A cable fault intelligent diagnosis method and system based on multi-algorithm fusion

Abstract

The application discloses a kind of cable fault intelligent diagnosis method and system based on multi-algorithm fusion, it is related to cable fault diagnosis technical field.For the problems of insufficient signal acquisition, large noise interference, low feature recognition and insufficient accuracy caused by diagnosis model hyperparameter dependence on artificial experience in existing cable fault diagnosis, the present application includes laying multiple types of sensors at multiple key positions to synchronously collect signals;Using sparrow optimization algorithm and variational mode decomposition to jointly reduce noise, selecting effective components through correlation coefficient and combining adaptive wavelet threshold denoising, reconstructing the denoised signal;Extracting multi-dimensional features from time domain, frequency domain and time-frequency domain;Filtering core features through Relieff and mRMR joint screening and normalizing;Taking normalized features as input, using grey wolf optimization algorithm to globally optimize capsule network hyperparameters, constructing a fault diagnosis model, and outputting cable operating status and fault confidence probability.The present application improves signal denoising effect, feature representation ability and fault diagnosis accuracy.

Description

Technical Field

[0001] This invention relates to the field of cable fault diagnosis technology, and in particular to a method and system for intelligent cable fault diagnosis based on multi-algorithm fusion. Background Technology

[0002] As the core carrier of electrical energy transmission in power systems, cables are widely used in urban power distribution networks, industrial power supply, and rail transit. Their operating status directly affects the safety and reliability of the power system. However, during long-term service, cables are susceptible to faults such as wire breakage, partial discharge, single-phase grounding, and phase-to-phase short circuits due to factors such as insulation aging, mechanical damage, environmental corrosion, and overload operation. In severe cases, these faults may lead to power outages or even electrical fires.

[0003] Traditional cable fault diagnosis methods have many limitations: First, signal acquisition coverage is insufficient. Existing diagnostic systems often only deploy sensors at a single location on the cable, or only collect a single physical quantity (such as only current signals or only temperature signals), making it difficult to achieve synchronous monitoring of multiple physical quantities along the entire cable line. This "single-point, single-parameter" acquisition mode cannot cover fault characteristic information throughout the entire cable life cycle, especially lacking the ability to detect early, weak faults, which easily leads to missed diagnoses and misdiagnoses.

[0004] Secondly, the on-site signal is severely affected by noise interference. Complex electromagnetic interference and environmental noise are common in cable operation sites, resulting in low signal-to-noise ratios for the acquired raw signals, which severely limits the accuracy of subsequent fault feature extraction. Existing noise reduction methods mostly employ single filtering strategies, such as using only wavelet denoising or only mode decomposition. These methods lack the ability to adaptively optimize the core parameters of variational mode decomposition and struggle to accurately distinguish between effective fault components and noise components, leading to unsatisfactory noise reduction results.

[0005] Third, the fault characteristic representation dimension is limited. Existing methods mostly extract feature values ​​from a single domain (such as only the time domain or only the frequency domain), which makes it difficult to comprehensively depict the subtle differences between different fault types. They also have limited ability to distinguish fault categories with similar time domain waveforms (such as partial discharge and single-phase grounding).

[0006] Fourth, the feature set contains a large amount of redundant information. The initially extracted multidimensional feature set often contains redundant features with high correlation and low discriminative power. These features not only increase the computational burden of the diagnostic model and reduce the real-time performance of the diagnosis, but may also interfere with the model's discrimination decision, leading to a decrease in diagnostic accuracy.

[0007] Fifth, the hyperparameters of diagnostic models rely on manual experience for setting. The hyperparameters of existing intelligent diagnostic models (such as neural network models) are mostly determined through manual trial and error or empirical rules, lacking a global optimization mechanism, making it difficult to guarantee the model's generalization performance and diagnostic accuracy under different operating conditions. Especially in capsule neural networks, key hyperparameters such as the temperature coefficient of the squash compression function, the number of iterations in dynamic routing, and the coupling coefficient scaling factor have a significant impact on model performance, but existing research has not yet provided a systematic optimization scheme. Summary of the Invention

[0008] The technical problem to be solved and the technical task proposed by this invention is to improve and refine existing technical solutions, and to provide a cable fault intelligent diagnosis method based on multi-algorithm fusion, so as to achieve accurate and efficient cable fault diagnosis. To this end, this invention adopts the following technical solution.

[0009] Firstly, a method for intelligent cable fault diagnosis based on multi-algorithm fusion is provided, which includes the following steps: S1: Install a sensor group including temperature sensor, acceleration sensor and current sensor at key locations of the cable to be diagnosed, and simultaneously collect multi-channel raw signals to build a fault diagnosis dataset including samples of various working states. S2: The multi-channel original signal is processed by a joint denoising algorithm combining the Sparrow Optimization algorithm and Variational Mode Decomposition to obtain multiple intrinsic mode components; the Pearson correlation coefficient between each intrinsic mode component and the original signal is calculated, the effective intrinsic mode components are selected according to the correlation coefficient threshold, and adaptive wavelet threshold denoising is performed on each effective intrinsic mode component to reconstruct the final denoised signal; S3: Based on the final denoised signal, extract feature values ​​from three dimensions: time domain, frequency domain, and time-frequency domain to form an initial multi-dimensional feature set; S4: The initial multi-dimensional feature set is filtered for core features by combining the ReliefF algorithm and the mRMR algorithm to obtain the final core feature set; S5: Normalize the core feature set; S6: Using the normalized core feature set as input, the Grey Wolf optimization algorithm is used to globally optimize the hyperparameters of the capsule neural network to construct a fault diagnosis model; when cable fault diagnosis is required, the data of the cable to be diagnosed is input into the fault diagnosis model, and the fault diagnosis model outputs the cable operating status and the corresponding fault confidence probability.

[0010] This technical solution deploys a sensor array containing temperature, acceleration, and current sensors at key locations on the cable to be diagnosed, enabling simultaneous acquisition of multi-channel raw signals. This covers fault characteristic information across the entire cable line, significantly enhancing the detection capability for early, subtle faults and effectively reducing the risk of missed or false diagnoses. A joint denoising algorithm combining the Sparrow Optimization algorithm and Variational Mode Decomposition (VMD) is used to process the raw signals. The Sparrow Optimization algorithm adaptively determines the optimal parameter combination for VMD, overcoming the shortcomings of traditional methods that rely on manual experience for parameter setting. Simultaneously, effective intrinsic mode components are screened using Pearson correlation coefficients, noise-dominant components are eliminated, and secondary processing is performed using adaptive wavelet threshold denoising, forming a multi-level joint denoising system. This system effectively preserves fault characteristic information while maximally suppressing environmental noise and electromagnetic interference. Feature values ​​are extracted from the time domain, frequency domain, and time-frequency domain to form an initial multi-dimensional feature set, which comprehensively characterizes the differentiated information of various faults from different perspectives. It has a stronger ability to distinguish between different fault types with similar time-domain waveforms (such as partial discharge and single-phase grounding). A combination of the ReliefF algorithm and the mRMR algorithm is used to select core features from the initial multi-dimensional feature set. The ReliefF algorithm evaluates the contribution of each feature to the differentiation of the fault category, while the mRMR algorithm minimizes the redundancy between features while ensuring the maximum correlation between features and the fault category. The combination of the two can effectively remove redundant and useless features while retaining highly discriminative features, reducing the computational burden of the diagnostic model and improving the real-time performance of the diagnosis. The Grey Wolf Optimization Algorithm is used to globally optimize the hyperparameters of the capsule neural network, avoiding the limitations of traditional methods that rely on manual trial and error or empirical rules to set hyperparameters. The Grey Wolf Optimization Algorithm has the characteristics of fast convergence speed and high optimization accuracy, and can find the optimal hyperparameter combination of the capsule neural network, thereby fully releasing the performance potential of the capsule network in the fault identification task. The GWO-CapsNet fault diagnosis model outputs the cable operating status and the corresponding fault confidence probability, not only providing the diagnostic conclusion, but also providing the model's confidence level of the conclusion, providing richer reference information for the subsequent decision-making of maintenance personnel, and enhancing the credibility and interpretability of the diagnostic results. The system systematically solves the technical problems of incomplete signal acquisition, poor anti-interference ability, low feature recognition, high feature redundancy, and insufficient model parameter optimization in existing technologies.

[0011] As a preferred technical means: In step S1, the key locations include the cable head end, the middle fault-prone section and the end; the multiple working states include normal state, open circuit, partial discharge, single-phase grounding and phase-to-phase short circuit; temperature signals, vibration signals and current signals during cable operation are collected synchronously by each group of sensors to form a multi-channel raw signal dataset, to create a cable fault diagnosis dataset, and the samples under each working state are divided into training set and test set according to the proportion.

[0012] This technical solution achieves blind-spot-free coverage of key nodes along the entire cable line, avoiding the omission of fault information due to single-point monitoring; it adopts five working states (normal, open circuit, partial discharge, single-phase grounding, and phase-to-phase short circuit) to cover the most important operating scenarios and fault types of cables, providing a complete labeling system for subsequent supervised learning; it divides the training set and test set proportionally to ensure the data independence of model training and validation, which is conducive to objectively evaluating the generalization performance of the model.

[0013] As a preferred technical means, step S2 specifically includes: S21: Set the sparrow population size, maximum number of iterations, proportion of discoverers, and proportion of vigilants; set the search range for the number of decomposition modes and the search range for the penalty factor in variational mode decomposition; each individual sparrow corresponds to a set of variational mode decomposition parameter combinations, including the number of decomposition modes and the penalty factor; S22: Based on the variational mode decomposition parameter combination corresponding to each individual sparrow, the original signal is subjected to variational mode decomposition to obtain multiple intrinsic eigenmode components. The original signal is equal to the sum of the intrinsic eigenmode components. S23: Calculate the Pearson correlation coefficient between each intrinsic mode component and the original signal, set a correlation coefficient threshold, retain effective intrinsic mode components with a correlation coefficient greater than or equal to the correlation coefficient threshold, and remove noise-dominant components with a correlation coefficient less than the correlation coefficient threshold. S24: For each effective intrinsic mode component after screening, use the Symlet4 wavelet basis function to perform 4-level discrete wavelet decomposition to obtain high-frequency wavelet coefficients and low-frequency wavelet coefficients. S25: The high-frequency wavelet coefficients are filtered using an adaptive soft threshold function. The processed wavelet coefficients are equal to the positive value obtained by multiplying the sign of the original wavelet coefficients by the absolute value of the original wavelet coefficients minus the adaptive threshold. The adaptive threshold is determined by the noise standard deviation, the signal length, and the maximum amplitude of the wavelet coefficients of the current intrinsic mode components. Its value decreases adaptively as the amplitude of the wavelet coefficients increases. S26: Perform inverse wavelet transform on the low-frequency wavelet coefficients and the filtered high-frequency wavelet coefficients of each intrinsic eigenmode component to obtain the denoised intrinsic eigenmode components. S27: Construct the fitness function, which is equal to the signal-to-noise ratio multiplied by the first weight, minus the root mean square error multiplied by the second weight, plus the noise suppression ratio multiplied by the third weight; with the goal of maximizing the fitness function value, perform the position updates of the discoverer, followers, and guards, and determine whether the maximum number of iterations has been reached. If not, return to step S22 to continue iterating; after reaching the maximum number of iterations, output the optimal combination of variational mode decomposition parameters corresponding to the maximum fitness function value, including the optimal number of decomposition modes and the optimal penalty factor; S28: Based on the optimal number of decomposed modes and the optimal penalty factor, perform variational mode decomposition on the original signal to obtain the optimal number of intrinsic mode components; after filtering each optimal intrinsic mode component by Pearson correlation coefficient and adaptive wavelet threshold denoising, superimpose and reconstruct all the denoised optimal intrinsic mode components to obtain the final denoised signal.

[0014] This technical solution utilizes the Sparrow Optimization algorithm to globally optimize the number of decomposition modes and the penalty factor in Variational Mode Decomposition (VMD), solving the technical challenge of determining VMD parameters and avoiding the instability of noise reduction effects caused by manual parameter trial and error. By screening effective intrinsic mode components using the Pearson correlation coefficient, it can accurately distinguish between components dominated by fault characteristics and components dominated by noise from a quantitative perspective, which is more scientific and accurate than relying on empirical judgment or simple energy thresholding methods. The solution employs Symlet4 wavelet basis functions for four-level discrete wavelet decomposition combined with adaptive soft thresholding filtering. The Symlet4 wavelet possesses good approximate symmetry and approximate linear phase characteristics, which can maximize the preservation of signal waveform distortion during noise reduction. The adaptive threshold dynamically adjusts with the wavelet coefficient amplitude, providing greater suppression for small coefficients and less suppression for large coefficients, thus preserving more refined abrupt details in fault characteristics compared to fixed thresholding methods. A comprehensive fitness function is constructed that includes signal-to-noise ratio, root mean square error, and noise suppression ratio to overcome the noise reduction bias that may be caused by optimizing a single index (such as excessive signal smoothing if only the signal-to-noise ratio is optimized). The three indices counterbalance each other to ensure that the optimized VMD parameters can achieve the best balance between signal fidelity and noise suppression.

[0015] As a preferred technical means: In step S4: The ReliefF algorithm is used to calculate the distinguishing ability of each feature for samples of different categories, obtain the feature weight, set the weight threshold, and select features with weights greater than or equal to the weight threshold to form the ReliefF preferred feature set. The mutual information between features and fault categories, as well as the mutual information between features, is calculated using the mRMR algorithm, and an objective function is constructed. The objective function is based on the maximum correlation criterion and the minimum redundancy criterion, where the maximum correlation criterion value is the average mutual information between each feature and fault category within the feature subset, and the minimum redundancy criterion value is the average mutual information between each feature within the feature subset. The objective function is equal to the maximum correlation criterion value minus the minimum redundancy criterion value. An incremental search method is used for iterative optimization to obtain the optimal mRMR feature set. The union of the ReliefF preferred feature set and the mRMR preferred feature set is taken as the final core feature set.

[0016] The ReliefF algorithm excels at evaluating the contribution of a single feature to class differentiation, while the mRMR algorithm excels at filtering redundancy among features. Both filter features from different perspectives, and taking the union can achieve the dual goals of high discriminative power and low redundancy. Compared with a single algorithm, the filtering is more comprehensive and reliable, avoiding the omission of important features and eliminating redundant features.

[0017] As a preferred technical means: in step S6, the capsule neural network includes a convolutional layer, a main capsule layer, a category capsule layer, and a decoder reconstruction layer; Both the main capsule layer and the category capsule layer use a squash compression function with a temperature coefficient to perform nonlinear compression on the capsule output vector. Specifically, the compressed output vector of the j-th capsule is equal to the square of the L2 norm of the total input vector of the capsule divided by the sum of the square of the L2 norm and the square of the temperature coefficient, and then multiplied by the total input vector of the capsule divided by its L2 norm. The capsule neural network adopts an improved dynamic routing mechanism to realize information transmission from lower-level capsules to higher-level capsules. The routing logic value update method is as follows: the routing logic value update from lower-level capsule i to higher-level capsule j is its original value plus the routing scaling factor multiplied by the dot product of the prediction vector of lower-level capsule i to higher-level capsule j after affine transformation and the output vector of higher-level capsule j.

[0018] Traditional squash compression functions use a fixed temperature coefficient of 1, lacking flexibility. This solution introduces an adjustable temperature coefficient, allowing the sensitive region of the compression function to be adaptively adjusted according to the specific task, avoiding loss of effective information due to excessive compression or noise amplification due to insufficient compression. A routing scaling factor is introduced to adjust the step size of the coupling coefficient update, effectively suppressing the violent oscillations of the coupling coefficient in the early stages of training, which is common in traditional dynamic routing mechanisms. This makes the routing convergence process smoother and more stable, improving the robustness of model training and the final classification accuracy.

[0019] As a preferred technical approach: In step S6, the specific steps for globally optimizing the hyperparameters of the capsule neural network using the gray wolf optimization algorithm include: The gray wolf population size and maximum number of iterations are set. The learning rate, weight decay coefficient, dynamic routing iterations, routing scaling factor, capsule dimension of the main capsule layer, capsule dimension of the category capsule layer, batch size, temperature coefficient of the squash compression function, and reconstruction loss weight of the capsule neural network are encoded into multi-dimensional position vectors of individual gray wolves. Continuous parameters are encoded with real numbers, and discrete parameters are encoded with integers. Construct a fitness function, which is equal to the validation set fault classification accuracy multiplied by the first weight, plus the validation set F1 score multiplied by the second weight, plus the reciprocal of the validation set total loss function value plus 1 multiplied by the third weight. The α wolf with the best fitness, the β wolf with the second best fitness, and the δ wolf with the third best fitness are determined based on the fitness ranking. The α wolf, β wolf and δ wolf jointly guide the remaining gray wolf individuals to update their positions. After iterating to the maximum number of iterations, the optimal hyperparameter combination corresponding to the maximum fitness value is output. The optimal hyperparameter combination is used to perform extended training of the capsule neural network, and random translation, rotation data augmentation and L2 regularization are introduced to obtain the optimal GWO-CapsNet fault diagnosis model.

[0020] Capsule neural networks contain multiple key hyperparameters, and it is difficult to find the globally optimal combination through manual parameter tuning. The Grey Wolf optimization algorithm is characterized by its simple structure, fast convergence speed, and strong global optimization ability, enabling it to efficiently search for optimal solutions in high-dimensional parameter spaces. Compared with grid search or random search, it has significantly improved efficiency and accuracy. The fitness function integrates three metrics: accuracy, F1 score, and loss function. Accuracy reflects the overall classification correctness, the F1 score balances precision and recall (particularly suitable for scenarios with imbalanced fault samples), and the loss function reflects the model's convergence state. The combination of these three metrics ensures that the selected hyperparameter combination performs well across multiple evaluation dimensions. Expanded rounds of refined training, combined with data augmentation (random translation and rotation) and L2 regularization, further enhance the model's generalization ability based on the optimal hyperparameters, effectively suppressing overfitting and making the model more robust in practical engineering applications.

[0021] As a preferred technical means: In step S3, the feature values ​​in the time domain include peak value, mean, standard deviation, variance, peak-to-peak value, kurtosis, skewness, waveform factor, peak factor, kurtosis factor, impulse factor, and mean factor; the feature values ​​in the frequency domain include centroid frequency, mean square frequency, root mean square frequency, frequency variance, and frequency standard deviation; the feature values ​​in the time-frequency domain include envelope entropy, sample entropy, approximate entropy, permutation entropy, and normalized wavelet packet energy of each frequency band after wavelet packet transformation.

[0022] Time-domain features reflect the amplitude distribution and waveform morphology of a signal, and are sensitive to amplitude variations; frequency-domain features reflect the frequency structure of a signal, and are sensitive to periodic fault components; time-frequency domain features (entropy features and wavelet packet energy) reflect the complexity of the signal's energy distribution and local time-varying characteristics, and are sensitive to non-stationary fault signals. These three domain features complement each other, forming a three-dimensional representation of the fault signal, enabling the full extraction of fault information hidden within the signal. Entropy features (envelope entropy, sample entropy, approximate entropy, and permutation entropy) characterize the complexity and regularity of the signal from different dimensions, exhibiting differentiated response sensitivity to different types of faults. Combinations of multiple entropy values ​​significantly improve the discriminative power of fault features.

[0023] As a preferred technical means: In step S5, the normalization processing method is as follows: the normalized feature value is equal to the original value of the feature in the sample minus the minimum value of the feature in all samples, and then divided by the maximum value of the feature in all samples minus the minimum value of the feature.

[0024] Normalization maps feature values ​​of different dimensions and orders of magnitude to the [0,1] interval, eliminating the impact of dimension differences on model training and avoiding the problem of large numerical features dominating the loss function and small numerical features being ignored. At the same time, it preserves the relative size relationship between each feature, and is simple to calculate and highly interpretable, making it suitable for scenarios with real-time requirements such as fault diagnosis.

[0025] Secondly, a cable fault intelligent diagnosis system based on multi-algorithm fusion is provided, which includes: The signal acquisition module is used to simultaneously acquire multi-channel raw signals at key locations of the cable to be diagnosed using a sensor group that includes temperature sensors, acceleration sensors, and current sensors, and to build a fault diagnosis dataset containing samples of various operating states. The joint noise reduction module is used to process the multi-channel original signal using a joint noise reduction algorithm that combines the sparrow optimization algorithm and variational mode decomposition. It filters the effective intrinsic mode components by Pearson correlation coefficient and combines them with adaptive wavelet thresholding to reconstruct the final noise-reduced signal. The feature extraction module is used to extract feature values ​​from the time domain, frequency domain, and time-frequency domain based on the final denoised signal to form an initial multi-dimensional feature set; The feature filtering module is used to filter the core features of the initial multi-dimensional feature set by combining the ReliefF algorithm and the mRMR algorithm to obtain the final core feature set. The normalization processing module is used to normalize the core feature set; The fault diagnosis module is used to globally optimize the hyperparameters of the capsule neural network using the gray wolf optimization algorithm to construct a fault diagnosis model. When cable fault diagnosis is required, the data of the cable to be diagnosed is input into the fault diagnosis model, and the fault diagnosis model outputs the cable operating status and the corresponding fault confidence probability.

[0026] The signal acquisition module deploys a sensor group containing three types of sensors—temperature, acceleration, and current—at key locations on the cable to be diagnosed, enabling simultaneous acquisition of multi-channel raw signals. This covers fault characteristic information across the entire cable line, significantly enhancing the detection capability for early, subtle faults and effectively reducing the risk of missed or false diagnoses. The joint noise reduction module employs a combined noise reduction algorithm that integrates the sparrow optimization algorithm and variational mode decomposition to process the raw signal. The sparrow optimization algorithm adaptively determines the optimal parameter combination for variational mode decomposition, eliminating noise-dominant components and retaining effective intrinsic mode components through Pearson correlation coefficients. A secondary fine-tuning process is then performed using adaptive wavelet threshold denoising, with the threshold dynamically adjusted according to the wavelet coefficient amplitude. Small coefficients are suppressed more significantly, while large coefficients are suppressed less. Through a multi-stage cascaded noise reduction system, environmental noise and electromagnetic interference are suppressed to the maximum extent while effectively preserving fault characteristic information. The feature extraction module extracts feature values ​​from three dimensions: time domain, frequency domain, and time-frequency domain. These three domain features complement each other, forming a three-dimensional representation of the fault signal. This allows for a comprehensive representation of the differentiated information of various faults from different perspectives, and provides stronger distinguishing ability for different fault types with similar time-domain waveforms. The feature selection module uses a combination of the ReliefF algorithm and the mRMR algorithm to select core features from the initial multi-dimensional feature set. Both algorithms select features from different angles, avoiding the omission of important features and eliminating redundant features, effectively reducing the computational burden of the diagnostic model and improving the real-time performance of the diagnosis. The fault diagnosis module uses the Grey Wolf optimization algorithm to globally optimize the hyperparameters of the capsule neural network. The capsule neural network, through vectorized representation and dynamic routing mechanisms, can effectively model the spatial hierarchical relationships between features. The Grey Wolf optimization algorithm avoids the limitations of traditional methods that rely on manual trial and error or empirical rules to set hyperparameters, and can search for the optimal hyperparameter combination of the capsule neural network, thereby fully releasing the performance potential of the capsule network in fault identification tasks.

[0027] As a preferred technical means: the multiple working states include normal state, open circuit, partial discharge, single-phase grounding, and phase-to-phase short circuit; the multi-dimensional feature values ​​include peak value, mean, standard deviation, variance, peak-to-peak value, kurtosis, skewness, waveform factor, peak factor, kurtosis factor, impulse factor, and mean factor in the time domain; centroid frequency, mean square frequency, root mean square frequency, frequency variance, and frequency standard deviation in the frequency domain; and envelope entropy, sample entropy, approximate entropy, permutation entropy, and normalized wavelet packet energy of each frequency band after wavelet packet transformation in the time and frequency domain.

[0028] Five states cover the most common operating conditions of cables in actual operation. This state coverage provides a complete labeling system for subsequent supervised learning, ensuring that the diagnostic model can effectively distinguish between normal operating conditions and various typical faults. A complete feature system is constructed from three dimensions: time domain, frequency domain, and time-frequency domain. At the time domain feature level, peak value and peak-to-peak value reflect the amplitude range of the signal, mean reflects the DC component level of the signal, standard deviation and variance reflect the dispersion of the signal, kurtosis and skewness reflect the probability distribution of the signal (kurtosis is sensitive to impulsive faults, and skewness is sensitive to waveform asymmetry). Dimensionless indicators such as waveform factor, peak factor, kurtosis factor, impulse factor, and mean factor reflect the waveform morphology characteristics of the signal, exhibiting good robustness to operating condition fluctuations such as load changes. At the frequency domain level, the centroid frequency reflects the dominant frequency position of the signal power spectrum, the mean square frequency and root mean square frequency reflect the central tendency of the signal frequency, and the frequency variance and frequency standard deviation reflect the dispersion of the signal spectrum. These frequency domain features have a keen response to periodic fault components (such as poor contact caused by mechanical vibration). At the time-frequency domain level, the envelope entropy is sensitive to the complexity of the signal envelope and is suitable for detecting amplitude modulation faults. The sample entropy can still maintain a stable estimate when the data length is short and is suitable for fault analysis under limited sample conditions. The approximate entropy measures the complexity of the signal sequence from a statistical perspective. The permutation entropy reveals the temporal structure regularity of the signal by comparing the magnitude order of adjacent values. Compared with the traditional wavelet transform, the wavelet packet transform can simultaneously perform fine decomposition of low-frequency and high-frequency parts and has a stronger ability to capture the time-varying features of non-stationary fault signals. The three types of features—time domain, frequency domain, and time-frequency domain—complement each other. Time domain features focus on amplitude distribution and waveform morphology, frequency domain features focus on spectral structure and frequency components, and time-frequency domain features focus on the local time-varying characteristics of non-stationary signals. These three types of features together constitute a holographic representation of fault signals from different physical perspectives, which can fully explore the multi-dimensional fault information hidden in the signal and significantly improve the distinguishability between different fault types.

[0029] Beneficial effects: 1. Full-position blind-spot-free monitoring: Multiple types of sensors are deployed at key locations at the beginning, middle and end of the cable to simultaneously collect temperature, acceleration and current signals, achieving full life cycle fault feature coverage, effectively capturing early weak faults, and reducing missed detection and diagnostic errors; 2. By using SSA to optimize VMD parameters and combining Pearson correlation coefficient screening with adaptive wavelet denoising, a multi-level noise reduction system is constructed, which significantly improves the signal-to-noise ratio while fully preserving effective information and suppressing environmental and electromagnetic noise. 3. Extract signal feature values ​​from multiple dimensions, including time domain, frequency domain, and time-frequency domain, to comprehensively characterize the differences between various faults and enhance the ability to distinguish fault features; 4. By integrating ReliefF and mRMR algorithms, key features are selected and redundancy is eliminated, ensuring the accuracy of fault identification and improving the real-time performance of model training and diagnosis. 5. The Grey Wolf algorithm is used to optimize the hyperparameters of the capsule neural network. Combined with improved routing and joint loss function, the fault confidence probability is output, providing a reliable basis for decision-making. Attached Figure Description

[0030] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0031] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings.

[0032] Example 1: This embodiment provides a cable fault intelligent diagnosis method based on multi-algorithm fusion.

[0033] This embodiment uses a 10kV cross-linked polyethylene (XLPE) cable in a city's power distribution network as the diagnostic object. The cable is 1000m long, and the fault-prone sections are set at 300m, 600m, and 800m from the beginning. The specific implementation steps are as follows: S1: Deploy a sensor array, including temperature, acceleration, and current sensors, at key locations on the cable to be diagnosed. Simultaneously acquire multi-channel raw signals to construct a fault diagnosis dataset containing samples of various operating states. The specific steps are as follows: S11: Each sensor group integrates a PT100 platinum resistance temperature sensor (-40℃~120℃, ±0.1℃), a MEMS triaxial accelerometer ADXL345 (±16g, 1000Hz), and a Rogowski coil current sensor RCT-100 (0~1000A, ±0.5%), respectively to collect cable sheath temperature, vibration, and current signals. A total of 5 sensor groups are deployed, located at the cable's beginning (0m), middle sections (300m, 600m, 800m), and end (1000m), respectively. Data is collected synchronously via an NI cDAQ-9178 module at a sampling frequency of 1000Hz, with 1024 sampling points per group and a 1-second sampling interval, achieving blind-zone-free synchronous monitoring of multiple physical quantities and key points. Temperature, vibration, and current signals of the cable under five different operating conditions were simultaneously collected by five sets of sensors. The analog signals were converted into digital signals by the data acquisition module and integrated to form a multi-channel raw signal dataset. The raw signal x(t) of each channel is a 1×1024 vector (t=1,2,...,1024). The raw signals of each channel were classified and labeled as follows: normal state is labeled as 0, open circuit fault is labeled as 1, partial discharge fault is labeled as 2, single-phase ground fault is labeled as 3, and phase-to-phase short circuit fault is labeled as 4. The dataset was divided into training set and test set in an 8:2 ratio and saved in CSV format for subsequent processing.

[0034] S12: For each channel sensor, sample the original signal of a certain channel once as x(t) ( ,in The total number of sampling points is used for analysis.

[0035] S2: A joint denoising algorithm combining Sparrow Optimization Algorithm (SSA) and Variational Mode Decomposition (VMD) is used to process the multi-channel original signal x(t) to obtain multiple intrinsic mode components (IMCs). The Pearson correlation coefficient between each IMC and the original signal is calculated. Effective IMCs are selected based on the correlation coefficient threshold, and adaptive wavelet threshold denoising is applied to each effective IMC to reconstruct the final denoised signal. The specific steps are as follows: S21: SSA and VMD parameter initialization. Set the sparrow population size N=50 and the maximum number of iterations. The proportion of discoverers is p=0.2, and the proportion of vigilants is q=0.1; the search interval for the VMD decomposition mode number K is set. The search interval for the penalty factor α Each individual sparrow corresponds to a set of VMD parameters. 50 sets of parameter combinations are randomly generated, all of which fall within the set search range; S22: VMD parameter combinations based on individual sparrows The original signal x(t) is decomposed using VMD to obtain K intrinsic mode components (IMFs), achieving preliminary separation of fault characteristic signals from noise. The original signal is represented as the sum of the intrinsic mode components:

[0036] In the formula Let be the k-th intrinsic eigenmode component, and the frequencies of each component do not overlap. Some IMF components are dominated by fault characteristic signals, while others are dominated by noise. VMD decomposition is achieved by solving the following variational problem:

[0037] In the formula, t is the time derivative, ω k For IMF k The center frequency of (t), α is the penalty factor, and each IMF component is obtained by iterative solution using the alternating direction multiplier method (ADMM).

[0038] S23: Calculate the Pearson correlation coefficient r between each IMFk(t) and the original signal x(t). k :

[0039] in, Let x(t) be the mean. for The mean of the correlation coefficient, where T is the signal length; set the correlation coefficient threshold. =0.5, keep Effective IMF components, excluding The dominant noise component is identified, redundant noise interference is reduced, and the efficiency and effectiveness of subsequent denoising are improved; S24: For each effective IMF component after screening, use the Symlet4 wavelet basis function to perform 4-level discrete wavelet decomposition to obtain the high-frequency wavelet coefficients, which are mainly noise components, and the low-frequency wavelet coefficients, which are mainly fault characteristic components.

[0040] S25: High-frequency coefficients are filtered using an adaptive soft threshold function.

[0041] in, These are the original wavelet coefficients. These are the processed wavelet coefficients; The adaptive threshold is calculated using the following formula:

[0042] In the formula, The standard deviation of noise. For signal length, This represents the maximum amplitude of the wavelet coefficients of the current IMF component.

[0043] S26: Perform inverse wavelet transform on the low-frequency wavelet coefficients and filtered high-frequency wavelet coefficients of each IMF component to obtain the denoised IMF components. ; S27: Denoise all IMF components The signal is then superimposed and reconstructed to obtain the initial denoised signal. :

[0044] In the formula The SNR reflects the current VMD parameter combination. The higher the SNR, the better the parameter combination for noise reduction effect; S28: Construct a fitness function with improving the signal SNR as the core objective, and combine it with root mean square error (RMSE) and noise suppression ratio (NRR) for auxiliary optimization to ensure that the optimized VMD parameters can maximize the improvement of the denoised signal quality.

[0045] in:

[0046]

[0047]

[0048] In the formula, The standard deviation of noise before denoising. The standard deviation of the noise after denoising. Weighting coefficients , This ensures that the optimal combination of VMD parameters corresponds to the maximized SNR when the fitness function value is maximized. S29: Update the positions of the discoverer, followers, and watchdogs. After 100 iterations, output the optimal VMD parameter combination corresponding to the maximum fitness value. In this embodiment, the optimal parameters are: , .

[0049] S291: Update the discoverer's location. (In the front of the population) Each individual entity acts as a discoverer, updating its location (VMD parameters) through a global search. :

[0050] In the formula, For the first i The discoverer was in the first t The position of the next iteration (corresponding to a set) ） Q is a random number that follows a normal distribution. ), i This is the individual serial number of the discoverer.

[0051] S292: Update follower positions; remaining individuals are followers. Perform local search and refine parameters by tracking the discoverer's position.

[0052] in, The optimal position in the t-th iteration corresponds to the VMD parameter combination with the highest current SNR. , The worst position in the t-th iteration corresponds to the VMD parameter combination with the lowest current SNR, and F is a random number following a uniform distribution. ), Random numbers that follow a normal distribution ( ), where i is the individual follower number and N is the population size.

[0053] S293: Update the vigilant's position. When the vigilant (randomly selected)... When an individual detects that the SNR at its current position is no longer increasing (i.e., it is trapped in a local optimum), it moves to a safe region (the optimal parameter region), as follows:

[0054] in, For the first i The vigilant was on the first t The current position in the next iteration. Let F be the globally optimal position in the t-th iteration, and F be a random number following a uniform distribution. ); S294: Determine if the current iteration count has reached the maximum iteration count. This completes the parameter optimization process and outputs the optimal parameters. If this is not achieved, return to step S22 and use the updated sparrow individual parameters. Re-execute S22-S29; if successful, output the optimal VMD parameter combination corresponding to the maximum fitness function value. .

[0055] S210: Based on the optimal parameters in S294 Perform VMD decomposition on the original signal x(t) to obtain One IMF component.

[0056] S211: Perform S23-S27 on each IMF component to obtain the denoised optimal IMF component. Based on this, all Superposition and reconstruction yield the final denoised signal. :

[0057] S3: Based on the final noise-reduced signal in step S211 Feature values ​​were extracted from three dimensions: time domain, frequency domain, and time-frequency domain, forming an initial multi-dimensional feature set, with a total of 29 feature values ​​extracted (12 in the time domain, 5 in the frequency domain, and 12 in the time-frequency domain). The specific steps are as follows: S31: Temporal feature extraction. Extracting the denoised signal. Peak value, mean, standard deviation, variance, peak-to-peak value, kurtosis, skewness, waveform factor, peak factor, kurtosis factor, impulse factor, and mean factor.

[0058] S32: Frequency domain feature extraction. First, the denoised signal... Perform a Fourier transform to obtain the power spectrum. :

[0059]

[0060] Based on this, calculate the centroid frequency, mean square frequency, root mean square frequency, frequency variance, and frequency standard deviation; S33: Time-frequency domain feature extraction: First, calculate the envelope entropy, sample entropy, approximate entropy and permutation entropy; then use wavelet packet transform to process the signal, with a wavelet basis of 10dB and a decomposition level of 3, to obtain the wavelet packet normalized energy corresponding to 8 frequency bands.

[0061] S4: The initial multi-dimensional feature set is filtered for core features using a combination of the ReliefF algorithm and the mRMR algorithm to obtain the final core feature set. Specifically, this includes the following steps: S41: For a certain channel, 250 signals under 5 working states are processed sequentially according to S2-S3 to construct an eigenvalue matrix F (1250 samples × 29).

[0062] S42: Calculate the discriminative power of each feature for different categories of samples using the ReliefF algorithm to obtain the feature weights:

[0063] in, For the first The weights of each feature; This represents the number of nearest neighbor samples; For the first The first sample Normalized values ​​of each feature; For the first The nearest neighbor samples of the same type of sample. Normalized values ​​of each feature; For the first The least neighbor samples of the different classes of the nth sample Normalized values ​​of each feature; For the first The range of values ​​for each feature ( ); For category The percentage of samples; For the first The category of each sample.

[0064] S43: Set weight threshold =0.3, filter out the weights The 16 features are used as the core feature set selected by the ReliefF algorithm. .

[0065] S44: mRMR screening of feature values ​​first involves mutual information calculation: feature is calculated using the following formula. Mutual information with fault category C With features and mutual information :

[0066] In the formula ： For joint probability density, 、 Marginal probability density; S45: Calculate the average mutual information between features and fault categories based on the maximum correlation criterion.

[0067] In the formula, S is the subset of features to be selected, and |S| is the number of features in the subset.

[0068] S46: According to the minimum redundancy criterion, the average mutual information R(S) among features within a feature subset is calculated using the following formula:

[0069] S47: Constructing the mRMR objective function:

[0070] S48: The incremental search method is used for iterative optimization 20 times, and the convergence threshold of the objective function is 1×10⁻⁶. -5 Fourteen features were selected to obtain the optimal feature set for mRMR. ; S49: Obtain the core feature set selected by the ReliefF algorithm. The core feature set selected by the mRMR algorithm Union of:

[0071] Final core feature set It contains 18 features, and the feature dimensions are compressed from 29 to 18, a compression ratio of about 38%, effectively eliminating redundant features.

[0072] S5: Normalize the core feature set. Specifically, this includes the following steps: S51: Based on The eigenvalue matrix constructed in S41 is reduced to obtain a 1250×18 eigenvalue matrix. S52: Employ the min-max normalization method for the core feature set under all operating conditions. Normalization is performed to eliminate the influence of different feature dimensions, mapping all feature values ​​to the [0,1] interval. The normalization formula is as follows:

[0073] in For the first The first sample The original values ​​of each feature, These are the normalized eigenvalues. For the first of all samples The maximum value of each feature, For the first of all samples The minimum value of each feature. ( (Total sample size: 1250) (h represents the total number of features, 18), ultimately yielding the normalized core feature set F'. core ; The samples were divided into a training set (1000 samples) and a test set (250 samples) in a ratio of S53:8:2.

[0074] S6: Using the normalized core feature set as input, the Grey Wolf optimization algorithm is used to globally optimize the hyperparameters of the capsule neural network to construct a fault diagnosis model. When cable fault diagnosis is required, the data of the cable to be diagnosed is input into the fault diagnosis model, and the fault diagnosis model outputs the cable operating status and the corresponding fault confidence probability. The specific steps are as follows: S61: Using the normalized core feature set F' from step S53 core The selected training set is used as input to construct a capsule network containing convolutional layers (3×3 kernel size, 32 kernels, stride 1), primary capsule layers (64 capsules, 16 dimensions), category capsule layers (5 capsules, corresponding to 5 working states, 32 dimensions), and decoder reconstruction layers (fully connected layers, output dimension 18, consistent with the core feature dimension). S62: The main capsule layer and the category capsule layers use a squash compression function with a temperature coefficient.

[0075] In the formula: : The output vector of the j-th capsule after squash compression; The total input vector of the j-th capsule is obtained by weighted summation of the lower-level capsules; Vector L2 norm, characterizing the activation intensity of capsules; : Squash compression temperature coefficient, a hyperparameter to be optimized.

[0076] S63: Employs an improved dynamic routing mechanism to achieve information transfer from lower-level capsules to higher-level capsules, and introduces a routing scaling factor. To suppress coupling coefficient oscillations, the formula for updating the routing logical value is:

[0077] In the formula: This is the routing logic value from lower-level capsule i to higher-level capsule j; is the scaling factor for the routing coupling coefficient, and is the hyperparameter to be optimized; is the prediction vector of the higher-level capsule j after the lower-level capsule i is mapped by the affine transformation matrix; Let j be the output vector of the high-level capsule. This is a vector dot product operation; S64: The coupling coefficients are normalized by softmax and then weighted and summed to obtain the total input of the category capsule. After iterative routing, the fault category confidence vector is output.

[0078] S65: The network employs joint optimization of classification margin loss and reconstruction loss, with the total loss being:

[0079] In the formula: Loss is the total loss function of the capsule network; For classification interval loss; The reconstructed loss weight coefficients are denoted as , and is the hyperparameter to be optimized. The mean squared error between the input feature vector and the feature vector reconstructed by the decoder; The core feature vector input to the model; This is the feature vector reconstructed from the category capsule by the decoder.

[0080] S66: The Grey Wolf Optimization (GWO) algorithm is used to globally optimize key parameters of the capsule network, including learning rate, weight decay coefficient, number of dynamic route iterations, route scaling factor, PrimaryCaps capsule dimension, DigitCaps capsule dimension, batch size, squash temperature coefficient, and reconstruction loss weight.

[0081] Set the initial range of parameters as follows: learning rate [0.001, 0.01], weight decay coefficient [0.0001, 0.001], dynamic routing iteration count [3, 7], routing scaling factor λ [0.5, 1.5], PrimaryCaps capsule dimensions [8, 32], DigitCaps capsule dimensions [16, 64], batch size [16, 64], squash temperature coefficient τ [0.5, 2.0], and reconstruction loss weights. [0.001, 0.01].

[0082] S67: Perform population initialization and parameter encoding: Set the gray wolf population size N to 30; maximum number of iterations. The value is 80. Real-number encoding is used for continuous parameters and integer encoding is used for discrete parameters. A 9-dimensional gray wolf position vector is generated, meaning the total dimension of the hyperparameters to be optimized is 9. This is mapped to a set of capsule network parameter combinations. S68: Construct a fitness function to verify set accuracy. The comprehensive fitness is constructed by combining scores and loss functions, and is adopted in fault diagnosis scenarios:

[0083] In the formula This represents the fitness value of an individual gray wolf. The fault classification accuracy of the model on the validation set; Fault classification of the model on the validation set The score balances precision and recall. This represents the total loss function value of the model on the validation set.

[0084] S69: Determine based on fitness ranking 、 、 Optimal gray wolf. Convergence factor a:

[0085] In the formula: a is the convergence factor; t is the current iteration number of the Grey Wolf algorithm; This represents the maximum number of iterations. S610: Update the Gray Wolf's position using the following formula:

[0086] in:

[0087] In the formula: 、 、 For the current individual gray wolf, respectively with 、 、 The distance vector of the wolf; 、 、 For the current iteration 、 、 The wolf's position vector (corresponding to the optimal three sets of parameter combinations); Let i be the position vector of the i-th gray wolf in the current iteration; Let i be the updated position vector of the i-th gray wolf; , This is the coefficient vector of the Grey Wolf algorithm; , The values ​​are random numbers uniformly distributed within the interval [0,1]. After updating, boundary constraints and discrete parameter rounding are applied to ensure that the parameters are within the valid search interval.

[0088] S611: After 80 iterations, the optimal hyperparameter combination corresponding to the maximum output fitness value is: learning rate 0.0035, weight decay coefficient 0.0005, dynamic routing iterations 5, routing scaling factor λ=0.9, PrimaryCaps capsule dimension 16, DigitCaps capsule dimension 32, batch size 32, squash temperature coefficient τ=1.2, and reconstruction loss weights. =0.005. Using this parameter, the number of iterations was increased to 200, and data augmentation (random translation and rotation) and L2 regularization were introduced to obtain the optimal GWO-CapsNet fault diagnosis model.

[0089] S612: The model diagnostic effect is verified using the test set. The magnitude of the output vector of the category capsule is used as the discrimination criterion, and the category with the largest magnitude is the diagnostic result. The output cable operating status is: normal state, open circuit, partial discharge, single-phase grounding, and phase-to-phase short circuit, and the fault confidence probability is given. The fault diagnosis results on the test set in this embodiment are shown in Table 1.

[0090] Table 1 Diagnostic Performance Tables for Various Operating Conditions

[0091] As shown in Table 1, the method of the present invention achieves an average diagnostic accuracy of 98.4% and an average fault confidence probability of 95.98% under five working conditions, demonstrating high diagnostic accuracy across all working conditions.

[0092] To verify the superiority of the method of the present invention, a comparative experiment was conducted between the GWO-CapsNet model of the present invention, a CNN model, and an unoptimized CapsNet model. The results are shown in Table 2.

[0093] Table 2 Performance Comparison of Different Diagnostic Models

[0094] As shown in Table 2, the GWO-CapsNet model outperforms the other two comparative models in terms of diagnostic accuracy, F1 score, and total loss. The diagnostic accuracy is improved by 5.6% compared to the CNN model and by 3.2% compared to the unoptimized CapsNet model, fully validating the effectiveness and superiority of the method proposed in this invention.

[0095] This embodiment focuses on 10kV cross-linked polyethylene cables. By simultaneously acquiring signals from multiple key locations and various types of sensors, and employing SSA-VMD combined noise reduction, multi-dimensional feature extraction, ReliefF-mRMR feature filtering, and GWO-CapsNet fault identification, it successfully achieved accurate diagnosis of both normal cable conditions and various faults, verifying the feasibility, effectiveness, and practicality of the method. The entire diagnostic process exhibits high accuracy, low false negative and false positive rates, and strong anti-interference capabilities and engineering feasibility, making it directly applicable to the daily operation and maintenance and fault diagnosis of urban power distribution network cables.

[0096] It should be noted that this embodiment is only a preferred embodiment of the present invention and is not intended to limit the present invention. In practical applications, the selection, placement, sampling parameters and algorithm parameters of the sensor can be adjusted according to the actual needs such as cable specifications, application scenarios, and fault types. For example, adjusting the sampling frequency of the sensor, the population size and iteration number of SSA and GWO, and the threshold of feature screening can all achieve the technical effects of the present invention and the adjustments fall within the protection scope of the present invention.

[0097] Example 2: This embodiment provides a cable fault intelligent diagnosis system based on multi-algorithm fusion.

[0098] The cable fault diagnosis system in this embodiment includes six functional modules: signal acquisition module, joint noise reduction module, feature extraction module, feature filtering module, normalization processing module, and fault diagnosis module.

[0099] During system operation, the signal acquisition module outputs multi-channel raw signals to the joint noise reduction module. The joint noise reduction module outputs the final noise-reduced signal to the feature extraction module. The feature extraction module outputs an initial multi-dimensional feature set to the feature filtering module. The feature filtering module outputs the core feature set to the normalization processing module. The normalization processing module outputs the normalized core feature set to the fault diagnosis module. The fault diagnosis module finally outputs the cable operating status and fault confidence probability. The specific functions and operating methods of each module are as follows.

[0100] I. Signal Acquisition Module The signal acquisition module runs on a data acquisition hardware platform and is used to synchronously acquire multi-channel raw signals at key locations of the cable to be diagnosed through a sensor group including temperature sensors, acceleration sensors and current sensors, to build a fault diagnosis dataset containing samples of various operating states.

[0101] In this embodiment, the signal acquisition module uses an NI cDAQ-9178 data acquisition card paired with an NI 9219 analog input module, connected to the edge computing node via a USB interface. Five sensor groups are deployed, located at the cable's beginning (0m), the middle fault-prone sections (300m, 600m, and 800m), and the end (1000m). Each sensor group integrates three sensors: a PT100 platinum resistance temperature sensor (-40℃~120℃, ±0.1℃), a MEMS triaxial accelerometer ADXL345 (±16g, 1000Hz), and a Rogowski coil current sensor RCT-100 (0~1000A, ±0.5%), respectively acquiring cable sheath temperature, vibration, and current signals. The sampling frequency is set to 1000Hz, with 1024 sampling points per sample and a 1-second sampling interval.

[0102] The signal acquisition module synchronously collects temperature, vibration, and current signals of the cable under five operating conditions (normal state, open circuit fault, partial discharge fault, single-phase grounding fault, and phase-to-phase short circuit fault) using five sets of sensors. The analog signals are converted to digital signals by a data acquisition card and integrated to form a multi-channel raw signal dataset. The raw signal x(t) for each channel is a 1×1024 vector (t=1,2,...,1024). The raw signals for each channel are categorized and labeled: normal state is labeled 0, open circuit fault is labeled 1, partial discharge fault is labeled 2, single-phase grounding fault is labeled 3, and phase-to-phase short circuit fault is labeled 4. 250 samples are collected for each operating condition, totaling 1250 samples, which are divided into a training set (1000 samples) and a test set (250 samples) at an 8:2 ratio.

[0103] II. Combined Noise Reduction Module The joint noise reduction module receives the original signal x(t) output by the signal acquisition module as input, and uses a joint noise reduction algorithm combining the sparrow optimization algorithm and variational mode decomposition to process the multi-channel original signal. The effective intrinsic mode components are screened by Pearson correlation coefficient and combined with adaptive wavelet threshold denoising to reconstruct the final noise-reduced signal.

[0104] The joint denoising module first uses a sparrow search algorithm to globally optimize the number of decomposition modes K and the penalty factor α of VMD, with the comprehensive fitness function (SNR, RMSE, NRR weighted) as the optimization objective. Then, based on the optimal parameters, it performs VMD decomposition on the original signal, filters effective IMF components using the Pearson correlation coefficient, and removes noise-dominant components. Next, it performs Symlet4 wavelet 4-level decomposition and adaptive soft thresholding denoising on each effective component. Finally, it superimposes and reconstructs the denoised IMF components to output the final denoised signal. The specific processing flow of the joint denoising module is exactly the same as step S2 in Example 1, and will not be repeated here. The denoised signal is output to the feature extraction module.

[0105] III. Feature Extraction Module The feature extraction module receives the final denoised signal output by the joint denoising module. As input, feature values ​​are extracted from the signal in three dimensions: time domain, frequency domain, and time-frequency domain, forming an initial multi-dimensional feature set.

[0106] In this embodiment, the feature extraction module performs the following three types of feature extraction: Temporal feature extraction: Extract 12 temporal feature values ​​of the denoised signal, including peak value, mean, standard deviation, variance, peak-to-peak value, kurtosis, skewness, waveform factor, peak factor, kurtosis factor, impulse factor, and mean factor.

[0107] Frequency domain feature extraction: First, perform Fourier transform on the denoised signal to obtain the power spectrum P(ω). Then, calculate five frequency domain feature values ​​based on the power spectrum: centroid frequency, mean square frequency, root mean square frequency, frequency variance, and frequency standard deviation.

[0108] Time-frequency domain feature extraction: Calculate four entropy features—envelope entropy, sample entropy, approximate entropy, and permutation entropy; then use wavelet packet transform to process the signal, with db10 as the wavelet basis and three decomposition layers to obtain the normalized wavelet packet energy corresponding to eight frequency bands.

[0109] A total of 29 feature values ​​were extracted from each sample to form an initial multi-dimensional feature set, which was then output to the feature selection module.

[0110] IV. Feature Filtering Module The feature filtering module receives the 1250×29 feature value matrix output by the feature extraction module as input, and uses a combination of the ReliefF algorithm and the mRMR algorithm to filter the core features of the initial multi-dimensional feature set to obtain the final core feature set.

[0111] The feature selection module runs the ReliefF algorithm and the mRMR algorithm respectively: the ReliefF algorithm selects high-discrimination features based on feature weight thresholds, and the mRMR algorithm selects low-redundancy features based on the maximum relevance and minimum redundancy criterion; the union of the two selection results is taken as the core feature set, and a 1250×18 core feature value matrix is ​​output to the normalization processing module. The specific processing flow of the feature selection module is the same as step S4 in Example 1, and will not be repeated here.

[0112] V. Normalization Processing Module The normalization processing module receives the 1250×18 core feature value matrix output by the feature selection module as input, and uses the min-max normalization method to normalize the core feature values ​​under all working conditions. After normalization, all feature values ​​are mapped to the [0,1] interval, and are divided into a training set (1000 samples) and a test set (250 samples) in an 8:2 ratio, and then output to the fault diagnosis module.

[0113] VI. Fault Diagnosis Module The fault diagnosis module receives the normalized core feature set output by the normalization processing module as input, and uses the Grey Wolf optimization algorithm to globally optimize the hyperparameters of the capsule neural network to construct the GWO-CapsNet fault diagnosis model. The normalized core feature set of the cable to be diagnosed is input into the GWO-CapsNet fault diagnosis model for forward inference. The magnitude of the output vector of the category capsule is used as the discrimination criterion, and the category with the largest magnitude is the diagnosis result. The module outputs the cable operating status and the corresponding fault confidence probability.

[0114] The fault diagnosis module first constructs a capsule neural network containing convolutional layers, a main capsule layer, category capsule layers, and a decoder, employing an improved squash function with a temperature coefficient τ and an improved dynamic routing mechanism with a scaling factor λ. Then, the Grey Wolf optimization algorithm is used to globally optimize nine key hyperparameters, with the comprehensive fitness function (accuracy, F1 score, and loss value weighted) as the optimization objective. After optimization, the optimal parameter combination is used for extended-round fine training, incorporating data augmentation and L2 regularization to obtain the final GWO-CapsNet model. During diagnosis, the magnitude of the category capsule output vector is used as the discrimination criterion to output the cable fault type and its corresponding confidence probability. The specific processing flow of the fault diagnosis module (including capsule network construction, GWO hyperparameter optimization, model training, and inference) is the same as step S6 in Example 1, and will not be repeated here.

[0115] The above are specific embodiments of the present invention, which demonstrate the substantial features and progress of the present invention. Equivalent modifications can be made to them according to actual usage needs, under the guidance of the present invention, and all such modifications are within the scope of protection of this solution.

Claims

1. A cable fault intelligent diagnosis method based on multi-algorithm fusion, characterized in that, Includes the following steps: S1: Install a sensor group including temperature sensor, acceleration sensor and current sensor at key locations of the cable to be diagnosed, and simultaneously collect multi-channel raw signals to build a fault diagnosis dataset including samples of various working states. S2: The multi-channel original signal is processed by a joint denoising algorithm combining the Sparrow Optimization algorithm and Variational Mode Decomposition to obtain multiple intrinsic mode components; the Pearson correlation coefficient between each intrinsic mode component and the original signal is calculated, the effective intrinsic mode components are selected according to the correlation coefficient threshold, and adaptive wavelet threshold denoising is performed on each effective intrinsic mode component to reconstruct the final denoised signal; S3: Based on the final denoised signal, extract feature values ​​from three dimensions: time domain, frequency domain, and time-frequency domain to form an initial multi-dimensional feature set; S4: The initial multi-dimensional feature set is filtered for core features by combining the ReliefF algorithm and the mRMR algorithm to obtain the final core feature set; S5: Normalize the core feature set; S6: Using the normalized core feature set as input, the Grey Wolf Optimization Algorithm is used to globally optimize the hyperparameters of the capsule neural network and construct a fault diagnosis model. When cable fault diagnosis is required, the data of the cable to be diagnosed is input into the fault diagnosis model, and the fault diagnosis model outputs the cable operating status and the corresponding fault confidence probability.

2. The intelligent cable fault diagnosis method based on multi-algorithm fusion according to claim 1, characterized in that, In step S1, the key locations include the cable head end, the middle fault-prone section, and the end; the multiple operating states include normal state, open circuit, partial discharge, single-phase grounding, and phase-to-phase short circuit; temperature signals, vibration signals, and current signals during cable operation are collected synchronously by each group of sensors to form a multi-channel raw signal dataset, a cable fault diagnosis dataset is created, and the samples under each operating state are divided into training set and test set according to the proportion.

3. The intelligent cable fault diagnosis method based on multi-algorithm fusion according to claim 1, characterized in that, Step S2 specifically includes: S21: Set the sparrow population size, maximum number of iterations, proportion of discoverers, and proportion of vigilants; set the search range for the number of decomposition modes and the search range for the penalty factor in variational mode decomposition; each individual sparrow corresponds to a set of variational mode decomposition parameter combinations, including the number of decomposition modes and the penalty factor; S22: Based on the variational mode decomposition parameter combination corresponding to each individual sparrow, the original signal is subjected to variational mode decomposition to obtain multiple intrinsic eigenmode components. The original signal is equal to the sum of the intrinsic eigenmode components. S23: Calculate the Pearson correlation coefficient between each intrinsic mode component and the original signal, set a correlation coefficient threshold, retain effective intrinsic mode components with a correlation coefficient greater than or equal to the correlation coefficient threshold, and remove noise-dominant components with a correlation coefficient less than the correlation coefficient threshold. S24: For each effective intrinsic mode component after screening, use the Symlet4 wavelet basis function to perform 4-level discrete wavelet decomposition to obtain high-frequency wavelet coefficients and low-frequency wavelet coefficients. S25: The high-frequency wavelet coefficients are filtered using an adaptive soft threshold function. The processed wavelet coefficients are equal to the positive value obtained by multiplying the sign of the original wavelet coefficients by the absolute value of the original wavelet coefficients minus the adaptive threshold. The adaptive threshold is determined by the noise standard deviation, the signal length, and the maximum amplitude of the wavelet coefficients of the current intrinsic mode components. Its value decreases adaptively as the amplitude of the wavelet coefficients increases. S26: Perform inverse wavelet transform on the low-frequency wavelet coefficients and the filtered high-frequency wavelet coefficients of each intrinsic eigenmode component to obtain the denoised intrinsic eigenmode components. S27: Construct the fitness function, which is equal to the signal-to-noise ratio multiplied by the first weight, minus the root mean square error multiplied by the second weight, plus the noise suppression ratio multiplied by the third weight; with the goal of maximizing the fitness function value, perform the position updates of the discoverer, followers, and guards, and determine whether the maximum number of iterations has been reached. If not, return to step S22 to continue iterating; after reaching the maximum number of iterations, output the optimal combination of variational mode decomposition parameters corresponding to the maximum fitness function value, including the optimal number of decomposition modes and the optimal penalty factor; S28: Based on the optimal number of decomposed modes and the optimal penalty factor, perform variational mode decomposition on the original signal to obtain the optimal number of intrinsic mode components; after filtering each optimal intrinsic mode component by Pearson correlation coefficient and adaptive wavelet threshold denoising, superimpose and reconstruct all the denoised optimal intrinsic mode components to obtain the final denoised signal.

4. The intelligent cable fault diagnosis method based on multi-algorithm fusion according to claim 1, characterized in that, In step S4: The ReliefF algorithm is used to calculate the distinguishing ability of each feature for samples of different categories, obtain the feature weight, set the weight threshold, and select features with weights greater than or equal to the weight threshold to form the ReliefF preferred feature set. The mutual information between features and fault categories, as well as the mutual information between features, is calculated using the mRMR algorithm, and an objective function is constructed. The objective function is based on the maximum correlation criterion and the minimum redundancy criterion, where the maximum correlation criterion value is the average value of the mutual information between each feature and fault category within the feature subset, and the minimum redundancy criterion value is the average value of the mutual information between each feature within the feature subset. The objective function is equal to the maximum correlation criterion value minus the minimum redundancy criterion value. An incremental search method is used for iterative optimization to obtain the optimal feature set for mRMR; The union of the ReliefF preferred feature set and the mRMR preferred feature set is taken as the final core feature set.

5. The intelligent cable fault diagnosis method based on multi-algorithm fusion according to claim 1, characterized in that, In step S6, the capsule neural network includes a convolutional layer, a main capsule layer, a category capsule layer, and a decoder reconstruction layer; Both the main capsule layer and the category capsule layer use a squash compression function with a temperature coefficient to perform nonlinear compression on the capsule output vector. Specifically, the compressed output vector of the j-th capsule is equal to the square of the L2 norm of the total input vector of the capsule divided by the sum of the square of the L2 norm and the square of the temperature coefficient, and then multiplied by the total input vector of the capsule divided by its L2 norm. The capsule neural network adopts an improved dynamic routing mechanism to realize information transmission from lower-level capsules to higher-level capsules. The routing logic value update method is as follows: the routing logic value update from lower-level capsule i to higher-level capsule j is its original value plus the routing scaling factor multiplied by the dot product of the prediction vector of lower-level capsule i to higher-level capsule j after affine transformation and the output vector of higher-level capsule j.

6. The intelligent cable fault diagnosis method based on multi-algorithm fusion according to claim 5, characterized in that, In step S6, the specific steps for globally optimizing the hyperparameters of the capsule neural network using the Grey Wolf optimization algorithm include: The gray wolf population size and maximum number of iterations are set. The learning rate, weight decay coefficient, dynamic routing iterations, routing scaling factor, capsule dimension of the main capsule layer, capsule dimension of the category capsule layer, batch size, temperature coefficient of the squash compression function, and reconstruction loss weight of the capsule neural network are encoded into multi-dimensional position vectors of individual gray wolves. Continuous parameters are encoded with real numbers, and discrete parameters are encoded with integers. Construct a fitness function, which is equal to the validation set fault classification accuracy multiplied by the first weight, plus the validation set F1 score multiplied by the second weight, plus the reciprocal of the validation set total loss function value plus 1 multiplied by the third weight. The α wolf with the best fitness, the β wolf with the second best fitness, and the δ wolf with the third best fitness are determined based on the fitness ranking. The α wolf, β wolf and δ wolf jointly guide the remaining gray wolf individuals to update their positions. After iterating to the maximum number of iterations, the optimal hyperparameter combination corresponding to the maximum fitness value is output. The optimal hyperparameter combination is used to perform extended training of the capsule neural network, and random translation, rotation data augmentation and L2 regularization are introduced to obtain the optimal GWO-CapsNet fault diagnosis model.

7. The intelligent cable fault diagnosis method based on multi-algorithm fusion according to claim 1, characterized in that, In step S3, the time-domain feature values ​​include peak value, mean, standard deviation, variance, peak-to-peak value, kurtosis, skewness, waveform factor, peak factor, kurtosis factor, impulse factor, and mean factor; the frequency-domain feature values ​​include centroid frequency, mean square frequency, root mean square frequency, frequency variance, and frequency standard deviation; the time-frequency domain feature values ​​include envelope entropy, sample entropy, approximate entropy, permutation entropy, and normalized wavelet packet energy of each frequency band after wavelet packet transform.

8. The intelligent cable fault diagnosis method based on multi-algorithm fusion according to claim 1, characterized in that, In step S5, the normalization process is as follows: the normalized feature value is equal to the original value of the feature in the sample minus the minimum value of the feature in all samples, and then divided by the maximum value of the feature in all samples minus the minimum value of the feature.

9. A cable fault intelligent diagnosis system based on multi-algorithm fusion, characterized in that, include: The signal acquisition module is used to simultaneously acquire multi-channel raw signals at key locations of the cable to be diagnosed using a sensor group that includes temperature sensors, acceleration sensors, and current sensors, and to build a fault diagnosis dataset containing samples of various operating states. The joint noise reduction module is used to process the multi-channel original signal using a joint noise reduction algorithm that combines the sparrow optimization algorithm and variational mode decomposition. It filters the effective intrinsic mode components by Pearson correlation coefficient and combines them with adaptive wavelet thresholding to reconstruct the final noise-reduced signal. The feature extraction module is used to extract feature values ​​from the time domain, frequency domain, and time-frequency domain based on the final denoised signal to form an initial multi-dimensional feature set; The feature filtering module is used to filter the core features of the initial multi-dimensional feature set by combining the ReliefF algorithm and the mRMR algorithm to obtain the final core feature set. The normalization processing module is used to normalize the core feature set; The fault diagnosis module is used to globally optimize the hyperparameters of the capsule neural network using the gray wolf optimization algorithm to build a fault diagnosis model. When cable fault diagnosis is required, the data of the cable to be diagnosed is input into the fault diagnosis model, and the fault diagnosis model outputs the cable operating status and the corresponding fault confidence probability.

10. The intelligent cable fault diagnosis system based on multi-algorithm fusion according to claim 9, characterized in that: The various operating states include normal state, open circuit, partial discharge, single-phase grounding, and phase-to-phase short circuit; the multi-dimensional feature values ​​include peak value, mean, standard deviation, variance, peak-to-peak value, kurtosis, skewness, waveform factor, peak factor, kurtosis factor, impulse factor, and mean factor in the time domain; centroid frequency, mean square frequency, root mean square frequency, frequency variance, and frequency standard deviation in the frequency domain; and envelope entropy, sample entropy, approximate entropy, permutation entropy, and normalized wavelet packet energy of each frequency band after wavelet packet transform in the time and frequency domain.

Images