence. The sampling frequency is, for example, set to... This ensures that the high-frequency injection signal band can be covered and that transient fault characteristics can be captured.

[0026] Step S1.3, process the acquired background noise sequence Perform a Fast Fourier Transform (FFT) to obtain its spectrum. A frequency band far removed from the power frequency (50Hz) and its integer multiples of harmonics is selected as the candidate injection band. Within this band, frequencies with lower energy and relatively flat spectra are chosen. As the frequency of the high-frequency injection signal, it is necessary to ensure that the spectrum of the high-frequency injection signal is separated from the background noise as much as possible.

[0027] Step S2, high-frequency injection signal and adaptive noise cancellation includes: Step S2.1: The high-frequency signal source outputs a high-frequency injection signal, which is injected through a coupling capacitor clip at the marine or land end into a closed loop formed by the sheath and / or armor of the coupling capacitor clip, the sampling resistor, and the grounding point, forming a loop current and a loop voltage. Due to the distributed capacitance and loop inductance of the actual loop, this loop constitutes an LCR filter network, which has a strong suppression capability for power frequency and low-to-mid-order harmonics in the low-frequency band, thus helping to reduce power frequency interference.

[0028] In step S2.2, the data acquisition and control unit uses Kelvin connection to perform differential sampling at the high-frequency power supply output and sampling resistor to obtain the main input signal and reference noise signal at each measurement point. Based on the reference noise vector and error signal, an improved Normalized Least Mean Square (NLMS) adaptive filter is constructed to perform adaptive noise cancellation. The weight vector is iteratively updated until the variance of the error signal converges to below the threshold.

[0029] The output signal of the improved Normalized Least Mean Square (NLMS) adaptive filter The formula is as follows: ; in, For the first The transpose of the weight vector of the improved normalized minimum mean square adaptive filter at time step. For the first The reference noise vector at time step; ; in, For the first The error signal between the main input signal at time t and the output of the improved normalized minimum mean square adaptive filter. Main input signal; The weight update formula for the improved normalized least mean square adaptive filter is as follows: ; in, For the first The weight vector of the improved normalized minimum mean square adaptive filter at time step. For the first The weight vector of the improved normalized minimum mean square adaptive filter at time step. For the first The adaptive step size factor at time, For the first The squared energy of the reference noise vector at time step [time]. This is the regularization parameter.

[0030] Step S2.3: The error signal obtained by the improved Normalized Least Mean Square (NLMS) adaptive filter is used as the enhanced signal component. Its signal-to-noise ratio is significantly improved compared with the original signal, laying the foundation for subsequent digital quadrature demodulation and parameter estimation.

[0031] Step S2.4, Recording and Extraction of Adaptive Filtering State Features: During the adaptive filter iteration process described in step S2.2, the error signal and adaptive step size factor at each sampling time are recorded synchronously, and the sequence of changes in the error signal and adaptive step size factor over time is used as the adaptive filter state characteristics. Within each analysis time window, the instantaneous amplitude, variance, and peak-to-peak value of the error signal are calculated to characterize the intensity of non-steady-state interference in the main input signal; the mean and number of jumps of the adaptive step size factor are calculated to characterize the convergence state changes of the improved normalized minimum mean square adaptive filter. The adaptive filter state characteristics are used to provide fault transient discrimination information for subsequent grounding condition identification.

[0032] Step S3, digital quadrature demodulation and sheath / armor current distribution ratio include: Step S3.1, in the signal processing module of the signal processing and deep learning diagnostic unit, based on the frequency of the high-frequency injected signal... and sampling period In the digital domain, a sinusoidal reference signal and a cosine reference signal, in phase and frequency with the high-frequency injected signal, are constructed. The enhanced signal components are mixed with the sinusoidal and cosine reference signals respectively and low-pass filtered. Through digital quadrature demodulation, the amplitude and phase of the high-frequency voltage and the amplitude and phase of the high-frequency current at each measurement point are obtained. Based on the amplitude and phase of the high-frequency voltage and the amplitude and phase of the high-frequency current, the equivalent impedance and grounding resistance of the grounding loop are calculated.

[0033] The formulas for the sine reference signal and the cosine reference signal are as follows: ; in, For the first The sinusoidal reference signal at time [time] Angular frequency, For the first The cosine reference signal at time; The amplitude and phase of the demodulated signal are given by the following formulas: ; Where A is the amplitude of the demodulated signal. I represents the phase of the demodulated signal, where I is the in-phase component and Q is the quadrature component.

[0034] The equivalent impedance of the grounding loop is given by the following formula: ; in, This is the equivalent impedance of the grounding loop. This refers to the amplitude of the high-frequency voltage. This represents the amplitude of the high-frequency current. The imaginary unit, It is a phase factor used to represent the phase relationship between high-frequency voltage and high-frequency current.

[0035] Grounding resistance, the formula is as follows: ; in, For grounding resistance, The equivalent impedance of the grounding loop The real part, This is the sampling resistor.

[0036] Step S3.2: Input the submarine cable structural parameters (resistivity and relative permeability of the core, sheath, and armor materials, as well as resistivity and relative permittivity of the main insulation, shielding layer, and filling layer) into the electromagnetic transient simulation program (ATP-EMTP) or equivalent calculation program to obtain the series impedance matrix. and parallel admittance matrix Based on this, a π-type equivalent circuit model of a three-conductor system consisting of a core, sheath, and armor is established.

[0037] Based on the core-sheath-armor three-conductor π-type equivalent circuit model of the aforementioned submarine single-core cable (see...), Figure 2 The unit-length series impedance matrix of this model. and parallel admittance matrix Represented as: ; in, For series impedance matrix, This is the impedance per unit length between wire cores. This is the series impedance per unit length from the conductor to the sheath. The series impedance per unit length from the conductor to the armor is given. This is the series impedance per unit length from the sheath to the conductor. The series impedance per unit length from sheath to sheath. The series impedance per unit length from the sheath to the armor. The series impedance per unit length from the armor to the conductor is given. This is the series impedance per unit length from armor to sheath. The series impedance per unit length from armor to armor is given. For parallel admittance matrix, The parallel admittance per unit length of the wire core is... For the parallel admittance per unit length of the sheath, The parallel admittance per unit length of the armor.

[0038] Step S3.3: Solve the equations of the three-conductor network simultaneously to determine the core current at each cross-section. Sheath current and armored current Calculate the sheath grounding current distribution ratio (at the marine end and onshore end), the armor grounding current distribution ratio, and the sheath and armor grounding loss power.

[0039] Step S4, multi-channel feature tensor and deep learning intelligent diagnosis includes: Step S4.1: In the feature fusion module of the signal processing and deep learning diagnostic unit, the physical features obtained in real time online in step S3 (grounding resistance, sheath current distribution ratio, armor current distribution ratio, amplitude and phase of high-frequency voltage, amplitude and phase of high-frequency current), the adaptive filtering state features, and the time-domain statistical features of the signal within the same analysis time window (the changes of the sheath grounding current distribution ratio and the armor grounding current distribution ratio over time, and the mean and variance of the grounding resistance), frequency-domain energy features (frequency-domain energy distribution and harmonic content obtained by spectral analysis of the enhanced signal components), and time-frequency domain statistical features (time-frequency local energy distribution obtained by wavelet packet decomposition of the enhanced signal components) are fused and standardized to construct a multi-channel feature tensor. .in, For real numbers, This represents the total number of feature channels. This represents the length of the time series.

[0040] Step S4.2: In the model training module of the signal processing and deep learning diagnostic unit, a convolutional-attention deep learning model is trained using the training labeled dataset. The cross-entropy loss function is used as the target, and the network parameters are updated through backpropagation and stochastic gradient descent algorithms to finally obtain the trained convolutional-attention deep learning model. This model consists of multi-layer one-dimensional convolutional and pooling structures, several attention modules, and a fully connected classifier, used to extract local temporal features and model the global correlation between different features. The training labeled dataset is constructed by extracting grounding resistance, sheath grounding current distribution ratio, armored grounding current distribution ratio, amplitude and phase of high-frequency voltage, amplitude and phase of high-frequency current, adaptive filtering state features, time-domain statistical features, frequency-domain energy features, and time-frequency domain features from historical operating data and simulation-generated data samples, and constructing a multi-channel feature tensor. The multi-channel feature tensor constructed using historical operating data and simulation-generated data samples, along with their corresponding grounding condition category labels, are used as training samples.

[0041] Step S4.3: In the online diagnostic module of the signal processing and deep learning diagnostic unit, the multi-channel feature tensor constructed in step S4.1 is... Input the data into the convolutional-attention deep learning model trained in step S4.2, and output a fault category probability vector. The fault category probability vector represents the classification probability of each classification result in the grounding condition. The classification results include: normal operating condition, metallic grounding fault, transitional grounding fault, intermittent grounding fault, and abnormal grounding resistance.

[0042] Step S5, Fault Alarm and Decision Support: Step S5.1: When the fault category probability vector output by the trained convolutional-attention deep learning model exceeds the set threshold, and indicators such as grounding resistance, sheath current distribution, and armor current distribution ratio deviate significantly from the normal range, the alarm and human-machine interaction module determines that a grounding fault or grounding structure defect has occurred, and issues alarm information to the local human-machine interface of the offshore / onshore substation and the remote monitoring center, prompting maintenance personnel to conduct further inspections or take measures such as load reduction or shutdown.

[0043] Step S5.2: When the comprehensive judgment determines that a grounding structure defect exists, the following defect location determination is further performed: Distribution ratio of sheath grounding current The deviation from its normal reference value is recorded as the sheath deviation, and the armor grounding current distribution ratio is calculated. The deviation from its normal reference value is denoted as the armor deviation. The sheath grounding current distribution ratio under normal operating conditions is the reference value. The armored grounding current distribution ratio is based on the following value: The deviations of the sheath and the armor are respectively: ; ; in, For the first The sheath deviation within each analysis time window. For the first Armor deviation within each analysis time window.

[0044] The ratios of sheath deviation to armor deviation, and the ratios of armor deviation to sheath deviation, are calculated. When the ratio of sheath deviation to armor deviation exceeds a preset first threshold, the grounding structure defect is determined to be located on the sheath side; when the ratio of armor deviation to sheath deviation exceeds a preset second threshold, the grounding structure defect is determined to be located on the armor side; otherwise, the degree of grounding structure defect is determined to be equivalent on both the sheath and armor sides. The alarm information includes the location determination information of the grounding structure defect.

Claims

1. An artificial intelligence ground fault detection method for a submarine single-core power transmission cable, characterized by, The methods include: Step S1: Arrange coupling capacitor clips on the outside of the sheath and / or armor of the submarine single-core cable. When no high-frequency injection signal is injected, collect the background voltage signal and / or background current signal at each measuring point through the coupling capacitor clips; perform spectrum analysis on the background voltage signal and / or background current signal, select the frequency band far away from the power frequency and integer multiples of the power frequency as the candidate injection frequency band, and determine the frequency of the high-frequency injection signal within the candidate injection frequency band; Step S2: A high-frequency injection signal is injected into the sheath and / or armor on which the coupling capacitor clips are arranged through a high-frequency signal source. The main input signals of loop voltage and loop current, as well as the reference noise signal, are sampled using Kelvin connection method. An improved adaptive filter is constructed to cancel the noise of the main input signal to obtain the enhanced signal component. At the same time, the error signal and adaptive step size factor of the improved adaptive filter at each moment during the noise cancellation process are recorded. The sequence of changes of the error signal and the adaptive step size factor over time is used as the adaptive filter state feature. Step S3: In the digital domain, the enhanced signal components are digitally quadrature demodulated to obtain the amplitude and phase of the high-frequency voltage and the amplitude and phase of the high-frequency current at each measurement point; the equivalent impedance and grounding resistance of the grounding loop are calculated based on the amplitude and phase of the high-frequency voltage and the amplitude and phase of the high-frequency current; simultaneously, based on the π-type equivalent circuit model of the core-sheath-armor three conductors of the submarine single-core cable, the sheath grounding current distribution ratio and the armor grounding current distribution ratio are calculated. Step S4: Combine the grounding resistance, the sheath grounding current distribution ratio, the armor grounding current distribution ratio, the amplitude and phase of the high-frequency voltage and the amplitude and phase of the high-frequency current, and the adaptive filtering state features with the time-domain statistical features, frequency-domain energy features, and time-frequency domain statistical features within the same analysis time window to construct a multi-channel feature tensor; input the multi-channel feature tensor into the trained convolutional-attention deep learning model to output the classification result and classification probability of the grounding condition; Step S5: Based on the classification probability output by the trained convolutional-attention deep learning model, the grounding resistance, the sheath grounding current distribution ratio, and the armor grounding current distribution ratio, a comprehensive judgment is made on their deviation from the normal reference value. When it is determined that there is a grounding fault or a grounding structure defect, an alarm message is output. The adaptive filtering state features include: Within the analysis time window, the instantaneous amplitude, variance, and peak-to-peak value of the error signal are calculated to reflect the intensity of non-steady-state interference in the main input signal. Within the analysis time window, the mean and number of jumps of the adaptive step size factor are calculated to reflect the convergence state changes of the improved normalized minimum mean square adaptive filter. When the classification result of the grounding condition is an intermittent grounding fault, the error signal exhibits periodic pulses and the adaptive step size factor undergoes synchronous jumps; the adaptive filtering state features capture the periodic pulses and synchronous jumps to provide fault transient discrimination basis for the convolution-attention deep learning model.

2. The artificial intelligence grounding fault detection method for submarine single-core power transmission cables according to claim 1, characterized in that, The improved adaptive filter mentioned in step S2 is an improved normalized least mean square adaptive filter, and the weight update formula of the improved normalized least mean square adaptive filter is: ； in, For the first The weight vector of the improved normalized minimum mean square adaptive filter at time step. For the first The weight vector of the improved normalized minimum mean square adaptive filter at time step. For the first The adaptive step size factor at time, For the first The error signal between the main input signal at time t and the output of the improved normalized minimum mean square adaptive filter. For the first The reference noise vector at time step, For the first The squared energy of the reference noise vector at time step [time]. This is the regularization parameter.

3. The artificial intelligence grounding fault detection method for submarine single-core power transmission cables according to claim 1, characterized in that, The series impedance matrix per unit length of the π-type equivalent circuit model of the core-sheath-armor three conductors of the submarine single-core cable described in step S3. and parallel admittance matrix Represented as: ； in, For series impedance matrix, This is the impedance per unit length from core to core. This is the series impedance per unit length from the conductor to the sheath. The series impedance per unit length from the conductor to the armor is given. This is the series impedance per unit length from the sheath to the conductor. The series impedance per unit length from sheath to sheath. The series impedance per unit length from the sheath to the armor. The series impedance per unit length from the armor to the conductor is given. This is the series impedance per unit length from armor to sheath. The series impedance per unit length from armor to armor is given. For parallel admittance matrix, The parallel admittance per unit length of the wire core is... For the parallel admittance per unit length of the sheath, The parallel admittance per unit length of the armor is given; the series impedance matrix and the parallel admittance matrix are obtained by electromagnetic transient simulation calculation based on the conductor and insulation material parameters and structural dimensions of the submarine single-core cable.

4. The artificial intelligence grounding fault detection method for submarine single-core power transmission cables according to claim 1, characterized in that, The convolutional-attention deep learning model described in step S4 includes, in sequence, a one-dimensional convolutional layer, a pooling layer, an attention layer, and a fully connected layer.

5. The artificial intelligence grounding fault detection method for submarine single-core power transmission cables according to claim 4, characterized in that, The attention layer in the convolutional-attention deep learning model described in step S4 employs a multi-head self-attention mechanism.

6. The artificial intelligence grounding fault detection method for submarine single-core power transmission cables according to claim 1, characterized in that, The spectrum analysis in step S1 is achieved by performing a fast Fourier transform on the background voltage signal and / or background current signal.

7. The artificial intelligence grounding fault detection method for submarine single-core power transmission cables according to claim 1, characterized in that, The classification results of the grounding conditions in step S4 include: normal operating conditions, metallic grounding faults, transitional grounding faults, intermittent grounding faults, and abnormal grounding resistance.

8. The artificial intelligence grounding fault detection method for submarine single-core power transmission cables according to claim 1, characterized in that, The time-domain statistical features mentioned in step S4 include: the changes in the sheath grounding current distribution ratio and the armor grounding current distribution ratio over time, as well as the mean and variance of the grounding resistance; The frequency domain energy characteristics include: the frequency domain energy distribution and harmonic content obtained by spectral analysis of the enhanced signal components; The time-frequency domain statistical features include: the time-frequency local energy distribution obtained by wavelet packet decomposition of the enhanced signal components.

9. The artificial intelligence grounding fault detection method for submarine single-core power transmission cables according to claim 1, characterized in that, In step S5, when the comprehensive discrimination determines that a grounding structure defect exists, the following defect location determination is further performed: The deviation of the sheath grounding current distribution ratio from its normal reference value is recorded as the sheath deviation, and the deviation of the armor grounding current distribution ratio from its normal reference value is recorded as the armor deviation. Compare the sheath deviation with the armor deviation: when the ratio of the sheath deviation to the armor deviation exceeds a preset first ratio threshold, it is determined that the grounding structure defect is located on the sheath side; When the ratio of the armor deviation to the sheath deviation exceeds a preset second ratio threshold, it is determined that the grounding structure defect is located on the armor side; otherwise, it is determined that the grounding structure defect is of the same degree on both the sheath side and the armor side. The alarm information includes information on the location of the grounding structure defect.

Images