Cable early fault intelligent identification method, device and equipment based on dynamic threshold value and storage medium

Abstract

The invention relates to the technical field of power grids, in particular to a dynamic threshold-based cable early fault intelligent identification method, device and equipment and a storage medium. Cable current disturbance is judged and early fault original data are identified by acquiring cable current sampling signals and adopting cable overcurrent transient process detection driven by a dynamic threshold value; performing S transformation on the original data to construct an initial feature set; selecting an optimal feature subset by adopting a Pearson's correlation coefficient method; and inputting the feature subset into a CNN-GRU-Attention neural network model to realize intelligent fault identification and classification. According to the method, the temperature compensation coefficient and the load rate deviation coefficient are introduced to dynamically adjust the threshold value, and the defect that a traditional fixed threshold value is poor in adaptability to environment and working condition changes is overcome; through combination of deep learning and time-frequency analysis, multiple types of cable early faults such as partial discharge, arc faults, insulation degradation and mixed faults can be accurately identified, and the accuracy and efficiency of fault identification are improved.

Description

Technical Field

[0001] The present invention relates to the field of power grid technology, and in particular to a method, device, equipment and storage medium for intelligently identifying early-stage cable faults based on a dynamic threshold. Background Art

[0002] Distribution cables are key components of power systems, transporting electricity from substations to end users, ensuring consistent power distribution across diverse regions and demand. In actual operation, cables can experience unpredictable incipient failures, often caused by localized insulation discharge (LD) due to aging, before they permanently fail.

[0003] Currently, cable fault identification technologies fall into two main categories: threshold methods and deep learning algorithms. The threshold method primarily processes fault signals and sets a threshold value to determine the fault type. In practice, this method relies on the operator's experience and expertise to set the threshold, which is highly subjective. Due to the fixed nature of the threshold setting, this method is less adaptable to changes in system topology and operating conditions, and cannot quickly respond to dynamic changes during system operation.

[0004] Deep learning algorithms, by employing neural network structures to automatically learn and extract key features from fault signals, enable end-to-end fault identification without the need for manual threshold setting. These algorithms offer greater adaptability and generalization, enabling them to handle diverse and complex cable fault patterns, thereby improving fault identification accuracy. However, existing deep learning methods still have room for improvement in feature extraction and selection, and most fail to fully account for the impact of cable operating environments and load variations. Summary of the Invention

[0005] In view of the problems existing in the prior art, the present invention is proposed.

[0006] Therefore, the problem to be solved by the present invention is how to solve the drawbacks of the traditional threshold method that the threshold needs to be set manually, as well as the problems of excessive output data, large dimensions, and insufficient fault identification accuracy. By introducing a dynamic threshold setting mechanism, the present invention can automatically adjust the threshold parameters according to environmental conditions and system operating status, thereby improving the environmental adaptability and sensitivity of the threshold setting. At the same time, feature selection is performed through the S transform and Pearson correlation coefficient method, which effectively reduces the dimension of the input data and improves the fitting speed. Combined with the CNN-GRU-Attention deep learning model, the present invention significantly improves the accuracy and efficiency of early cable fault identification, providing a strong guarantee for the safe and stable operation of the power system.

[0007] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0008] In a first aspect, an embodiment of the present invention provides a method for intelligently identifying early-stage cable faults based on a dynamic threshold, which includes obtaining a cable current sampling signal;

[0009] Cable overcurrent transient process detection based on dynamic threshold drive, judging cable current disturbance, and identifying raw data of early cable faults;

[0010] Performing S transformation on the original data to construct an initial feature set;

[0011] Perform feature selection on the initial feature set by using the Pearson correlation coefficient method to select the optimal feature subset;

[0012] The optimal feature subset is input into the CNN-GRU-Attention neural network model to achieve intelligent recognition and classification of early cable faults.

[0013] As a preferred solution of the method for intelligently identifying early-stage cable faults based on dynamic thresholds according to the present invention, the cable overcurrent transient process detection based on dynamic threshold driving includes:

[0014] A composite criterion is formed by the high-frequency detail coefficient energy criterion and the low-frequency approximation coefficient root mean square value criterion;

[0015] When any one of the high-frequency detail coefficient energy criterion or the low-frequency approximation coefficient root mean square value criterion is met, it is determined that a sudden disturbance occurs in the cable current;

[0016] Determine whether a second transient current change occurs within 0.5 to 5 cycles after the first disturbance mutation occurs;

[0017] When the second disturbance mutation occurs, the current data between the two disturbance mutations are used as the original data for early cable fault identification.

[0018] As a preferred solution of the dynamic threshold-based intelligent identification method for early cable faults described in the present invention, the threshold of the high-frequency detail coefficient energy criterion includes a temperature compensation coefficient, and the temperature compensation coefficient is dynamically adjusted according to the real-time monitoring temperature of the cable surface and the reference benchmark temperature.

[0019] As an optimal solution of the intelligent identification method for early cable faults based on dynamic thresholds described in the present invention, the threshold of the low-frequency approximation coefficient root mean square value criterion includes a temperature deviation coefficient and a load rate deviation coefficient, and is dynamically adjusted according to the ambient temperature, cable thermal resistance coefficient, cable rated operating temperature and load rate.

[0020] As a preferred solution of the intelligent identification method for early cable faults based on dynamic thresholds described in the present invention, the initial feature set constructed by the S transform includes features in three aspects: statistics, entropy and energy, specifically including six feature quantities: mean, standard deviation, skewness, kurtosis, root mean square value and logarithmic energy entropy.

[0021] As a preferred solution of the method for intelligently identifying early-stage cable faults based on dynamic thresholds according to the present invention, the CNN-GRU-Attention neural network model includes:

[0022] The convolutional neural network (CNN) part is used to extract the spatial information of the input features;

[0023] The gated recurrent unit (GRU) part is used to process time series data;

[0024] The attention mechanism is used to weight the feature sequence output by GRU to improve the discrimination of feature importance;

[0025] The fully connected layer is used to output the recognition and classification results of early cable faults.

[0026] As a preferred solution of the method for intelligently identifying early-stage cable faults based on dynamic thresholds according to the present invention, the following is mentioned:

[0027] In a second aspect, an embodiment of the present invention provides a dynamic threshold-based intelligent cable early fault identification system, which includes a signal acquisition module for acquiring a cable current sampling signal;

[0028] The disturbance detection module is used to detect the transient process of cable overcurrent based on dynamic threshold drive, judge the cable current disturbance, and identify the raw data of early cable faults;

[0029] A feature extraction module is used to perform S transformation on the original data to construct an initial feature set;

[0030] A feature selection module is used to perform feature selection on the initial feature set by using the Pearson correlation coefficient method to select the optimal feature subset;

[0031] The fault identification module is used to input the optimal feature subset into the CNN-GRU-Attention neural network model to realize intelligent identification and classification of early cable faults.

[0032] In a third aspect, an embodiment of the present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program instructions are executed by the processor, the steps of the method for intelligently identifying early cable faults based on dynamic thresholds as described in the first aspect of the present invention are implemented.

[0033] In a fourth aspect, an embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program instructions are executed by a processor, the steps of the method for intelligently identifying early cable faults based on dynamic thresholds as described in the first aspect of the present invention are implemented.

[0034] The beneficial effects of the present invention are as follows: by introducing a cable overcurrent transient process detection mechanism based on dynamic threshold drive, and dynamically adjusting the judgment threshold in combination with the temperature compensation coefficient and the load rate deviation coefficient, the present invention realizes the adaptive ability of cable fault detection to the environment and operating status. Compared with the traditional fixed threshold method, this technology overcomes the disadvantage that the threshold needs to be manually set, and can automatically optimize the detection parameters according to factors such as the real-time monitoring temperature of the cable surface, the ambient temperature, and the load rate. When the cable operating environment changes, the system can adjust the thresholds of the high-frequency detail coefficient energy judgment and the low-frequency approximation coefficient root mean square value judgment in real time to maintain the best detection sensitivity. This adaptive mechanism significantly improves the environmental adaptability and sensitivity of the threshold setting, reduces the false alarm rate and missed detection rate, makes the early fault detection of the cable more accurate and reliable, and provides a strong guarantee for the safe and stable operation of the power system. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0036] Figure 1 The flowchart of the intelligent identification method of early cable faults based on dynamic threshold;

[0037] Figure 2 A computer device diagram of an intelligent identification method for early cable faults based on dynamic thresholds;

[0038] Figure 3 This is the S-modulus time-frequency matrix diagram of the intelligent identification method for early cable faults based on dynamic thresholds;

[0039] Figure 4 This is a flow chart of the dynamic threshold cable overcurrent transient process detection method of the intelligent identification method of cable early fault based on dynamic threshold;

[0040] Figure 5 This is the GRU loop structure diagram of the intelligent identification method for early cable faults based on dynamic thresholds;

[0041] Figure 6 Flowchart for calculating the attention mechanism weights for the intelligent cable early fault identification method based on dynamic thresholds

[0042] Figure 7 Schematic diagram of the CNN-GRU model for intelligent identification of early cable faults based on dynamic thresholds

[0043] Figure 8 The flowchart of the cable early fault identification method based on dynamic threshold is shown. DETAILED DESCRIPTION

[0044] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0045] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0046] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it constitute a separate or selective embodiment that is mutually exclusive with other embodiments.

[0047] Example 1

[0048] Reference Figures 1 and 2 , which is the first embodiment of the present invention, provides a method for intelligently identifying early-stage cable faults based on a dynamic threshold, comprising:

[0049] S100: Acquire cable current sampling signal;

[0050] S200: Acquires raw data for detecting transient overcurrent conditions in cables based on dynamic threshold driving, determines cable current disturbances, and identifies early-stage cable faults.

[0051] S300: Obtaining and performing S transformation on the original data to construct an initial feature set;

[0052] S400: performing feature selection on the initial feature set by using the Pearson correlation coefficient method to select the optimal feature subset;

[0053] S500: The optimal feature subset is input into a CNN-GRU-Attention neural network model to achieve intelligent recognition and classification of early cable faults.

[0054] During long-term operation, distribution cables in power systems are prone to early-stage failures due to factors such as insulation aging, environmental impacts, and load fluctuations. If these early-stage failures are not discovered and handled in a timely manner, they will evolve into more serious permanent failures, leading to system power outages, equipment damage, and even safety accidents. In step S100, the acquisition of cable current sampling signals in existing technologies often lacks adaptability and cannot accurately capture fault characteristics under different operating environments. In step S200, traditional fault detection methods mainly rely on fixed thresholds and cannot cope with the dynamic changes of cables under different temperature and load conditions, resulting in high false alarm rates or serious missed detections. In step S300, common signal analysis methods such as Fourier transform can only obtain frequency domain information and cannot simultaneously reflect the time domain characteristics of the fault signal, making it difficult to accurately identify transient faults. In step S400, there is a lack of effective feature selection methods, which often retains too many redundant features or omits key features, affecting the accuracy of subsequent identification. In step S500, traditional models cannot simultaneously process the spatial and temporal characteristics of cable fault signals, and do not adequately distinguish the importance of different features. This technical solution solves the above problems and achieves high-precision identification of early cable faults through the organic combination of dynamic threshold driving, S-transform feature extraction, Pearson correlation coefficient feature selection and CNN-GRU-Attention deep learning model.

[0055] Claim 1 of the present invention forms a complete intelligent identification solution for early cable faults through the coordinated work of five key steps. First, by acquiring the cable current sampling signal, a data foundation for fault identification is established; second, the disturbance detection method based on dynamic threshold overcomes the shortcomings of the traditional threshold method in poor environmental adaptability, and can automatically adjust the threshold parameters according to factors such as temperature and load, accurately capturing the disturbance changes of the cable current; third, the S transform is used to perform time-frequency analysis on the original data, and the time domain and frequency domain characteristics of the signal are obtained at the same time, and a feature set containing multi-dimensional information is constructed; fourth, feature selection is performed through the Pearson correlation coefficient method, redundant features are eliminated, the data dimension is reduced, and computational efficiency is improved; finally, the optimized features are input into the CNN-GRU-Attention neural network model, giving full play to the advantages of convolutional neural networks in extracting spatial features, gated recurrent units in processing time series data, and attention mechanisms in enhancing the weight of key information, thereby achieving accurate identification and classification of various early fault types such as partial discharge, arc fault, and insulation degradation. The entire technical solution can complete the entire process from signal acquisition to fault classification without human intervention, greatly improving the automation level, accuracy and real-time performance of cable fault identification, and can independently solve the technical problem of difficult early fault detection in distribution cables.

[0056] In this embodiment, the S transform is a time-frequency analysis technology. Compared with the traditional Fourier transform, the S transform can simultaneously provide time domain and frequency domain information of the signal, retain the phase information of the signal, and has a frequency-adaptive window width characteristic, which is suitable for analyzing non-stationary signals such as cable faults.

[0057] CNN (Convolutional Neural Network) is a deep learning model specifically designed to process data with grid structures. Through features such as local connections, weight sharing, and multi-layer stacking, it can automatically extract the spatial features of signals and is suitable for processing local features and spatial relationships in cable fault signals.

[0058] GRU (Gated Recurrent Unit) is a variant of recurrent neural network that can effectively capture long-term dependencies of data. It controls the flow of information by updating and resetting gates, solving the gradient vanishing problem of traditional RNN and is suitable for processing the timing characteristics of cable fault signals.

[0059] Attention is a mechanism that can highlight important information and suppress irrelevant information. By assigning different weights to different time steps or features, the model pays more attention to the signal parts that are critical to fault identification, thereby improving the interpretability and accuracy of the model.

[0060] Example 2

[0061] Reference Figures 1-8 , which is the second embodiment of the present invention.

[0062] In the embodiment of the present application, obtaining the cable current sampling signal in step S100 includes the following steps A1-A2:

[0063] A1: Use distribution line current sensors to collect cable three-phase current sampling signals with a sampling frequency of 10 kHz and a sampling time of no less than 10 power cycles.

[0064] In a preferred embodiment, the current sensor used is a Hall effect current sensor with a measurement range of 0-1000A, an accuracy of ±0.5%, and a bandwidth of 0-100kHz. The sampling frequency can be adjusted according to actual needs. When capturing high-frequency transient signals, the sampling frequency can be increased to 50kHz to ensure that the high-frequency details of the fault are captured. Hall effect current sensors also offer advantages such as good isolation, high linearity, and strong anti-interference capabilities, making them suitable for use in the complex electromagnetic environment of power systems.

[0065] In another optional embodiment, a Rogowski coil current sensor can be used. This sensor has the advantage of requiring no power outage for installation and is suitable for monitoring operating cables. It offers a measurement range of 0-500A, an accuracy of ±1.0%, and a bandwidth of 0-30kHz. For medium- or high-voltage cables, a fiber-optic current sensor can be used. Its excellent insulation and resistance to electromagnetic interference make it particularly suitable for operation in high-voltage environments.

[0066] It should be noted that the cyclical characteristics of the power system must be considered during current sampling. When the power frequency is 50Hz, one cycle is 20ms. The sampling duration should include at least 10 power cycles, or 200ms. This ensures that sufficient data is collected to analyze the cyclical changes in cable current and possible fault characteristics. In actual applications, sampling parameters can be flexibly adjusted according to specific monitoring needs and system conditions to achieve optimal monitoring results.

[0067] A2: Preprocess the collected raw current signal, including filtering, denoising and standardization.

[0068] In a preferred embodiment, the original signal is first filtered using a Butterworth low-pass filter with a cutoff frequency set to 0.4 times the sampling frequency to eliminate high-frequency noise and sampling aliasing; then, a wavelet threshold denoising method is used for denoising, the db4 wavelet is selected, the number of decomposition layers is 4, and a soft threshold function is used, which can effectively remove Gaussian white noise in the signal while retaining the important features of the signal; finally, the signal is normalized, and the processed signal amplitude is normalized to the range of [-1,1] to eliminate the influence of signal amplitude differences at different times and under different load conditions.

[0069] In another optional embodiment, a Kalman filter algorithm can be used to preprocess the signal. This method can effectively suppress noise interference while preserving the signal's dynamic characteristics. For signals containing impulse noise, a combination of median filtering and mean filtering can be used: median filtering is used to remove impulse interference, and mean filtering is used to smooth the signal. In addition, a high-pass filter can be used to correct for signal baseline drift that may occur during long-term monitoring, ensuring signal consistency and comparability.

[0070] It should be noted that the purpose of signal preprocessing is to improve the accuracy of subsequent analysis. Different preprocessing methods have different effects on the signal, and the appropriate preprocessing method should be selected based on the actual cable operating environment and monitoring requirements. In environments with strong electromagnetic interference, the filtering strength can be increased; in situations where the signal changes rapidly, care should be taken to preserve the dynamic characteristics of the signal. The selection of preprocessing parameters directly affects the subsequent extraction of fault characteristics and requires careful setting and verification.

[0071] In the embodiment of the present application, step S200 obtains the raw data of the cable overcurrent transient process detection based on dynamic threshold driving, determines the cable current disturbance, and identifies the early cable fault, including the following steps B1-B3:

[0072] B1: Cable overcurrent transient process detection based on dynamic threshold drive includes:

[0073] A composite criterion is formed by the high-frequency detail coefficient energy criterion and the low-frequency approximation coefficient root mean square value criterion;

[0074] When any one of the high-frequency detail coefficient energy criterion or the low-frequency approximation coefficient root mean square value criterion is met, it is determined that a sudden disturbance occurs in the cable current;

[0075] Determine whether a second transient current change occurs within 0.5 to 5 cycles after the first disturbance mutation occurs;

[0076] When the second disturbance mutation occurs, the current data between the two disturbance mutations are used as the original data for early cable fault identification.

[0077] In a preferred embodiment, the preprocessed cable current signal is first decomposed using a six-layer discrete wavelet transform to obtain detail components and approximate components. The db5 wavelet function is used. For a signal with a 10kHz sampling frequency, the frequency bands of the detail components in each layer correspond to 5-2.5kHz, 2.5-1.25kHz, 1.25-0.625kHz, 0.625-0.313kHz, 0.313-0.156kHz, and 0.156-0.078kHz, respectively. The sixth layer of approximate components corresponds to 0-0.078kHz. Early-stage cable faults, such as partial discharge, typically generate high-frequency signals and can be detected by sudden energy changes in the detail components in the first few layers. System overloads or short circuits, on the other hand, manifest primarily as changes in low-frequency signals and can be detected by changes in the detail and approximate components in the last few layers.

[0078] By calculating the dynamic thresholds for each component energy, the system adaptively determines whether a current disturbance has occurred. Upon detecting the first sudden disturbance, the system begins counting and monitors for a second sudden disturbance within a time window of 0.5 to 5 cycles (10-100ms). This time window is based on the typical characteristics of early-stage cable faults: self-recovering half-cycle and multi-cycle single-phase-to-ground faults typically clear automatically within a short period of time.

[0079] In another optional embodiment, the number of wavelet decomposition layers and the selected wavelet function can be flexibly adjusted for different cable types and operating environments. For example, in environments with high signal noise, the sym8 wavelet, with its better smoothing properties, can be selected; in situations requiring finer temporal resolution, the number of decomposition layers can be increased to 8 or 10. Furthermore, the time window for determining sudden disturbance changes can be adjusted based on actual operating experience. For older cables with poor insulation, it can be shortened to 0.3 to 3 cycles to improve detection sensitivity.

[0080] It should be noted that the design of the composite criterion takes into account the diverse manifestations of early-stage cable faults. By analyzing characteristics in both high-frequency and low-frequency bands, it can comprehensively capture the characteristics of different fault types. The high-frequency criterion is more sensitive when a cable experiences transient insulation breakdown, while the low-frequency criterion is more effective when the cable experiences gradual insulation degradation. The combined use of these two criteria significantly improves the comprehensiveness and reliability of fault detection, reducing missed and false detections.

[0081] In a preferred embodiment, the pre-processed cable current signal is first decomposed by a 6-layer discrete wavelet transform to obtain detail components and approximate components. The detail components and approximate components can be expressed as:

[0082] Based on the feature extraction of wavelet transformation, the sampled cable current sampling signal is decomposed into detail components d through 6 layers of discrete wavelet transformation. j,n and approximate component c j,n .

[0083] They can be expressed as:

[0084]

[0085] g(n) and h(n) are high-pass and low-pass filters, i is the serial number of the sampling point in the current window, j is the number of wavelet coefficient layers of the signal, and n is the discretization degree of the wavelet function.

[0086] The cable current signal is decomposed into wavelet features in different frequency bands as the basis for measuring the current transient process. The energy E of the j-th layer detail coefficient within its frequency band is d,j It can be expressed as:

[0087]

[0088] For the detail component d j,n and approximate component c j,n , i is the number of the sampling point in the current window, j is the number of wavelet coefficient layers of the signal, h and g are high-pass and low-pass filters, respectively. This multi-scale decomposition can decompose the original signal into different frequency bands, making it easier to capture fault characteristics within different frequency ranges.

[0089] In cable early fault detection, the high-frequency detail coefficient energy criterion is primarily used to identify faults that generate high-frequency pulses, such as partial discharge, while the low-frequency approximation coefficient root mean square value criterion is suitable for detecting low-frequency signal changes caused by insulation aging and other factors. The combination of these two criteria forms a composite criterion that comprehensively covers different types of early fault characteristics.

[0090] After detecting the first sudden disturbance, the system begins monitoring for a second disturbance within a time window of 0.5 to 5 cycles. This time window is based on the typical characteristics of early-stage cable faults: self-recovering half-cycle and multi-cycle single-phase-to-ground faults typically clear automatically within a short period of time. If a second disturbance is detected within this time window, it is considered an early-stage cable fault, and the current data between the two disturbances is used as the raw data for fault identification.

[0091] In another optional embodiment, wavelet decomposition parameters can be optimized based on the cable type and operating environment. For older cables, the threshold can be lowered to improve detection sensitivity; for newly installed cables, the threshold can be appropriately increased to reduce false positives. Furthermore, different wavelet basis functions have different adaptability to different fault signals. Wavelet basis functions such as db4, sym8, or coif3 can be compared to select the most appropriate one for the current monitoring target.

[0092] It should be noted that the dynamic threshold-driven cable overcurrent transient detection method overcomes the limitations of traditional fixed threshold methods and can automatically adjust detection sensitivity based on the cable's operating environment and status. The design of the composite criterion considers the various manifestations of early-stage cable faults, improving the comprehensiveness and accuracy of detection. The time window setting of 0.5 to 5 cycles captures the entire fault process while avoiding misidentifying normal operating disturbances as faults. This method is particularly suitable for identifying self-recovering transient faults, providing the possibility of early detection of potential cable problems.

[0093] B2: The threshold value of the high-frequency detail coefficient energy criterion includes a temperature compensation coefficient, which is dynamically adjusted according to the real-time monitored temperature of the cable surface and the reference benchmark temperature.

[0094] In a preferred embodiment, the high-frequency detail coefficient energy criterion adopts a dynamic threshold design, and the threshold calculation refers to the following formulas (3)(4)(5).

[0095] The root mean square value R of the approximation coefficient of the jth layer c,j It can be expressed as:

[0096]

[0097] After the cable current signal is decomposed by multi-scale wavelet, the wavelet coefficient modulus maximum point corresponds to the current signal mutation point, which can detect the mutation moment of the cable transient overcurrent. The wavelet function is selected as the first-order derivative of the smoothing function, and the wavelet transform modulus maximum M is d,j It is the maximum value of the convolution of the sampling signal and the wavelet function.

[0098] M d,j =max(|d j,n |) (4)

[0099] Because early-stage cable faults are self-recovering half-cycle and multi-cycle single-phase grounding faults that automatically clear within 0.5 to 5 cycles, a composite criterion is constructed based on the relationship between high-frequency detail energy, low-frequency trend RMS, and dynamic threshold. If either of criteria 1 or 2 is met, the cable current is considered to have experienced a sudden disturbance.

[0100] Criterion 1: High-frequency detail coefficient energy criterion.

[0101]

[0102] l is the current sampling window number, E d (n) is the high-frequency detail energy value when the current window is n, is the current mutation criterion 1 threshold.

[0103]

[0104] σ(E d (n~ln)) is the standard deviation of the energy in the window n~ln, is the mean value of energy in window n~ln, T env is the temperature compensation coefficient, and α is a set constant.

[0105] Among them, the temperature compensation coefficient formula is:

[0106] T env =β·(T current -T ref ) (7)

[0107] Where: T current Real-time monitoring of cable surface temperature; T ref is the reference temperature; β is the temperature compensation coefficient, which is determined by the cable material properties and experimental calibration. Typical values ​​for common cables are: XLPE cable: β = 0.03 ~ 0.12 J / ℃; PVC cable: β = 0.05 ~ 0.15 J / ℃. Under extreme conditions, β can be halved to avoid overcompensation.

[0108] The constant α is set in the range of [0.1-0.5]. The optimal value can be determined by ROC curve analysis. For every 10% increase in load rate, α decreases by 2%.

[0109] The introduction of a temperature compensation factor is based on the principle that the electrical properties of cable insulation materials change with temperature. As temperature rises, the dielectric constant and loss tangent of the insulation material increase, leading to a decrease in insulation performance and a greater susceptibility to premature failures such as partial discharge. By monitoring the cable surface temperature in real time and comparing it with a reference temperature, the system automatically adjusts the fault judgment threshold, improving the detection's environmental adaptability.

[0110] In practical applications, β values ​​vary for different cable types: for XLPE cables, the β value range is 0.03–0.12 J / °C; for PVC cables, the β value range is 0.05–0.15 J / °C. For cables operating under extreme temperature conditions, the β value can be halved to avoid overcompensation. The constant β is set within the range of [0.1–0.5], and the optimal value can be determined through ROC curve analysis.

[0111] In another optional embodiment, the influence of cable load changes can be factored in. As the cable load factor increases, the β value needs to be appropriately reduced to prevent oversensitivity and false alarms under high load conditions. Specifically, for every 10% increase in load factor, the β value decreases by approximately 2%. Furthermore, the effect of cable aging can be considered, and the threshold can be appropriately lowered for older cables to improve detection sensitivity.

[0112] It should be noted that the dynamic adjustment mechanism of the temperature compensation coefficient is one of the key innovations of this invention. It enables the system to automatically optimize fault detection parameters based on the actual cable operating environment, significantly improving detection accuracy and reliability. This dynamic adjustment is particularly important in environments with large temperature fluctuations, effectively avoiding false alarms or missed alarms caused by temperature changes. The system should regularly calibrate the temperature sensor to ensure the accuracy of temperature measurement and thus the rationality of threshold calculation.

[0113] B3: The threshold of the low-frequency approximation coefficient RMS value criterion includes the temperature deviation coefficient and the load rate deviation coefficient, and is dynamically adjusted according to the ambient temperature, cable thermal resistance coefficient, cable rated operating temperature, and load rate.

[0114] In a preferred embodiment, the low-frequency approximation coefficient root mean square value criterion adopts a dynamic threshold design.

[0115] Criterion 2: Low-frequency approximation coefficient root mean square value criterion.

[0116]

[0117] m is the number of sampling points in one cycle, is the threshold of criterion 2, and its calculation method is shown in formula (8).

[0118]

[0119] ΔT is the temperature deviation coefficient, ΔL is the load rate deviation coefficient, and χ and δ are correction weight coefficients.

[0120] In different environments, when the harmonic distortion rate is greater than 20%, the load factor deviation coefficient is amplified by 10% to 20%, and a 10-second sliding average of ΔL is performed to suppress the impact of transient fluctuations. Load fluctuations, cable aging, harmonic content, and three-item imbalance all affect the load factor deviation coefficient. The load factor deviation coefficient formula is as follows:

[0121]

[0122] Where: N is the number of sampling points in the sliding window; I k is the effective value of current; I rated is the rated current carrying capacity of the cable; is the average load rate within the window.

[0123] In response to different working environments, T is updated every 24 hours based on the ambient temperature. ref ; Ambient temperature, load current, cable material and heat dissipation conditions will affect the temperature deviation coefficient. The temperature deviation coefficient formula is as follows:

[0124]

[0125] Where: T ambient is the ambient temperature; R thermal is the thermal resistance coefficient of the cable; T ref is the rated operating temperature of the cable.

[0126] The weight coefficient does not change much due to the influence of environment and working conditions; the temperature correction weight coefficient χ is generally taken as 0.01-0.05, which can be determined by temperature rise test; the load rate weight coefficient δ is generally taken as 0.1-0.3, which can be determined by load step test.

[0127] In the overcurrent transient detection process, the occurrence of the first transient current rise is first determined, and then it is determined whether the second transient current change occurs within the next 0.5 to 5 cycles. If the judgment condition for the first current mutation is met, the current window time t1 is recorded; then, within 0.5 to 5 cycles after t1, it is determined whether the second current mutation occurs. The judgment criterion is the same as the first judgment, but the threshold value is different. The threshold calculation method of judgment criterion 1 in the second judgment is shown in formula (9).

[0128]

[0129] The threshold calculation method of criterion 2 in the second discrimination is shown in formula (10).

[0130]

[0131] Similarly, the time t2 when the second current mutation occurs is recorded, and the current data between t1 and t2 is used as the original data for early cable fault identification.

[0132] The system adjusts parameters based on specific environments. When harmonic distortion exceeds 20%, the load factor deviation coefficient is amplified by 10% to 20%, and a 10-second sliding average of λL is applied to mitigate the impact of transient fluctuations. Every 24 hours, the system updates the temperature deviation coefficient based on ambient temperature to ensure long-term monitoring accuracy.

[0133] The selection of weight coefficients is crucial to threshold calculation: the temperature correction weight coefficient α1 is generally 0.01-0.05, which can be determined through temperature rise tests; the load rate weight coefficient α2 is generally 0.1-0.3, which can be determined through load step tests.

[0134] In another optional embodiment, the threshold calculation can be further optimized based on the actual cable installation environment and operating conditions. For direct-buried cables, the soil moisture factor can be added. Increased humidity can lead to poor heat dissipation, requiring a lower threshold. For overhead cables, the wind speed factor can be added. Increased wind speed improves heat dissipation, requiring a higher threshold. Furthermore, for power cables that frequently start and stop equipment, a startup surge compensation factor can be added to prevent startup transient currents from being misidentified as faults.

[0135] It should be noted that the low-frequency approximation coefficient RMS value criterion is primarily used to detect abnormal power-frequency current in cables and is sensitive to impedance changes caused by insulation aging, poor connector contact, and other factors. By introducing temperature and load factor deviation coefficients, this criterion dynamically adjusts the threshold based on the cable's actual operating environment, avoiding the limitations of fixed thresholds under varying operating conditions. This dynamic threshold design significantly improves the system's adaptability to changes in cable operating conditions, making fault detection more accurate and reliable, and particularly suitable for long-term online monitoring applications.

[0136] In the embodiment of the present application, the step S300 is to perform S transformation on the original data to construct an initial feature set, including the following steps C1-C2:

[0137] C1: The initial feature set constructed by S transformation includes three aspects of features: statistics, entropy and energy, specifically including six feature quantities: mean, standard deviation, skewness, kurtosis, root mean square value and logarithmic energy entropy.

[0138] The original data of cable early fault identification is subjected to S-transformation. The S-transformation result of the continuous-time signal x(t) is shown below.

[0139]

[0140] ω(τ-t,f) is the Gaussian window function, τ is the center position of the Gaussian window function, which is used to control the translation of the Gaussian window, f is the frequency, and δ(f) is the Gaussian window width. Different time-frequency resolutions can be obtained by changing the value of frequency f.

[0141] After the original signal undergoes S-transformation, a complex time-frequency matrix will be generated, in which rows and columns correspond to different frequencies and different moments respectively. After taking the module of the matrix, the corresponding module time-frequency matrix is ​​obtained. The row vector in the matrix represents the change process of the signal amplitude in the time series at the frequency corresponding to the current row, and the column vector represents the amplitude of the signal at different frequencies at the current moment.

[0142] As δ(f) decreases, the window width becomes narrower and the time resolution improves, which is suitable for capturing the rapidly changing components in the transient signal of the cable fault. As δ(f) increases, the window width becomes wider and the frequency resolution improves, which is suitable for analyzing low-frequency background noise or steady-state signals. The window width selection is driven by the signal characteristics. A high-frequency narrow window can accurately locate the arrival time of the fault wave head, and a low-frequency wide window can suppress the low-frequency oscillation interference caused by the cable distributed parameters.

[0143] τ→kT; T is the sampling interval, N is the total number of sampling points; k, m = 0, 1, 2, ..., N-1 and n = 1, 2, ..., N-1.

[0144] The specific implementation steps of S transform in cable fault signal processing are as follows:

[0145] Obtain the original data of the cable fault and convert the continuous signal into a discrete sequence. The sampling frequency must satisfy the Nyquist theorem.

[0146] According to formula (15), the Gaussian window function is constructed, and the window width is dynamically adjusted according to formula (16);

[0147] Each discrete frequency is calculated according to formula (14) to generate a complex time-frequency matrix.

[0148] After the original signal undergoes S-transformation, a complex time-frequency matrix will be generated, in which rows and columns correspond to different frequencies and different times respectively. After taking the module of this matrix, the corresponding module time-frequency matrix can be obtained.

[0149] The row vectors in the matrix represent how the signal amplitude changes over time at a specific frequency. For example, a higher frequency at f = 1 MHz reflects the transient impact of the traveling wave from a cable fault, while a lower frequency at f = 1 kHz reflects low-frequency harmonics or power-frequency current disturbances caused by cable insulation degradation. The column vectors represent the signal's spectral distribution at a specific moment. For example, at the moment of a fault, the high-frequency component suddenly increases, while the low-frequency component exhibits transient oscillations. At steady-state moments, the high-frequency component decays, while the low-frequency harmonic amplitudes continue to increase.

[0150] The S-transform of the original signal generates a complex time-frequency matrix, with rows and columns corresponding to different frequencies and different times, respectively. Taking the modulus of this matrix yields a modular time-frequency matrix. The row vectors in this matrix represent the time series variation of the signal amplitude at the frequency corresponding to the current row, and the column vectors represent the amplitude of the signal at different frequencies at the current moment. By analyzing these time-frequency characteristics, different types of cable faults can be effectively distinguished.

[0151] Based on the modular time-frequency matrix, the following six feature quantities are calculated to construct the initial feature set:

[0152] Mean: reflects the average energy level of the signal at a specific frequency.

[0153] Standard deviation: Indicates the degree of dispersion of signal energy distribution.

[0154] Skewness: describes the asymmetry of the signal distribution.

[0155] Kurtosis: reflects the sharpness of the signal distribution.

[0156] Root mean square value: represents the effective energy of the signal.

[0157] Logarithmic energy entropy: a measure of the complexity of a signal.

[0158] The calculation formulas of these six characteristic quantities are shown in Table 1 below, which describe the statistical characteristics, energy distribution and complexity of cable fault signals from different perspectives.

[0159] Figure 4 The corresponding relationship between the matrix and fault characteristics:

[0160] Cable short circuit, disconnection or partial discharge: A single peak surge occurs at a certain time point in the high frequency line f>100kHz. If secondary peaks of similar amplitude appear at subsequent times, it indicates that there are multiple reflections at the fault point.

[0161] Cable insulation aging, accumulated partial discharge, or ground fault: In the row vectors corresponding to the 50 Hz power frequency and odd harmonics, the amplitude is continuously higher than the background level, and attenuated oscillations occur in the 1 kHz-10 kHz frequency band.

[0162] Poor cable connector contact and arc discharge: In the 10kHz-1MHz frequency band, multiple consecutive frequency lines have random amplitude fluctuations without a clear dominant frequency, and the noise amplitude changes synchronously with the power frequency cycle.

[0163] After obtaining the time-frequency matrix, the signal amplitude at each frequency is calculated in terms of statistics, entropy, and energy. A total of 6 feature quantities are obtained for each frequency, as shown in Table 1.

[0164] Table 1 Extracted feature quantities

[0165]

[0166] To further reduce the matrix dimension, it is necessary to extract the main information features of the matrix to reduce the dimension. The Pearson correlation coefficient method is used to screen the optimal feature subset to improve recognition accuracy and fitting speed. It is used to measure the linear correlation between two variables. Its value range is between -1 and 1. The calculation formula is shown in Equation (14).

[0167]

[0168] In another optional embodiment, the S-transform parameters can be optimized based on different cable fault types. For partial discharge faults, the high-frequency component is more pronounced, so the number of sampling points in the high-frequency region can be increased. For insulation aging faults, the low-frequency characteristics are more pronounced, so the resolution in the low-frequency region can be increased. Furthermore, the frequency range can be adjusted based on the cable's operating voltage level; the characteristic fault frequency of high-voltage cables is generally higher than that of low-voltage cables.

[0169] It should be noted that the S-transform, as an advanced time-frequency analysis tool, offers significant advantages over the traditional Fourier transform. It not only preserves the signal's phase information but also features a frequency-adaptive window width. The six features constructed using the S-transform can comprehensively describe the various characteristics of cable fault signals, providing a rich source of information for subsequent feature selection and fault identification. The combination of these features can effectively distinguish different types of early-stage cable faults, improving the accuracy and reliability of fault identification.

[0170] C2: Based on the S-transform results, calculate the feature quantities at different frequencies to form a complete initial feature set matrix.

[0171] In a preferred embodiment, the characteristics of early cable faults vary in different frequency bands, so it is crucial to extract full-band features from the S-transform results. Based on the modulus time-frequency matrix obtained by the S-transform, multiple representative frequency points can be selected to calculate the above six feature quantities. For a signal with a sampling frequency of 10kHz, the following frequency points can be selected: 50Hz (power frequency), 150Hz (3rd harmonic), 250Hz (5th harmonic), 500Hz, 1kHz, 2kHz, and 5kHz, a total of 7 frequency points, with 6 feature quantities at each frequency point, forming an initial feature set matrix of 7×6=42 dimensions.

[0172] Different types of cable faults exhibit distinct frequency spectrum characteristics. Partial discharge faults manifest primarily as a sudden surge in high-frequency energy (above 1kHz); faults caused by insulation aging manifest as increased harmonic content, primarily in the low-frequency range (50Hz-500Hz); and poor cable joint contact manifests as energy fluctuations in the mid-frequency range (500Hz-2kHz). By extracting features across the entire frequency range, the system captures the frequency domain characteristics of different fault types, providing comprehensive information for subsequent fault classification.

[0173] During feature extraction, specific frequency bands can be analyzed for focus. For example, for partial discharge faults, additional sampling points can be added to the high-frequency band (1kHz-5kHz); for insulation aging faults, additional sampling points can be added to the low-frequency harmonics (150Hz-350Hz). This targeted frequency selection can improve the accuracy of identifying specific fault types.

[0174] In another optional embodiment, the interrelationships between frequencies can be considered to calculate frequency ratio features, such as the ratio of high-frequency energy to low-frequency energy. This is particularly effective in distinguishing external interference from internal faults. The rate of change of the characteristic quantity at each frequency point can also be calculated to reflect the dynamic process of fault development. For intermittent faults, the coefficient of fluctuation of the characteristic quantity (the ratio of the standard deviation to the mean) can be calculated to capture the instability characteristics of the fault.

[0175] It should be noted that constructing the initial feature set matrix forms the foundation for subsequent feature selection and fault identification. By combining multiple frequency points and multiple features, a high-dimensional feature space is formed that comprehensively describes the characteristics of cable faults. While this comprehensive feature extraction strategy increases computational complexity, it significantly improves the accuracy and robustness of fault identification, particularly in complex electromagnetic environments. However, high-dimensional features also introduce data redundancy and the "curse of dimensionality," necessitating optimization through subsequent feature selection steps. This is precisely the role of the Pearson correlation coefficient method in the next step.

[0176] In the embodiment of the present application, step S400 obtains the optimal feature subset and inputs it into the CNN-GRU-Attention neural network model to achieve intelligent recognition and classification of early-stage cable faults, including the following steps D1-D2:

[0177] D1: CNN-GRU-Attention neural network model includes:

[0178] The convolutional neural network (CNN) part is used to extract the spatial information of the input features;

[0179] The gated recurrent unit (GRU) part is used to process time series data;

[0180] The attention mechanism is used to weight the feature sequence output by GRU to improve the discrimination of feature importance;

[0181] The fully connected layer is used to output the recognition and classification results of early cable faults.

[0182] Combining CNNs with GRUs leverages the strengths of both. CNNs extract features from time series data through convolution and pooling, discovering hidden patterns and associations within the data and reducing prediction uncertainty when power fluctuates significantly. GRUs excel at processing time series data and capturing long-term dependencies. By using CNN-extracted features as input to the GRU, the GRU can better understand and utilize these features, supplementing any potential omissions in its information mining efforts and obtaining more comprehensive information from the input, enabling accurate fault classification.

[0183] The CNN-GRU structure is divided into three parts. In the first part, data feature information is extracted through convolution and a convolutional neural network is constructed.

[0184] The second part is a network that uses GRU (Gated Recurrent Unit) to process time series data.

[0185] In the third part, the Attention mechanism is added to perform weighted average calculation on the output of GRU to obtain more accurate cable early fault identification accuracy.

[0186] The CNN network structure typically consists of several key components, including the input layer, convolutional layers, activation function layers, pooling layers, fully connected layers, and output layers. During the forward propagation process, the input data is processed sequentially through multiple convolutional and pooling layers to extract and refine key feature vectors. These feature vectors are then passed to the fully connected layers, where they undergo a series of weight adjustments and activation function processing, resulting in the final output.

[0187] The calculation formula of the output layer of the GRU neural network is shown in (15):

[0188] y=σ(w y h t ) (18)

[0189] Where y is the output of the neural network, w y is the weight, h t Candidate states of the network hidden layer.

[0190] The attention mechanism can improve the generalization ability of the model. By focusing on key information, the attention mechanism can help the prediction model learn more essential and universal feature representations, thereby improving the generalization performance of the model in new scenarios and new tasks. It can also deal with information overload problems, thereby improving processing efficiency. This is of great significance for processing complex and nonlinear cable early fault data. Its calculation formula is shown in Equation (16).

[0191]

[0192] Where h t is the output of the hidden layer of the prediction model, α t is the weight value, for h t Perform weighted summation.

[0193] Table 2. CNN-GRU network parameters

[0194]

[0195]

[0196] For each sample, 1000 data points were randomly selected as the experimental dataset, with 800 used as the training set and 200 used as the test set. The model parameters were: an initial learning rate of 0.004, a 4×4 convolution kernel, 109 neurons, 200 iterations, all-zero padding for the second convolution layer and the max pooling layer, the Adam gradient descent optimizer, and the ReLU activation function. The L2 norm was introduced to suppress weights during training.

[0197] The specific steps are as follows:

[0198] (1) Data preprocessing and feature extraction: The original disturbance signal is extracted from the collected signal using the dynamic threshold method, and the feature vector is constructed using the S change and Pearson correlation coefficient method.

[0199] (2) CNN feature extraction: CNN receives the data obtained in step 1 and extracts feature vectors through convolution operation.

[0200] (3) GRU fault identification: Utilize the time series processing capability of GRU to capture the dynamic changes in the sequence and obtain preliminary fault identification results.

[0201] (4) Introducing the Attention mechanism: After the GRU processes the feature sequence, the Attention mechanism is introduced to weight the feature sequence output by the GRU. The weighted feature sequence can better highlight the information that has important contributions to the prediction results. The Attention mechanism can significantly enhance the influence of key time steps in the GRU, thereby effectively reducing the prediction error of the model.

[0202] (5) Final recognition result: The weighted feature sequence is further processed and the fault recognition result is finally output.

[0203] The specific identification steps are as follows: Figure 8 shown.

[0204] Table 2 defines the fault dataset:

[0205] Table 3 Fault type definition

[0206] Fault type Scenario Example Signal characteristics Partial discharge Air gap defects, metal particles High frequency pulse (1-30MHz) Arc fault Poor contact, disconnection Broadband noise (10kHz-1MHz) Insulation degradation Water tree growth and sheath damage Low frequency resonance (50Hz-5kHz) Mixed failure Discharge + arc, arc + insulation degradation Multi-band superposition Normal state / Power frequency + white noise (<1kHz)

[0207] Comparative experimental design:

[0208] Comparative test of feature selection schemes: The mutual information method and the random forest feature importance method were selected as the comparison methods. The same batch of test set samples were used, and the recognition model used was the model in this paper. Finally, the evaluation indicators classification accuracy and F1-score were used to analyze the results.

[0209] Deep learning model comparison test: CNN, GRU and transformer were selected as the comparison groups. The feature datasets were all datasets after Pearson dimensionality reduction, and the evaluation indicators were still accuracy and F1-score.

[0210] Evaluation of recognition performance under different fault types and degrees: Based on the fault type definitions in Table 2, a three-level quantitative standard is designed.

[0211] Level 1: Signal amplitude <5% full scale, characteristic frequency energy ratio <10%;

[0212] Level 2: Amplitude 5% to 20%, energy ratio 10% to 30%;

[0213] Level 3: Amplitude > 20%, Energy Ratio > 30%

[0214] After determining the data set and quantitative standards, we conduct recognition performance evaluation. We calculate the precision and recall of each subclass to evaluate the recognition performance of different fault types. We also calculate the F1-score variance of different levels of the same fault type to evaluate the recognition performance of different degrees of the same fault. The smaller the variance, the better the performance.

[0215] The contribution of the dynamic threshold module can be analyzed based on three indicators: (1) effective signal ratio = number of signals in the trigger period / total number of signals; (2) noise suppression rate = 1-(noise energy in the trigger period / noise energy in the entire period); (3) fault missed detection rate = number of real fault segments that are not triggered / total number of fault segments.

[0216] Contribution analysis of CNN-GRU-Attention: The CNN-GRU-Attention model in this paper was removed one module at a time, forming three comparison groups: CNN-GRU, CNN-Attention, and GRU-Attention. The classification accuracy score Macro-F1, temporal localization error, and F1 decay rate were calculated for each group to evaluate the contribution of each module.

[0217] Performance evaluation and stability analysis of the system under different operating conditions: Environmental, electrical, and signal conditions are changed in sequence, and performance evaluation is performed by calculating the following performance indicators:

[0218] Detection accuracy: number of correctly identified fault samples / total number of samples;

[0219] Response time: the time from signal input to output result;

[0220] Missed detection rate: number of unidentified real fault samples / total number of fault samples;

[0221] False alarm rate: the proportion of normal samples that are mistakenly judged as faults;

[0222] The stability indicators are as follows:

[0223] Long-term drift: the accuracy drop rate after 72 hours of continuous operation;

[0224] Temperature drift coefficient: the accuracy change ratio for every 10°C temperature change;

[0225] Load sensitivity: The rate of change of false alarm rate when the load changes from 0% to 150%.

[0226] D2: Training and optimization process of CNN-GRU-Attention neural network model.

[0227] In a preferred embodiment, model training uses mini-batch stochastic gradient descent (SGD) with the Adam optimizer for weight updates. The initial learning rate is set to 0.004, and as training progresses, a learning rate decay strategy is adopted, reducing the learning rate by a factor of 0.8 every 50 epochs. The training batch size is set to 32, and the total number of training epochs is 200. To prevent overfitting, the network uses L2 regularization with a regularization coefficient of 0.0001. A dropout layer with a dropout rate of 0.5 is introduced before the fully connected layer.

[0228] The loss function uses cross entropy loss, which is suitable for multi-classification problems. For cases of class imbalance, class weighting can be introduced to increase the loss weight of minority class samples, improving the model's ability to identify minority classes. For example, if the number of mixed fault samples is smaller than that of other types of faults, the weight of the mixed fault class can be appropriately increased.

[0229] During training, we used an early stopping strategy to avoid overfitting. We split the dataset into a training set and a validation set in an 8:2 ratio. We evaluated model performance on the validation set after every five epochs. If performance on the validation set did not improve after 10 consecutive epochs, we stopped training and saved the model parameters with the best validation set performance.

[0230] Model evaluation uses a comprehensive approach using multiple metrics, including accuracy, precision, recall, and F1-score. For multi-classification problems, performance metrics for each category, as well as overall macro-average and micro-average metrics, are calculated to comprehensively assess model performance. For cable fault identification, particular attention is paid to the confusion matrix, analyzing the confusion between different fault types and enabling targeted model improvements.

[0231] In another optional embodiment, a k-fold cross validation method can be used to evaluate the generalization performance of the model, typically with k = 5 or 10. By training and evaluating the model multiple times on different training and test sets, a more stable and reliable performance estimate can be obtained. In addition, ensemble learning techniques such as voting or bagging can be used to integrate the prediction results of multiple models to improve overall prediction performance and stability.

[0232] In the embodiment of the present application, step S500 obtains the optimal feature subset and inputs it into the CNN-GRU-Attention neural network model to achieve intelligent recognition and classification of early-stage cable faults, including the following steps E1-E2:

[0233] E1: The CNN-GRU-Attention neural network model is used to identify different types of cable early faults, such as partial discharge, arc fault, insulation degradation, and mixed faults.

[0234] In a preferred embodiment, early-stage cable faults are mainly divided into four types: partial discharge, arc fault, insulation degradation and mixed fault, and each fault has different characteristic manifestations and mechanisms.

[0235] Partial discharge (PD) faults typically occur in cable insulation caused by air pockets, foreign matter, or partial breakdown, manifesting as high-frequency (1-30 MHz) pulse signals. Initially, these faults manifest as intermittent, small discharges. Over time, as insulation deteriorates, the discharge intensity and frequency gradually increase. The characteristics of PD faults are typically evident in the high-frequency region of the S-transform, including high kurtosis, sudden energy increases, and significant changes in entropy. The CNN-GRU-Attention model can capture the spatiotemporal characteristics of these high-frequency pulses and accurately identify PD faults.

[0236] Arc faults primarily occur at poor contact areas at cable joints or connection points, manifesting as broadband noise (10kHz-1MHz). These faults are typically caused by mechanical stress, overheating, or corrosion. Initially, they manifest as intermittent poor contact, generating a weak arc. As the contact deteriorates, the arc intensity and duration increase, ultimately leading to a complete circuit break. The characteristics of arc faults are evident in the intermediate frequency region of the S-transform, including randomly fluctuating energy distribution, high entropy, and high standard deviation. The model can effectively distinguish arc faults from other types of faults by analyzing these time-frequency characteristics.

[0237] Insulation degradation faults are caused by the gradual aging of cable insulation materials over time or by environmental factors (such as moisture and chemical intrusion). They manifest as low-frequency resonance (50Hz-5kHz) and increased harmonics. This fault develops slowly, initially manifesting as a decrease in insulation resistance and increased dielectric loss, which gradually leads to increased leakage current and a decrease in insulation strength. The characteristics of insulation degradation faults are primarily concentrated in the low-frequency region of the S-transform, including increased harmonic content, elevated low-frequency energy, and changes in statistical characteristics. By analyzing these low-frequency characteristics, the model can detect insulation degradation trends early on.

[0238] Mixed faults refer to complex situations where multiple fault types coexist, such as partial discharge accompanied by insulation degradation or arc faults accompanied by partial discharge. The characteristics of these faults are more complex, manifesting as the superposition of features from multiple frequency bands. The CNN-GRU-Attention model, through multi-level feature extraction and an attention mechanism, can simultaneously capture fault features from different frequency bands and assign appropriate weights to different features, enabling accurate identification of mixed faults.

[0239] The CNN-GRU-Attention model is adaptable to the time-frequency characteristics of different fault types. The CNN extracts spatial features and local patterns, making it particularly suitable for capturing the pulse characteristics of high-frequency partial discharges. The GRU captures temporal dependencies, making it suitable for analyzing the development of insulation degradation and arc faults. The Attention component automatically adjusts its focus on different time steps and features based on the fault type, improving the model's sensitivity to various fault types.

[0240] In another optional embodiment, fault types and levels can be further subdivided. For example, partial discharge can be classified into three levels based on discharge intensity: mild, moderate, and severe; and insulation degradation can be classified into initial, developing, and critical stages based on the degree of aging. This more refined classification can provide a more detailed assessment of cable condition, providing more basis for maintenance decisions.

[0241] For new cable materials and special operating environments, transfer learning can be used to leverage pre-trained models and fine-tune them for the new situation. This approach can significantly reduce the workload of data collection and model training, allowing for rapid adaptation to new application scenarios.

[0242] It should be noted that different types of cable early-stage faults exhibit significant differences in their physical mechanisms and signal characteristics, requiring models with sufficient flexibility and expressiveness to distinguish these differences. The CNN-GRU-Attention model, through its deep feature extraction and time series analysis capabilities, can effectively identify various types of early-stage faults, particularly subtle and complex ones that are difficult to detect using traditional methods. This comprehensive fault identification capability provides strong support for cable condition monitoring and preventive maintenance, significantly reducing operational risks and maintenance costs in power systems.

[0243] E2: Deployment and application scenarios of the CNN-GRU-Attention neural network model.

[0244] In a preferred embodiment, the trained CNN-GRU-Attention model can be deployed on edge computing devices or cloud servers to enable real-time monitoring and early warning of early cable faults. For edge deployment, the model can be converted to a lightweight version, using techniques such as model pruning, knowledge distillation, or quantization to reduce computational complexity and memory usage, enabling it to run on resource-constrained terminal devices. This solution is suitable for remote cable monitoring points, enabling local preliminary fault detection and transmitting only critical information back to the central control center, saving communication bandwidth and data storage space.

[0245] Cloud deployments can leverage more powerful computing resources to run the full model, enabling more sophisticated and comprehensive analysis. Cloud systems can simultaneously process data from multiple monitoring points, enabling comprehensive monitoring and analysis of large-scale cable networks. Through integration with power system management software, fault detection results can be directly used for operations and maintenance scheduling and decision support, achieving closed-loop management of cable condition monitoring.

[0246] In practice, the system can implement a multi-level early warning mechanism. For minor early signs of failure, the system issues a monitoring reminder, recommending increased monitoring frequency and increased attention. For clear early failures, the system issues a maintenance warning, recommending inspection and maintenance during the next scheduled power outage. For severe, worsening failures, the system issues an emergency warning, recommending immediate inspection and necessary emergency treatment. This tiered early warning mechanism balances operational efficiency and system reliability, enabling refined preventive maintenance.

[0247] As the system operates longer, more failure cases and operational data are collected, enabling the model to be continuously updated and optimized. Through incremental or online learning, the model can adapt to new failure modes and changes in cable characteristics, maintaining identification accuracy. Furthermore, the system incorporates information such as the cable's operating environment, load history, and installation age to construct a comprehensive assessment model to predict the cable's remaining service life and failure risk, providing a scientific basis for asset management and replacement planning.

[0248] In summary, by organically combining S-transform feature extraction, Pearson correlation coefficient feature selection, and the CNN-GRU-Attention deep learning model, the present invention constructs an efficient intelligent framework for early-stage cable fault identification. The S-transform can simultaneously provide time-domain and frequency-domain information of the signal, capturing the unique time-frequency characteristics of different types of faults. The Pearson correlation coefficient method effectively reduces the dimensionality of the input data and improves the fitting speed by screening key features. The CNN-GRU-Attention model fully leverages the respective advantages of convolutional neural networks in extracting spatial features, gated recurrent units in processing time series data, and the attention mechanism in highlighting key information. This innovative combination of technologies enables the system to accurately identify various types of early-stage cable faults, such as partial discharge, arc faults, insulation degradation, and mixed faults, maintaining high recognition accuracy even in complex electromagnetic environments and changing operating conditions. Compared with traditional methods, the present invention realizes intelligent and automated cable fault identification, significantly improving recognition accuracy and shortening response time. This provides a scientific basis for preventive cable maintenance, effectively extending cable service life, and reducing the operation and maintenance costs of power systems.

[0249] Example 3

[0250] The above is a schematic diagram of a method for intelligently identifying early-stage cable faults based on a dynamic threshold. It should be noted that the technical solution of this system for intelligently identifying early-stage cable faults based on a dynamic threshold is based on the same concept as the technical solution of the method for intelligently identifying early-stage cable faults based on a dynamic threshold. For details not described in detail in the technical solution of the system for intelligently identifying early-stage cable faults based on a dynamic threshold in this embodiment, please refer to the description of the technical solution of the method for intelligently identifying early-stage cable faults based on a dynamic threshold.

[0251] This embodiment further provides a system for intelligently identifying early-stage cable faults based on a dynamic threshold, including:

[0252] A signal acquisition module is used to obtain a cable current sampling signal;

[0253] The disturbance detection module is used to detect the transient process of cable overcurrent based on dynamic threshold drive, judge the cable current disturbance, and identify the raw data of early cable faults;

[0254] A feature extraction module is used to perform S transformation on the original data to construct an initial feature set;

[0255] A feature selection module is used to perform feature selection on the initial feature set by using the Pearson correlation coefficient method to select the optimal feature subset;

[0256] The fault identification module is used to input the optimal feature subset into the CNN-GRU-Attention neural network model to realize intelligent identification and classification of early cable faults.

[0257] This embodiment also provides an electronic device suitable for intelligent identification of early cable faults based on dynamic thresholds, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute computer-executable instructions to implement the method for intelligent identification of early cable faults based on dynamic thresholds proposed in the above embodiment.

[0258] This embodiment further provides a storage medium storing a computer program. When the program is executed by a processor, the method for intelligently identifying early-stage cable faults based on dynamic thresholds as proposed in the above embodiment is implemented.

[0259] The storage medium proposed in this embodiment and the method for realizing intelligent identification of early cable faults based on dynamic thresholds proposed in the above embodiment belong to the same inventive concept. For technical details not fully described in this embodiment, please refer to the above embodiment, and this embodiment has the same beneficial effects as the above embodiment.

[0260] Through the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented with the help of software and necessary general hardware, and of course can also be implemented by hardware. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory (FLASH), hard disk or optical disk, etc., including a number of instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute the methods of various embodiments of the present invention.

[0261] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for intelligently identifying early-stage cable faults based on dynamic thresholds, characterized by: Including, obtaining cable current sampling signal; Cable overcurrent transient process detection based on dynamic threshold drive, judging cable current disturbance, and identifying raw data of early cable faults; Performing S transformation on the original data to construct an initial feature set; Perform feature selection on the initial feature set by using the Pearson correlation coefficient method to select the optimal feature subset; The optimal feature subset is input into the CNN-GRU-Attention neural network model to achieve intelligent recognition and classification of early cable faults.

2. The method for intelligently identifying early-stage cable faults based on dynamic thresholds according to claim 1, characterized in that: The cable overcurrent transient process detection based on dynamic threshold driving includes: A composite criterion is formed by the high-frequency detail coefficient energy criterion and the low-frequency approximation coefficient root mean square value criterion; When any one of the high-frequency detail coefficient energy criterion or the low-frequency approximation coefficient root mean square value criterion is met, it is determined that a sudden disturbance occurs in the cable current; Determine whether a second transient current change occurs within 0.5 to 5 cycles after the first disturbance mutation occurs; When the second disturbance mutation occurs, the current data between the two disturbance mutations are used as the original data for early cable fault identification.

3. The method for intelligently identifying early-stage cable faults based on dynamic thresholds according to claim 2, characterized in that: The threshold value of the high-frequency detail coefficient energy criterion includes a temperature compensation coefficient, and the temperature compensation coefficient is dynamically adjusted according to the real-time monitoring temperature of the cable surface and the reference benchmark temperature.

4. The method for intelligently identifying early-stage cable faults based on dynamic thresholds according to claim 3, wherein: The threshold of the low-frequency approximation coefficient root mean square value criterion includes a temperature deviation coefficient and a load rate deviation coefficient, and is dynamically adjusted according to the ambient temperature, the cable thermal resistance coefficient, the cable rated operating temperature and the load rate.

5. The method for intelligently identifying early-stage cable faults based on dynamic thresholds according to claim 4, characterized in that: The initial feature set constructed by the S-transformation includes features in three aspects: statistics, entropy and energy, specifically including six feature quantities: mean, standard deviation, skewness, kurtosis, root mean square value and logarithmic energy entropy.

6. The method for intelligently identifying early-stage cable faults based on dynamic thresholds according to claim 5, characterized in that: The CNN-GRU-Attention neural network model includes: The convolutional neural network (CNN) part is used to extract the spatial information of the input features; The gated recurrent unit (GRU) part is used to process time series data; The attention mechanism is used to weight the feature sequence output by GRU to improve the discrimination of feature importance; The fully connected layer is used to output the recognition and classification results of early cable faults.

7. The method for intelligently identifying early-stage cable faults based on dynamic thresholds according to claim 6, characterized in that: The CNN-GRU-Attention neural network model is used to identify different types of cable early faults, such as partial discharge, arc fault, insulation degradation, and mixed faults.

8. A system for intelligently identifying early-stage cable faults based on a dynamic threshold, based on the method for intelligently identifying early-stage cable faults based on a dynamic threshold according to any one of claims 1 to 7, characterized in that: It also includes a signal acquisition module for acquiring a cable current sampling signal; The disturbance detection module is used to detect the transient process of cable overcurrent based on dynamic threshold drive, judge the cable current disturbance, and identify the raw data of early cable faults; A feature extraction module is used to perform S transformation on the original data to construct an initial feature set; A feature selection module is used to perform feature selection on the initial feature set by using the Pearson correlation coefficient method to select the optimal feature subset; The fault identification module is used to input the optimal feature subset into the CNN-GRU-Attention neural network model to realize intelligent identification and classification of early cable faults.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method for intelligently identifying early-stage cable faults based on dynamic thresholds according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for intelligently identifying early-stage cable faults based on dynamic thresholds according to any one of claims 1 to 7 are implemented.

Images