A cable testing method and a small handheld cable tester

Abstract

This invention provides a cable testing method and a small handheld cable tester. The method includes: performing wavelet transform noise reduction, Z-score normalization, and time-series dynamic normalization on the raw cable data; then constructing a geometric deep learning network using the dispersed self-organizing structure of a rotating cube set to extract multimodal deep features; next, fusing different modal features through a spike neural network gating mechanism and a cross-modal attention mechanism; finally, performing fault classification based on a differential evolution optimized neural network, and achieving fault location by combining a temporal generative adversarial network and a dynamic probabilistic neighborhood growing clustering algorithm. The system also evaluates the contribution of each modality of data through information entropy and mutual information analysis, and optimizes model performance using adaptive weight allocation and lightweight neural network technology. This invention can accurately identify and locate various fault types such as cable breaks, short circuits, insulation aging, and poor contact, significantly improving the efficiency and accuracy of cable maintenance.

Description

Technical Field

[0001] This invention relates to the field of cable testing technology, and in particular to a cable testing method and a small handheld cable tester for multimodal data acquisition, analysis and fault diagnosis of cables. Background Technology

[0002] Common cable fault detection techniques mainly include time-domain reflectometry (TDR) and frequency-domain analysis. TDR locates the fault point by sending pulse signals and analyzing the reflected waveforms, while frequency-domain analysis determines the fault type by measuring the impedance characteristics of the cable at different frequencies. These traditional methods each have their advantages in a single operating mode, but they often exhibit limitations in complex environments.

[0003] Existing advanced cable testing technologies attempt to combine multiple testing methods to improve accuracy through simple data fusion. These technologies typically employ linear fusion or rule-based judgment methods, combining and analyzing data from different testing modules to enhance the reliability of fault diagnosis. However, this simple fusion approach cannot fully explore the deep correlations between multimodal data, especially when dealing with complex faults involving nonlinear relationships and high uncertainty, where accuracy and robustness significantly decrease.

[0004] In addition, the existing technology has the following technical defects: First, it lacks an effective integration mechanism for multi-source heterogeneous data, resulting in insufficient detection accuracy; second, the algorithm complexity is high, making it difficult to achieve real-time analysis on small handheld devices; and third, it is not adaptable to different types of faults, which is particularly evident in the detection of hidden faults such as cable aging and poor contact, thus restricting the application effect and scope of cable testing technology. Summary of the Invention

[0005] The purpose of this invention is to overcome the above-mentioned defects in the prior art and provide a cable testing method and a small handheld cable tester. Through multimodal data acquisition and fusion, geometric deep learning feature extraction, neural network gating mechanism and adaptive weight optimization, it can achieve high-precision identification and location of various fault types such as cable breakage, short circuit, insulation aging and poor contact.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] This invention provides a cable testing method, comprising:

[0008] Acquire raw cable data in multiple modes, perform multi-dimensional data acquisition on the cable, and obtain raw multi-mode cable data stream containing reflection waveform, frequency response characteristics, impedance change curve and insulation resistance value;

[0009] The original multimode cable data stream is subjected to noise reduction and standardization operations. Through wavelet transform noise reduction, Z-score standardization and time series dynamic normalization, a preprocessed multimode cable test dataset is obtained.

[0010] Based on the preprocessed multimodal cable test dataset, a geometric deep learning network is constructed using the dispersed self-organizing structure of a set of rotating cubes. The deep features of different modal data are extracted through geometric convolution operations and self-organizing learning mechanisms to obtain a multidimensional cable feature vector set.

[0011] Based on the multidimensional cable feature vector set, a spike neural network gating mechanism is constructed, and different modal features are fused through a cross-modal attention mechanism to obtain a unified feature representation after fusion;

[0012] Based on the unified feature representation after fusion, fault classification is performed using a backpropagation neural network model optimized by differential evolution, and fault localization is performed by combining temporal generative adversarial network to enhance samples and dynamic probabilistic neighborhood growth clustering algorithm to obtain fault type judgment results and fault location coordinates.

[0013] Based on the fault type judgment result and the fault location coordinates, the contribution of each modality data is evaluated through information entropy and mutual information analysis. The model is then optimized using adaptive weight allocation and lightweight neural network technology to obtain the final cable fault detection report.

[0014] Preferably, the noise reduction and standardization operations performed on the original multimode cable data stream, through wavelet transform noise reduction, Z-score standardization, and time series dynamic normalization, yield a preprocessed multimode cable test dataset, including:

[0015] Based on the original multimode cable data stream, an adaptive median filtering algorithm is used to remove environmental noise and measurement interference. Wavelet transform is used to decompose the signal and reconstruct the key frequency bands to obtain the noise-reduced cable test data.

[0016] The noise-reduced cable test data is processed using Z-score standardization to address the dimensional differences between different modal data, and the time series dynamic warping algorithm is used to solve the problem of asynchronous sampling of multimodal data, resulting in a standardized aligned dataset.

[0017] Based on the standardized aligned dataset, fault features are enhanced and extracted using autoencoder and signal decomposition techniques to reconstruct lost or incomplete data information, thereby obtaining the preprocessed multimode cable test dataset.

[0018] Preferably, the geometric deep learning network constructed using the distributed self-organizing structure of the rotating cube set, extracts deep features of different modalities through geometric convolution operations and self-organizing learning mechanisms, and obtains a multidimensional cable feature vector set, including:

[0019] Based on the preprocessed multimodal cable test dataset, a cube unit group with rotatability, orientation sensitivity and spatial positioning is designed. Through geometric convolution operation, a structured representation of different modal data is established in a high-dimensional feature space to obtain the initial geometric feature mapping.

[0020] For the initial geometric feature mapping, a distributed self-organizing learning mechanism is applied. Through competitive learning, lateral inhibition, and topology preservation mapping, the distribution and connection relationship of the cube units are dynamically adjusted to form a self-organizing network that adapts to the characteristics of cable faults, thus obtaining a self-organizing feature representation.

[0021] Based on the self-organizing feature representation, dedicated feature extraction networks are designed for different modal data of time-domain reflection, frequency-domain response, impedance characteristics, and insulation characteristics. Combined with multi-scale feature pyramids and attention mechanisms, the multi-dimensional cable feature vector set is obtained.

[0022] Preferably, the construction of the spike neural network gating mechanism, which fuses features from different modalities through a cross-modal attention mechanism to obtain a unified feature representation after fusion, includes:

[0023] Based on the multidimensional cable feature vector set, a bio-inspired spike neural network model is constructed. The feature values ​​are converted into pulse sequences using a time-coding method. By simulating the neuron firing mechanism, the ability to capture the instantaneous characteristics of cable faults is enhanced, resulting in a pulse-coded feature representation.

[0024] For the pulse code feature representation, the design of updating gate control new information on memory state, the impact of resetting gate control historical information on current state, and memory gate management long-term dependencies are implemented. By working together with update gate, reset gate, and memory gate, dynamic filtering of information is achieved, resulting in the feature representation after gating.

[0025] Based on the feature representation after gating, a cross-modal attention mechanism is implemented. By calculating the correlation matrix between features of different modalities and generating attention weights, complementary information is highlighted and redundant features are suppressed, resulting in the fused unified feature representation.

[0026] Preferably, the fault classification using a backpropagation neural network model optimized by differential evolution, combined with temporal generative adversarial network-enhanced samples and dynamic probabilistic neighborhood growing clustering algorithm for fault localization, to obtain fault type judgment results and fault location coordinates, includes:

[0027] Based on the unified feature representation after fusion, a backpropagation neural network model optimized by differential evolution is constructed. The network weights and bias parameters are optimized by the global search capability of the differential evolution algorithm, thereby improving the ability to identify fault types such as open circuit, short circuit, insulation aging, and poor contact, and obtaining a preliminary fault type probability distribution.

[0028] Based on the preliminary fault type probability distribution, a temporal generative adversarial network model is used to learn the temporal features of various faults, generate diverse synthetic fault samples for data augmentation, improve the ability to identify rare faults and boundary cases, and obtain an enhanced fault feature library.

[0029] Based on the enhanced fault feature library and the preliminary fault type probability distribution, the dynamic probabilistic neighborhood growth and edge-preserving maximum clustering algorithm is applied to accurately locate the fault and identify the boundary through seed point selection, neighborhood growth and edge preservation constraints, so as to obtain the fault type judgment result and the fault location coordinates.

[0030] Preferably, the step of evaluating the contribution of each modality's data through information entropy and mutual information analysis, and optimizing the model using adaptive weight allocation and lightweight neural network technology to obtain the final cable fault detection report includes:

[0031] Based on the fault type judgment result and the fault location coordinates, the contribution of each mode to fault detection and the complementarity between modes are quantified by calculating the information entropy, mutual information and conditional mutual information of each modality data, and a modality contribution score is obtained.

[0032] Based on the modal contribution score, an adaptive weight allocation strategy is designed in combination with cable type, environmental conditions and fault characteristics. The fusion weights of each modal data of time domain reflection, frequency domain analysis, impedance measurement and insulation resistance are dynamically adjusted to obtain an optimized weight allocation scheme.

[0033] Based on the optimized weight allocation scheme, a lightweight neural network model suitable for small handheld devices is constructed through network pruning, knowledge distillation, and quantization compression techniques. Combined with augmented reality technology, fault visualization and intelligent report generation are achieved to obtain the final cable fault detection report.

[0034] Preferably, in the decentralized self-organizing structure of the rotating cube set, the rotatability of the cube units enables them to undergo rotational transformations in a high-dimensional feature space to adapt to feature observations from different angles. The orientation sensitivity of the cube units enables the network to capture the directional features of cable faults, and the spatial positioning enables the cube units to establish geometric relationships between cable multimodal data.

[0035] Preferably, in the spike neural network gating mechanism, the update gate controls the degree of update of the memory state by the new information at the current moment through the sigmoid function, the reset gate controls the degree of influence of historical information on the current state, and the memory gate manages long-term dependencies through the tanh function. The update gate, reset gate and memory gate work together to achieve adaptive learning of the timing characteristics of cable faults.

[0036] Preferably, the dynamic probabilistic neighborhood growth and edge-preserving maximum clustering algorithm achieves accurate location of cable faults and complete boundary identification through dynamic seed point selection based on fault probability distribution, probabilistic neighborhood definition considering cable physical characteristics, neighborhood growth process that maintains clear fault boundaries, and edge-preserving constraints that prevent clustering regions from crossing actual boundaries.

[0037] The present invention also provides a small handheld cable tester, comprising:

[0038] The data acquisition module is used to acquire raw cable data in multiple modes, perform multi-dimensional data acquisition on the cable, and obtain raw multi-mode cable data streams including reflection waveforms, frequency response characteristics, impedance change curves, and insulation resistance values.

[0039] The data preprocessing module is used to perform noise reduction and standardization operations on the original multimode cable data stream. Through wavelet transform noise reduction, Z-score standardization and time series dynamic normalization, a preprocessed multimode cable test dataset is obtained.

[0040] The feature extraction module is used to construct a geometric deep learning network based on the preprocessed multimodal cable test dataset using the dispersed self-organizing structure of the rotating cube set, and extract the deep features of different modal data through geometric convolution operations and self-organizing learning mechanisms to obtain a multidimensional cable feature vector set.

[0041] The feature fusion module is used to construct a spike neural network gating mechanism based on the multidimensional cable feature vector set, and fuse different modal features through a cross-modal attention mechanism to obtain a unified feature representation after fusion;

[0042] The fault identification module is used to classify faults based on the fused unified feature representation, using a differential evolution optimized backpropagation neural network model, and to locate faults by combining temporal generative adversarial network enhanced samples and dynamic probabilistic neighborhood growth clustering algorithm, so as to obtain fault type judgment results and fault location coordinates.

[0043] The report generation module is used to evaluate the contribution of each modality data based on the fault type judgment result and the fault location coordinates, through information entropy and mutual information analysis, and to optimize the model using adaptive weight allocation and lightweight neural network technology to obtain the final cable fault detection report.

[0044] Preferably, the data acquisition module includes:

[0045] The time-domain reflectometer unit is used to emit electromagnetic pulses and receive reflected signals to detect cable impedance discontinuities and fault locations.

[0046] The frequency domain analyzer unit is used to measure the transmission characteristics and frequency response of cables at different frequencies.

[0047] Impedance measurement unit, used to measure the characteristic impedance and impedance change curve of a cable;

[0048] Insulation resistance testing unit is used to measure the insulation resistance and insulation performance parameters of cables.

[0049] Preferably, it further includes:

[0050] The display unit is used to display test data, analysis results, and fault diagnosis reports in real time.

[0051] Storage unit, used to store historical test data, model parameters, and test reports;

[0052] The communication unit is used for data transmission and remote diagnostics with external devices.

[0053] The power management unit is used to provide a stable power supply for each functional module.

[0054] Compared with the prior art, the present invention has the following beneficial effects:

[0055] 1. By acquiring and preprocessing multimodal data and comprehensively utilizing the advantages of different detection methods, the comprehensiveness and reliability of cable fault detection have been improved;

[0056] 2. A geometric deep learning network was constructed using a distributed self-organizing structure of a set of rotating cubes, which enabled effective feature extraction from multimodal cable data and enhanced the ability to identify complex faults.

[0057] 3. Cross-modal feature fusion was achieved by utilizing the peak neural network gating mechanism, which solved the problem of heterogeneous data fusion and improved the quality of feature representation;

[0058] 4. By combining differential evolution-optimized backpropagation neural networks and dynamic probabilistic neighborhood growth clustering algorithms, high-precision fault classification and localization were achieved;

[0059] 5. The algorithm performance was optimized through adaptive weight allocation and lightweight neural network technology, enabling it to run efficiently on small handheld devices. Attached Figure Description

[0060] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0061] Figure 1 This is a flowchart of the cable testing method of the present invention;

[0062] Figure 2 This is a structural block diagram of the small handheld cable tester of the present invention. Detailed Implementation

[0063] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0064] like Figure 1 As shown, a cable testing method includes:

[0065] Step S1: Acquire raw cable data for multiple modes, perform multi-dimensional data acquisition on the cable, and obtain raw multi-mode cable data stream containing reflection waveform, frequency response characteristics, impedance change curve and insulation resistance value;

[0066] Step S2: Perform noise reduction and standardization operations on the original multimode cable data stream. Through wavelet transform noise reduction, Z-score standardization and time series dynamic normalization, a preprocessed multimode cable test dataset is obtained.

[0067] Step S3: Based on the preprocessed multimodal cable test dataset, construct a geometric deep learning network using the dispersed self-organizing structure of the rotating cube set, and extract the deep features of different modal data through geometric convolution operations and self-organizing learning mechanisms to obtain a multidimensional cable feature vector set;

[0068] Step S4: Based on the multidimensional cable feature vector set, construct a spike neural network gating mechanism, and fuse different modal features through a cross-modal attention mechanism to obtain a unified feature representation after fusion;

[0069] Step S5: Based on the fused unified feature representation, fault classification is performed using a backpropagation neural network model optimized by differential evolution, and fault localization is performed by combining temporal generative adversarial network to enhance samples and dynamic probabilistic neighborhood growth clustering algorithm to obtain fault type judgment results and fault location coordinates.

[0070] Step S6: Based on the fault type judgment result and the fault location coordinates, evaluate the contribution of each modal data through information entropy and mutual information analysis, optimize the model using adaptive weight allocation and lightweight neural network technology, and obtain the final cable fault detection report.

[0071] In this embodiment, raw cable data in multiple modes is first acquired. Multi-dimensional data acquisition is performed on the cable to obtain a raw multi-mode cable data stream containing reflected waveforms, frequency response characteristics, impedance variation curves, and insulation resistance values. In this step, multi-mode data acquisition refers to simultaneously acquiring multiple physical characteristic data of the cable using different measurement methods and sensors. The reflected waveform is obtained using the time-domain reflectometry (TDR) method, which detects impedance discontinuities in the cable by sending electromagnetic pulses to the cable and analyzing the returned reflected signals. Frequency response characteristics are obtained through frequency domain analysis, measuring the cable's transmission characteristics at different frequencies and reflecting the cable's frequency-dependent losses and phase changes. The impedance variation curve is an impedance distribution map along the cable length, reflecting the cable's physical and electrical continuity. Insulation resistance values ​​are an indicator of the cable insulation layer's ability to impede current, typically measured using a high-precision resistance meter. These multi-dimensional data together constitute the basic data stream for a comprehensive assessment of the cable's condition.

[0072] Next, the original multimodal cable data stream undergoes noise reduction and standardization. Wavelet transform noise reduction, Z-score normalization, and time series dynamic warping are used to obtain a preprocessed multimodal cable test dataset. Wavelet transform noise reduction is a time-frequency domain analysis method that decomposes the signal into wavelet coefficients of different frequencies and time scales. Thresholding removes wavelet coefficients representing noise, and the signal is then reconstructed, effectively preserving signal edges and abrupt changes, which is crucial for detecting cable fault points. Z-score normalization is a commonly used data normalization method. By calculating each data point, subtracting the mean, and dividing by the standard deviation, data with different dimensions and ranges are transformed to the same statistical distribution, allowing for effective comparison and fusion of data from different modalities. Time series dynamic warping is a key technology for solving the problem of time asynchrony in multimodal data. It uses dynamic programming algorithms to find the optimal correspondence between different time series, ensuring that the data from each modality are aligned in the time dimension, laying the foundation for subsequent feature fusion.

[0073] Then, based on the preprocessed multimodal cable test dataset, a geometric deep learning network is constructed using the distributed self-organizing structure of a rotating cube ensemble. Deep features of different modalities are extracted through geometric convolution operations and a self-organizing learning mechanism to obtain a multidimensional cable feature vector set. The rotating cube ensemble is a novel geometric deep learning structure composed of cube units with rotatability, orientation sensitivity, and spatial localization. Rotatability allows the cube units to rotate and transform in a high-dimensional feature space, enabling observation and understanding of data features from multiple angles. Orientation sensitivity allows the network to capture the directional features of cable faults, such as attenuation or reflection characteristics in the signal propagation direction. Spatial localization allows the cube units to establish geometric relationships between multimodal cable data, mapping data with different physical meanings to a unified feature space. Geometric convolution operations are an extension of traditional convolution, considering the relative position and orientation of features in geometric space, enabling more effective extraction of structured features. The self-organizing learning mechanism includes three key components: competitive learning, lateral inhibition, and topology-preserving mapping. Through this mechanism, the network can dynamically adjust the distribution and connection relationships of the cube units, adaptively forming a feature extractor sensitive to specific cable fault modes.

[0074] Next, based on the multidimensional cable feature vector set, a spike neural network gating mechanism is constructed. This mechanism fuses features from different modalities through a cross-modal attention mechanism to obtain a unified feature representation. A spike neural network is a computational model inspired by biological neural networks, simulating the mechanism by which neurons transmit information through pulse sequences. Neurons in this network only generate output pulses when they receive sufficient input. This network has unique advantages in processing time-sensitive features, enabling it to accurately capture the instantaneous characteristics of cable faults. The gating mechanism comprises three key components: an update gate, a reset gate, and a memory gate. The update gate uses a sigmoid function to control the degree to which new information updates the memory state; the reset gate controls the degree to which historical information influences the current state; and the memory gate uses a tanh function to manage long-term dependencies. These three gating units work together to achieve dynamic information filtering, making it particularly suitable for handling time-series problems such as cable faults. The cross-modal attention mechanism generates attention weights by calculating the correlation matrix between features from different modalities, guiding the feature selection and fusion process. This mechanism can automatically learn "which features from which modalities are most important for the current judgment," highlighting complementary information and suppressing redundant features.

[0075] Then, based on the fused unified feature representation, a differential evolution-optimized backpropagation neural network model is used for fault classification. This is combined with temporal generative adversarial networks (GANs) to enhance samples and a dynamic probabilistic neighborhood growing clustering algorithm for fault localization, yielding fault type judgment results and fault location coordinates. The differential evolution-optimized backpropagation neural network (DE-BP) is a hybrid model combining global optimization and local search. The differential evolution algorithm is a swarm intelligence optimization method that searches for the global optimum in the solution space by simulating mutation, crossover, and selection operations in biological evolution. Using it to optimize the weights and bias parameters of the BP neural network can effectively avoid getting trapped in local optima and improve the ability to identify complex cable faults. The temporal generative adversarial network (SiTime-GAN) is a variant of the generative adversarial network designed for time-series data, consisting of a generator and a discriminator. The generator attempts to create realistic synthetic fault samples, while the discriminator attempts to distinguish between real and synthetic samples. Through adversarial training, the network can learn the temporal characteristics of cable faults, generating diverse synthetic fault samples and addressing the data scarcity problem. Dynamic Probabilistic Neighborhood Growth and Edge Preserving Maximum Clustering (DPNG-EPMC) is an advanced clustering method specifically designed for signal processing. It achieves accurate fault location and boundary identification by using dynamic seed point selection based on fault probability distribution, probabilistic neighborhood definition considering cable physical characteristics, neighborhood growth process that maintains clear fault boundaries, and edge preserving constraints that prevent clustered regions from crossing actual boundaries.

[0076] Finally, based on the fault type judgment result and the fault location coordinates, the contribution of each modal data is evaluated through information entropy and mutual information analysis. Adaptive weight allocation and lightweight neural network technology are then used for model optimization to obtain the final cable fault detection report. Information entropy is a measure of information uncertainty, calculated as the negative logarithmic expected value of the probability distribution of information, used to quantify the amount of information contained in each modal data. Mutual information is an indicator of the degree of interdependence between two random variables, which can be used to assess the correlation between modal data and fault judgment results, as well as the redundancy and complementarity between different modal data. Adaptive weight allocation is a method that dynamically determines the optimal combination of modal weights based on cable type, environmental conditions, fault characteristics, and modal contribution, enabling intelligent adjustment of the fusion strategy according to specific circumstances. Lightweight neural network technology includes three key techniques: network pruning, knowledge distillation, and quantization compression. Network pruning removes neurons and connections that contribute less to the output, reducing the network size; knowledge distillation extracts knowledge from a complex "teacher" model to train a lightweight "student" model; and quantization compression converts model parameters from floating-point numbers to low-bit integers, reducing storage and computational requirements. These technologies enable complex algorithms to run efficiently on small, resource-constrained handheld devices, achieving efficient, low-power fault detection and visualization report generation.

[0077] Through the synergistic effect of the above six key steps, this cable testing method can comprehensively collect and analyze cable data, accurately identify and locate various faults, and provide detailed test reports, offering strong technical support for cable maintenance. This method is particularly suitable for cable fault detection in complex environments, effectively identifying multiple fault types such as open circuits, short circuits, insulation aging, and poor contact, thus improving the efficiency and accuracy of cable maintenance.

[0078] Example 1:

[0079] In this embodiment, the original multimode cable data stream undergoes noise reduction and standardization operations. Through wavelet transform noise reduction, Z-score normalization, and time series dynamic normalization, a preprocessed multimode cable test dataset is obtained, including:

[0080] Based on the original multimode cable data stream, an adaptive median filtering algorithm is used to remove environmental noise and measurement interference. Wavelet transform is used to decompose the signal and reconstruct the key frequency bands to obtain the noise-reduced cable test data.

[0081] The noise-reduced cable test data is processed using Z-score standardization to address the dimensional differences between different modal data, and the time series dynamic warping algorithm is used to solve the problem of asynchronous sampling of multimodal data, resulting in a standardized aligned dataset.

[0082] Based on the standardized aligned dataset, fault features are enhanced and extracted using autoencoder and signal decomposition techniques to reconstruct lost or incomplete data information, thereby obtaining the preprocessed multimode cable test dataset.

[0083] In this embodiment, based on the original multimodal cable data stream, an adaptive median filtering algorithm is first used to remove environmental noise and measurement interference. Wavelet transform is then used to decompose the signal and reconstruct key frequency bands to obtain the denoised cable test data. The adaptive median filtering algorithm is a nonlinear filtering method that automatically adjusts the filter window size and threshold based on local statistical characteristics, effectively removing impulse noise and outliers while preserving signal edge features. The algorithm first calculates the median of pixels within the filter window, then determines whether a pixel is a noise point based on the difference between the median and the current pixel value. For points identified as noise, the median is used; for non-noise points, the original value is retained. This method can remove random noise while preserving signal details, making it particularly suitable for the preprocessing of cable test data. Wavelet transform is a time-frequency analysis method that can analyze signal features at different scales. Unlike the traditional Fourier transform, wavelet transform uses finite-length or rapidly decaying waveforms as basis functions, providing local information in both the time and frequency domains simultaneously. In cable testing, wavelet transform first decomposes the signal into wavelet coefficients across multiple frequency bands. These coefficients are then processed using soft or hard thresholding methods (removing small coefficients representing noise), and finally the signal is reconstructed to obtain the denoised result. The advantage of wavelet transform denoising lies in its ability to effectively preserve the edge and abrupt change characteristics of the signal, which is crucial for detecting cable fault points.

[0084] Next, the noise-reduced cable test data is processed using Z-score standardization to address the dimensional differences between different modal data. The dynamic time warping (DTW) algorithm is then used to solve the problem of asynchronous sampling of multimodal data, resulting in a standardized aligned dataset. Z-score standardization is a commonly used data normalization method. By subtracting the mean from each data point and dividing by the standard deviation, the data is converted into a standard normal distribution with a mean of 0 and a standard deviation of 1. The main advantage of this standardization method is that it can maintain the distribution shape of the data while eliminating the influence of different dimensions and ranges, allowing for effective comparison and fusion of data from different modalities (such as reflection waveforms, frequency response, impedance changes, and insulation resistance). Dynamic time warping (DTW) is an important technique for solving time series alignment problems. In multimodal cable testing, data from different modalities often have different sampling frequencies and time points, leading to data asynchrony. The DTW algorithm constructs a distance matrix and finds the optimal path to determine the best correspondence between different time series, achieving nonlinear time alignment. Unlike simple linear interpolation, DTW can handle scaling and offset on the time axis, more accurately aligning key events (such as fault points) in different modal data, laying the foundation for subsequent multimodal feature fusion.

[0085] Finally, based on the standardized aligned dataset, fault features are enhanced and extracted using autoencoders and signal decomposition techniques to reconstruct lost or incomplete data information, resulting in the preprocessed multimodal cable test dataset. An autoencoder is an unsupervised learning neural network consisting of an encoder and a decoder. The encoder compresses the input data into a low-dimensional latent space, while the decoder attempts to reconstruct the original input from this low-dimensional representation. By minimizing the reconstruction error, the autoencoder learns the intrinsic structure and features of the data. In cable test data processing, autoencoders are used for two key tasks: feature enhancement and data reconstruction. For feature enhancement, by designing a special loss function, the autoencoder can highlight fault-related feature patterns; for data reconstruction, the autoencoder can utilize the learned data distribution characteristics to fill in missing or incomplete measurement data. Signal decomposition techniques are methods that decompose complex signals into multiple simpler components, commonly including Empirical Mode Decomposition (EMD) and Variational Mode Decomposition (VMD). These methods can decompose cable test signals into multiple intrinsic mode functions representing different physical meanings, effectively separating fault features from background signals and improving the identifiability of fault features. By combining autoencoders and signal decomposition, fault characteristics can be enhanced, background noise can be suppressed, and potentially missing data can be supplemented, ultimately resulting in a high-quality preprocessed multimode cable test dataset.

[0086] Example 2:

[0087] In this embodiment, a geometric deep learning network is constructed using the dispersed self-organizing structure of a set of rotating cubes. Through geometric convolution operations and a self-organizing learning mechanism, deep features of different modalities are extracted to obtain a multidimensional cable feature vector set, including:

[0088] Based on the preprocessed multimodal cable test dataset, a cube unit group with rotatability, orientation sensitivity and spatial positioning is designed. Through geometric convolution operation, a structured representation of different modal data is established in a high-dimensional feature space to obtain the initial geometric feature mapping.

[0089] For the initial geometric feature mapping, a distributed self-organizing learning mechanism is applied. Through competitive learning, lateral inhibition, and topology preservation mapping, the distribution and connection relationship of the cube units are dynamically adjusted to form a self-organizing network that adapts to the characteristics of cable faults, thus obtaining a self-organizing feature representation.

[0090] Based on the self-organizing feature representation, dedicated feature extraction networks are designed for different modal data of time-domain reflection, frequency-domain response, impedance characteristics, and insulation characteristics. Combined with multi-scale feature pyramids and attention mechanisms, the multi-dimensional cable feature vector set is obtained.

[0091] In this embodiment, firstly, based on the preprocessed multimodal cable test dataset, a group of cube cells with rotatability, orientation sensitivity, and spatial positioning is designed. Through geometric convolution operations, a structured representation of different modal data is established in a high-dimensional feature space, yielding an initial geometric feature map. Rotatable cube cells are a novel type of geometric feature extractor that can undergo rotational transformations in a high-dimensional feature space. Rotatability means that these cells can automatically adjust their spatial orientation according to the characteristics of the input data, allowing them to "observe" and understand cable data characteristics from multiple angles. For example, when analyzing time-domain reflection waveforms, rotating cube cells can simultaneously focus on multiple dimensions of the waveform, such as amplitude, width, and slope, providing a more comprehensive feature representation than traditional convolution. Orientation sensitivity enables cube cells to capture the directional characteristics of cable faults, such as the direction of signal propagation and attenuation characteristics in the cable. This characteristic is particularly important for distinguishing different types of faults; for example, a broken wire fault manifests as a sudden reflection peak in the signal propagation direction, while insulation aging may exhibit attenuation characteristics in multiple directions. Spatial positioning allows cube cells to accurately locate themselves in the feature space and establish geometric relationships between different modal data. Geometric convolution is an extension of traditional convolution, considering not only the adjacency of features but also their relative orientation in cubic space. For input multimodal data, geometric convolution first maps the data to corresponding positions in cubic space using a modality-specific mapping function. Then, it defines a rotation transformation matrix in cubic space, enabling the network to learn the optimal feature observation angle. Finally, geometric convolution integrates the feature representations in cubic space to form an initial geometric feature map.

[0092] Next, a distributed self-organizing learning mechanism is applied to the initial geometric feature mapping. Through competitive learning, lateral inhibition, and topology-preserving mapping, the distribution and connectivity of the cube units are dynamically adjusted to form a self-organizing network adapted to the characteristics of cable faults, resulting in a self-organizing feature representation. Distributed self-organizing learning is a learning method inspired by biological neural networks, allowing the network structure to autonomously adjust according to the characteristics of the input data, forming a topology best suited for the current task. Competitive learning is its first key component, where cube units compete for feature representations, with the unit that responds most strongly to a specific input pattern gaining priority in updating its weights. Specifically, the similarity between the input feature vector and each cube unit is first calculated to determine the best-matching unit, and then the weights of that unit and its topological neighborhood are updated. Lateral inhibition is the second key component; the winning unit suppresses the activation of surrounding units, promoting sparsity of feature representations. This mechanism simulates the lateral inhibition phenomenon in biological neural networks, helping to enhance the network's ability to distinguish input patterns. Topology-preserving mapping is the third key component, maintaining the topological relationships of the input data space to ensure that similar fault patterns are mapped to nearby feature space locations. This mapping achieves a structured representation of cable fault modes by defining the topological neighborhood relationships between cube cells and maintaining these relationships during the learning process. Through this self-organizing learning, the cube cell network can form "expert cells" that specialize in responding to specific fault modes, enhancing its ability to distinguish between complex and mixed faults.

[0093] Finally, based on the self-organizing feature representation, dedicated feature extraction networks are designed for different modal data of time-domain reflection, frequency-domain response, impedance characteristics, and insulation characteristics. These networks, combined with multi-scale feature pyramids and attention mechanisms, yield the multi-dimensional cable feature vector set. The dedicated feature extraction networks are customized neural network structures designed for the characteristics of different modal data. For time-domain reflection waveforms, a one-dimensional convolutional neural network and peak detection algorithm are used to focus on capturing features such as the peak position, shape, amplitude, and time delay of the reflected wave. These features are crucial for fault location. For frequency-domain response, a spectrum analysis network is used, focusing on the resonance point, cutoff frequency, and attenuation characteristics of the frequency response curve. These features reflect the frequency-dependent losses and phase changes of the cable, and are particularly effective for detecting faults such as insulation degradation. For impedance characteristics, a sequence modeling network is used, focusing on the trend, abrupt change points, and periodic characteristics of impedance changes. These features reflect the physical and electrical continuity of the cable and are of significant value for identifying faults such as poor contact. For insulation resistance, a statistical feature network is employed to analyze the distribution characteristics, rate of change, and stability of resistance values. These features are direct indicators for assessing cable insulation condition and are particularly important for detecting insulation damage and aging. A multi-scale feature pyramid is a network architecture capable of simultaneously extracting local details and global structural information. It integrates features at different scales through top-down and bottom-up feature fusion paths. The top-down path constructs feature maps of different resolutions through continuous downsampling, capturing global semantic information; the bottom-up path fuses fine-grained local information through feature upsampling and lateral connections. An attention mechanism, by computing spatial and channel attention maps, highlights key feature regions and channels, suppresses redundant information, and further improves the quality of feature representation. Through the comprehensive application of these techniques, high-quality feature vector sets can be extracted from multimodal cable test data, providing a solid foundation for subsequent fault identification.

[0094] Example 3:

[0095] In this embodiment, in the distributed self-organizing structure of the rotating cube set, the rotatability of the cube units enables them to undergo rotational transformations in a high-dimensional feature space to adapt to feature observations from different angles. The orientation sensitivity of the cube units enables the network to capture the directional features of cable faults, and the spatial positioning enables the cube units to establish geometric relationships between cable multimodal data.

[0096] In this embodiment, the rotatability of the cube units is the first core characteristic of the rotating cube ensemble. This refers to the ability of these units to flexibly rotate and transform in a high-dimensional feature space to adapt to feature observation from different angles. In traditional neural networks, feature extractors typically have a fixed orientation, which limits their ability to understand input data. Rotating cube units, by introducing learnable rotation matrices, can automatically adjust their spatial orientation according to the characteristics of the input data, allowing them to "observe" and understand cable data features from multiple angles. Specifically, each cube unit is associated with a rotation matrix, which is a parameter learned by the network and optimized through backpropagation. This rotatability enables the network to adapt to various complex situations in cable fault detection. For example, when analyzing time-domain reflection waveforms, traditional convolutional neural networks can typically only extract features along the time axis, while rotating cube units can "observe" the waveform from any angle, simultaneously focusing on multi-dimensional characteristics such as amplitude, width, rising edge, and falling edge, providing a more comprehensive feature representation. This multi-angle feature observation capability allows the network to better distinguish similar but fundamentally different fault modes, such as localized insulation damage and poor contact, which may exhibit similar characteristics from a traditional single perspective.

[0097] The orientation sensitivity of cube cells is the second core characteristic of rotating cube assemblies, enabling the network to capture the directional characteristics of cable faults. In cable faults, the signal propagation direction and attenuation characteristics often contain important diagnostic information. Different types of faults exhibit different characteristic patterns in different directions, and orientation sensitivity allows the network to distinguish these directional features, improving the accuracy of fault type identification. Specifically, the cube cells contain orientation-sensitive filters that respond to feature variations in specific directions. For example, a wire break fault manifests as a sharp reflection peak in the signal propagation direction, with less variation in the dimension perpendicular to the propagation direction; while insulation aging may exhibit attenuation characteristics in multiple directions, with a relatively flat reflection peak. Orientation-sensitive cube cells can capture these directional differences, providing richer feature representations. This orientation sensitivity is particularly valuable for analyzing the fault propagation characteristics in cables. For example, by analyzing the attenuation rate of the fault signal in different directions, the severity and type of the fault can be determined, providing a basis for maintenance decisions.

[0098] Spatial localization is the third core characteristic of the rotating cube ensemble, enabling cube cells to establish geometric correlations between multimodal cable data. Although different modes of cable data (such as time-domain reflection, frequency-domain response, and impedance characteristics) exhibit different forms, they describe different aspects of the same physical entity. Spatial localization allows cube cells to establish geometric correspondences between these modal data in a high-dimensional feature space, forming a more complete fault representation. Specifically, each cube cell is assigned spatial coordinates, which are optimized during the learning process so that related features are mapped to nearby spatial locations. Through this mechanism, the correspondences between different modal data can be automatically learned; for example, associating the peak position in the time-domain reflection waveform with the frequency change point in the frequency-domain response, or linking impedance changes with a decrease in insulation resistance. This cross-modal spatial correlation is crucial for a comprehensive understanding of cable faults because data from different modes often provide complementary information. For example, time-domain reflection may provide accurate fault location, while frequency-domain analysis may provide clues to the fault type. Spatially correlating this information yields more accurate and comprehensive fault diagnosis results.

[0099] By combining these three properties, the decentralized self-organizing structure of the rotating cube set can more effectively extract and represent the complex features of cable faults, providing richer and more accurate feature information for subsequent fault identification and localization. This structure not only improves the ability to identify complex fault modes, but also enhances its adaptability to different test environments and cable types, enabling cable testing methods to cope with various complex real-world scenarios.

[0100] Example 4:

[0101] In this embodiment, the construction of the spike neural network gating mechanism, which fuses features from different modalities through a cross-modal attention mechanism to obtain a unified feature representation after fusion, includes:

[0102] Based on the multidimensional cable feature vector set, a bio-inspired spike neural network model is constructed. The feature values ​​are converted into pulse sequences using a time-coding method. By simulating the neuron firing mechanism, the ability to capture the instantaneous characteristics of cable faults is enhanced, resulting in a pulse-coded feature representation.

[0103] For the pulse code feature representation, the design of updating gate control new information on memory state, the impact of resetting gate control historical information on current state, and memory gate management long-term dependencies are implemented. By working together with update gate, reset gate, and memory gate, dynamic filtering of information is achieved, resulting in the feature representation after gating.

[0104] Based on the feature representation after gating, a cross-modal attention mechanism is implemented. By calculating the correlation matrix between features of different modalities and generating attention weights, complementary information is highlighted and redundant features are suppressed, resulting in the fused unified feature representation.

[0105] In this embodiment, a bio-inspired spike neural network model is first constructed based on the multidimensional cable feature vector set. A time-coding method is used to convert feature values ​​into pulse sequences. By simulating the neuron firing mechanism, the ability to capture the instantaneous characteristics of cable faults is enhanced, resulting in a pulse-coded feature representation. The spike neural network is a computational model inspired by biological neurons. Unlike traditional artificial neural networks, it simulates the mechanism by which biological neurons transmit information through discrete electrical impulses (action potentials). In this network, neurons only generate output pulses when they receive sufficient input stimuli to reach a threshold. This mechanism makes the network particularly sensitive to temporal characteristics. Time coding is a key concept in spike neural networks, referring to the encoding method that converts continuous feature values ​​into pulse occurrence times. In this embodiment, a hybrid strategy of rate coding and temporal coding is used. Rate coding focuses on the pulse frequency within a certain time window; the larger the feature value, the higher the firing frequency of the corresponding neuron. Temporal coding focuses on the precise occurrence time of the pulse; the larger the feature value, the earlier the corresponding neuron fires. For a feature value x, its encoding is a pulse time t = T. max * (1 - x / x max ), where T max It is the maximum time window, x max This represents the maximum value of the feature. The neuronal firing mechanism is the core of the spike neural network, involving the accumulation and reset of membrane potentials. When a neuron receives an input pulse, its membrane potential increases; when the membrane potential reaches a threshold, the neuron generates an output pulse and resets the membrane potential; in the absence of input, the membrane potential decays over time. This mechanism enables the network to accurately capture the instantaneous characteristics of cable faults, especially for abrupt events such as intermittent contact defects, offering significant advantages. Through this temporal encoding and neuronal firing mechanism, the multidimensional cable feature vector set is transformed into a pulse sequence with temporal characteristics, forming a pulse-coded feature representation.

[0106] Next, for the pulse-coded feature representation, an update gate is designed to control the update of the memory state by new information, a reset gate controls the influence of historical information on the current state, and a memory gate manages long-term dependencies. Through the collaborative work of the update gate, reset gate, and memory gate, dynamic information filtering is achieved, resulting in the gated feature representation. The gating mechanism is an information processing method inspired by the brain's working mechanism, used to control information flow and memory management. The update gate is the first key component of the gating mechanism, controlling the degree to which new information updates the memory state at the current moment through the sigmoid function. The output range of the sigmoid function is (0,1), suitable as a controller for the update ratio. The calculation of the update gate depends on the hidden state of the previous moment and the input features at the current moment. When its output value is close to 1, it indicates a high degree of update of the memory state by the current new information, suitable for detecting sudden changes in cable faults; when the output value is close to 0, it indicates the retention of historical information, suitable for maintaining a stable state. The reset gate is the second key component of the gating mechanism, controlling the degree of influence of historical information on the current state. The calculation of the reset gate also relies on the hidden state of the previous time step and the input features of the current time step. When its output value is close to 0, it indicates that historical information is "forgotten," allowing the network to focus on the current features; when the output value is close to 1, it indicates that the influence of historical information is retained. This mechanism is particularly useful for handling state transitions in cable faults. For example, when a network abruptly changes from a normal state to a fault state, the reset gate can help the network quickly adapt to the new state. The memory gate is the third key component of the gating mechanism, managing long-term dependencies through the tanh function. The output range of the tanh function is (-1, 1), suitable for representing bidirectional information flow. The memory gate is responsible for maintaining long-term memory, which is crucial for tracking the development trend of cable faults (such as the gradual process of insulation aging). These three gating units work together to achieve dynamic filtering of information, forming a gating-processed feature representation. This representation retains both important current information and maintains the necessary memory of the past, making it particularly suitable for handling problems with temporal characteristics, such as cable faults.

[0107] Finally, based on the gated feature representation, a cross-modal attention mechanism is implemented. By calculating the correlation matrix between features of different modalities and generating attention weights, complementary information is highlighted and redundant features are suppressed, resulting in the fused unified feature representation. The cross-modal attention mechanism is an efficient multimodal data fusion method that can automatically learn "which features of which modalities are most important for the current judgment." The correlation matrix is ​​the core component of this mechanism, used to quantify the interdependencies between features of different modalities. First, the gated features of each modality are projected into a common feature space, and then their dot product (or other similarity measure) is calculated to form the correlation matrix. Each element of this matrix represents the degree of correlation between a specific feature of one modality and a specific feature of another modality. The attention weights are generated based on the correlation matrix. The correlation values ​​are converted into a probability distribution using the softmax function, representing the degree of attention paid to each feature. The process of calculating the attention weights considers the importance and complementarity of features, enabling automatic focus on the feature combinations most valuable for the current judgment. For example, when a suspicious peak is detected in the time-domain reflection waveform, the attention mechanism automatically guides the model to focus on anomalies in the corresponding frequency band in the frequency domain features and corresponding changes in impedance characteristics, thereby comprehensively understanding the fault characteristics. Through this cross-modal attention mechanism, complementary information from different modalities is effectively integrated, while redundancy and noise are suppressed, forming a unified feature representation after fusion. This representation not only contains key information from each modality but also solves the problems of redundancy and inconsistency between modalities, providing a high-quality feature foundation for subsequent fault identification and localization.

[0108] Example 5:

[0109] In this embodiment, the spike neural network gating mechanism uses the sigmoid function to control the degree of update of the memory state by the new information at the current moment, the reset gate to control the degree of influence of historical information on the current state, and the memory gate to manage long-term dependencies through the tanh function. The update gate, reset gate and memory gate work together to achieve adaptive learning of the timing characteristics of cable faults.

[0110] In this embodiment, the specific implementation methods of the three gating units in the spike neural network gating mechanism and their collaborative working mechanism are further described in detail.

[0111] The update gate is the first key component in the gating mechanism of a spike neural network. Its main function is to control the degree to which new information updates the memory state. In implementation, the update gate maps the input to the (0,1) interval using the sigmoid function. This interval is suitable for representing the proportion or degree of update. The sigmoid function has a smooth S-shaped curve characteristic, capable of compressing any real value into the (0,1) interval, making it suitable as a proportional controller. The calculation formula of the update gate involves the hidden state of the previous time step and the pulse coding features of the current time step. These inputs are linearly transformed using a learnable weight matrix, and then the sigmoid function is applied to obtain the final gating value. When the output value of the update gate is close to 1, it indicates a high degree of update to the memory state from the current new information, and more attention will be paid to the current input; when the output value is close to 0, it indicates the retention of historical information, and more reliance will be placed on the previous state. This mechanism allows the network to dynamically adjust its acceptance of new information based on the characteristics of the input data, which is particularly important for cable fault detection. For example, when a potential fault abrupt change is detected (such as a significant peak in the reflected waveform), the update gate outputs a higher value to respond quickly to the change; while in a steady state, the update gate outputs a lower value to maintain the stability of the state and avoid over-responding to noise.

[0112] The reset gate is the second key component in the gated mechanism of a spike neural network. Its main function is to control the degree to which historical information influences the current state. The calculation of the reset gate is similar to that of the update gate. It uses a learnable weight matrix to linearly transform the hidden state from the previous time step and the pulse-coded features from the current time step, and then applies the sigmoid function. When the output value of the reset gate is close to 0, it indicates that historical information is "forgotten," allowing the network to focus on current features, suitable for handling new events unrelated to history. When the output value is close to 1, it indicates that the influence of historical information is retained, suitable for situations where historical context needs to be considered. This mechanism is particularly useful for handling state transitions in cable faults. For example, when a sudden change occurs from a normal state to a fault state, the reset gate helps the network quickly adapt to the new state, reducing interference from historical states. When analyzing fault development trends, the reset gate can maintain a memory of historical states, helping to understand the evolution of the fault. The synergistic work of the reset gate and the update gate enables the network to balance old and new information, allowing for rapid response to changes while maintaining necessary continuity. This is crucial for accurately identifying various types of cable faults (from transient to progressive faults).

[0113] The memory gate is the third key component in the gating mechanism of a spike neural network, and its main function is to manage long-term dependencies. Unlike update and reset gates, the memory gate maps the input to the (-1,1) interval using the tanh function. This interval is suitable for representing bidirectional information flow, including information enhancement and suppression. The tanh function has zero-centrality, making it perform well in handling both positive and negative values, and is particularly suitable for representing information enhancement (positive values) and suppression (negative values). The computation of the memory gate also involves a learnable weight matrix, but its output is not only used to control the direction of information flow, but also directly participates in the formation of the memory content. The memory gate is responsible for maintaining long-term memory, which is particularly important for tracking the development trend of cable faults. For example, when monitoring slowly developing faults such as insulation aging, the memory gate can maintain tracking of historical degradation trends, maintaining attention to long-term trends even in the face of short-term data fluctuations. Another important function of the memory gate is to manage information at different time scales, from rapidly changing instantaneous features to slowly evolving long-term trends, allowing for appropriate representation and processing.

[0114] The update gate, reset gate, and memory gate work collaboratively to achieve adaptive learning of the temporal characteristics of cable faults. When calculating the final hidden state, their roles are integrated: the update gate controls the acceptance of new information, the reset gate controls the retention of historical information, and the memory gate manages long-term dependencies and information flow. This collaborative mechanism enables the network to adaptively handle various complex cable fault conditions. For example, for sudden faults such as open circuits or short circuits, the update gate assigns high weight when the fault occurs, allowing the network to respond quickly; for intermittent contact failures, the reset gate helps the network quickly switch states when the fault appears and disappears; for slowly developing faults such as insulation aging, the memory gate keeps track of historical degradation trends; and for compound faults (such as the simultaneous presence of contact failure and insulation aging), the synergistic effect of the three gating mechanisms helps the network distinguish the impact of different fault characteristics and give appropriate attention. This adaptive learning capability significantly improves the identification and processing capabilities of various cable faults, enabling it to cope with complex fault conditions encountered in practical engineering.

[0115] Example 6:

[0116] In this embodiment, the fault classification using a backpropagation neural network model optimized by differential evolution, combined with temporal generative adversarial network-enhanced samples and dynamic probabilistic neighborhood growing clustering algorithm for fault localization, yields fault type determination results and fault location coordinates, including:

[0117] Based on the unified feature representation after fusion, a backpropagation neural network model optimized by differential evolution is constructed. The network weights and bias parameters are optimized by the global search capability of the differential evolution algorithm, thereby improving the ability to identify fault types such as open circuit, short circuit, insulation aging, and poor contact, and obtaining a preliminary fault type probability distribution.

[0118] Based on the preliminary fault type probability distribution, a temporal generative adversarial network model is used to learn the temporal features of various faults, generate diverse synthetic fault samples for data augmentation, improve the ability to identify rare faults and boundary cases, and obtain an enhanced fault feature library.

[0119] Based on the enhanced fault feature library and the preliminary fault type probability distribution, the dynamic probabilistic neighborhood growth and edge-preserving maximum clustering algorithm is applied to accurately locate the fault and identify the boundary through seed point selection, neighborhood growth and edge preservation constraints, so as to obtain the fault type judgment result and the fault location coordinates.

[0120] In this embodiment, a three-stage method is used for fault classification and localization based on the fused unified feature representation. This multi-stage method can progressively improve the accuracy and robustness of fault identification, especially for complex and rare fault types.

[0121] Specifically, firstly, based on the fused unified feature representation, a differential evolution-optimized backpropagation neural network model is constructed. The network weights and bias parameters are optimized through the global search capability of the differential evolution algorithm, improving the ability to identify fault types such as open circuits, short circuits, insulation aging, and poor contact, thus obtaining a preliminary fault type probability distribution. The differential evolution-optimized backpropagation neural network (DE-BP) is a hybrid model combining global optimization and local search. Backpropagation neural networks are a supervised learning algorithm based on gradient descent. By calculating the gradient of the loss function relative to the network parameters, weights are adjusted through backpropagation to minimize prediction errors. However, traditional BP neural networks are prone to getting trapped in local optima and are sensitive to initial parameters, potentially performing poorly when dealing with complex cable fault classification problems. The differential evolution algorithm is a swarm intelligence optimization method that searches for the global optimum in the solution space by simulating mutation, crossover, and selection operations in biological evolution. In the DE-BP model, the differential evolution algorithm first generates an initial population containing multiple candidate solutions (a set of neural network parameters), each candidate solution representing a possible set of network weights and bias parameters. Then, a mutation operation (adding the weighted difference of multiple existing solutions to the basis vectors) generates new candidate solutions. Next, a crossover operation combines the mutated candidate solutions with the original solutions to generate trial solutions. Finally, a selection operation compares the fitness (e.g., classification accuracy) of the trial solutions and the original solutions, retaining the better solution. Through this series of operations, the differential evolution algorithm can explore a broader parameter space, avoid getting trapped in local optima, and find neural network parameters more suitable for fault classification tasks. This global optimization method is particularly important for improving the ability to distinguish easily confused fault types (such as minor short circuits and severe contact defects), ultimately outputting a preliminary fault type probability distribution, representing the various types of faults that may exist in the cable and their probabilities.

[0122] Next, based on the preliminary fault type probability distribution, a temporal generative adversarial network (SiTime-GAN) model is used to learn the temporal characteristics of various faults, generating diverse synthetic fault samples for data augmentation. This improves the ability to identify rare faults and boundary cases, resulting in an enhanced fault feature library. SiTime-GAN is a variant of generative adversarial networks designed for time-series data, specifically for processing data with temporal characteristics, such as cable fault signals. A generative adversarial network consists of a generator and a discriminator, which mutually reinforce each other through adversarial training, ultimately enabling the generator to produce synthetic samples that closely approximate the real data distribution. In SiTime-GAN, the generator is a recurrent neural network structure that captures the temporal dependencies of cable fault signals, generating temporally coherent fault sequences; the discriminator learns to distinguish between real and synthetic fault samples, providing feedback to the generator. The training process of SiTime-GAN consists of two phases: pre-training and adversarial training. In the pre-training phase, an autoencoder and a recurrent network learn the basic characteristics of time-series data. In the adversarial training phase, the generator attempts to create realistic synthetic fault samples, while the discriminator tries to distinguish between real and synthetic samples. The two continuously compete and improve together. Through this training, SiTime-GAN can learn the temporal characteristics of cable faults and generate synthetic fault samples with real physical meaning. These synthetic samples are used for data augmentation, addressing the data scarcity problem in cable fault detection, especially for rare fault types and boundary cases, where it is often difficult to obtain enough real samples for model training and validation. The diverse synthetic samples generated by SiTime-GAN significantly improve the ability to identify rare faults and boundary cases, forming an enhanced fault feature library that provides richer reference information for subsequent fault localization.

[0123] Finally, based on the enhanced fault feature library and the preliminary fault type probability distribution, the Dynamic Probabilistic Neighborhood Growth and Edge-Preserving Maximum Clustering (DPNG-EPMC) algorithm is applied. Through seed point selection, neighborhood growth, and edge-preserving constraints, precise fault location and boundary identification are performed, yielding the fault type judgment result and the fault location coordinates. The Dynamic Probabilistic Neighborhood Growth and Edge-Preserving Maximum Clustering (DPNG-EPMC) algorithm is an advanced clustering method specifically designed for signal processing, suitable for tasks such as cable fault location that require precise identification of spatial location and boundaries. Seed point selection is the first step of the algorithm. Based on the fault type probability distribution and the enhanced fault feature library, high-confidence fault candidate points are identified as clustering seeds. Traditional clustering algorithms typically use fixed rules or random methods to select initial seed points, which can easily lead to unstable or inaccurate clustering results. The DPNG-EPMC algorithm, however, selects the most probable fault point as the clustering starting point based on fault probability and feature similarity, improving the accuracy of location. Neighborhood growth is the second step of the algorithm. A dynamic probabilistic neighborhood is defined for each seed point, considering the cable's physical characteristics and signal propagation patterns. Different types of cable faults exhibit different distribution characteristics in physical space. For example, open-circuit faults typically show point-like features, while insulation aging may be regional. The DPNG-EPMC algorithm achieves accurate clustering of different fault types by defining a neighborhood radius and probability threshold that adaptively adjusts based on the fault type. Starting from a seed point, the algorithm progressively expands the clustering region according to the probability threshold while maintaining the clarity of the fault boundaries. Edge-preserving constraints are the third step of the algorithm, preventing clustering regions from crossing actual boundaries. In fault localization, accurate identification of fault boundaries is crucial for repair work. The DPNG-EPMC algorithm introduces edge-preserving conditions to ensure that clustering does not cross regions where feature gradients change significantly, i.e., the actual boundaries of the fault. Finally, the maximum clustering principle is applied to merge related clustering regions to form a complete fault region representation. Through this multi-step clustering method, fault points can be accurately located and fault boundaries identified, ultimately outputting highly reliable fault type judgment results and fault location coordinates, providing accurate guidance for maintenance personnel.

[0124] Example 7:

[0125] In this embodiment, the dynamic probabilistic neighborhood growth and edge-preserving maximum clustering algorithm achieves accurate location of cable faults and complete boundary identification through dynamic seed point selection based on fault probability distribution, probabilistic neighborhood definition considering cable physical characteristics, neighborhood growth process that maintains clear fault boundaries, and edge-preserving constraints that prevent clustering regions from crossing actual boundaries.

[0126] In this embodiment, the four key components of the dynamic probabilistic neighborhood growth and edge-preserving maximum clustering algorithm and their role in cable fault location are further explained in detail.

[0127] Dynamic seed point selection based on fault probability distribution is the first key step in the Dynamic Probabilistic Neighborhood Growing and Edge-Preserving Maximum Clustering (DPNG-EPMC) algorithm, solving the problems of randomness and instability in seed point selection in traditional clustering algorithms. Seed points are the starting points of the clustering process, and their selection directly affects the results and quality of clustering. In cable fault location, ideal seed points should be located in the core of the fault area, rather than on the edge or in non-fault areas. The dynamic seed point selection method first establishes a fault probability map using the fault type probability distribution, which maps the fault probability of each point along the cable to a spatial distribution. The probability distribution originates from the output of the backpropagation neural network model optimized by differential evolution, combined with sample features enhanced by a temporal generative adversarial network. Based on this map, a peak detection algorithm is used to identify points with local probability maxima; these points are considered potential seed point candidates. Unlike traditional fixed threshold selection, dynamic seed point selection adaptively adjusts the selection criteria according to the overall characteristics of the probability distribution. For example, when the probability distribution is relatively flat (indicating a diffuse or multi-point fault area), the selection threshold is lowered and the number of seed points is increased; conversely, when the probability distribution is sharply concentrated (indicating a clear fault point), the selection threshold is raised and redundant seed points are reduced. Furthermore, spatial constraints are considered to ensure that the selected seed points maintain an appropriate distance from each other, avoiding over-concentration. This constraint is based on the physical characteristics of the cable, such as the typical spatial distribution range of different types of faults. Finally, the candidate seed points are scored and ranked, comprehensively considering fault probability values, local contrast (probability differences with surrounding areas), and feature stability (consistency across different modal data), to select the optimal set of points as the final seed points. This dynamic seed point selection method based on fault probability distribution significantly improves the quality and stability of the clustering starting point, laying a solid foundation for subsequent accurate fault localization.

[0128] The probabilistic neighborhood definition considering the physical characteristics of cables is the second key step in the DPNG-EPMC algorithm, solving the problem of single neighborhood definitions and lack of physical meaning in traditional clustering algorithms. The neighborhood definition determines the association between points during the clustering process. For objects like cables with specific physical structures, traditional Euclidean distance or fixed-radius neighborhoods are insufficient to accurately reflect the spatial propagation patterns of fault characteristics. Probabilistic neighborhood is a high-level neighborhood definition method that combines physical constraints with statistical characteristics, considering not only the spatial distance between points but also their similarity and physical connectivity in the feature space. In implementation, a physical constraint model is first established based on the cable type and structure. Different types of cables (such as oil-impregnated paper insulation and cross-linked polyethylene insulation) have different electrical characteristics and fault propagation characteristics. For example, insulation aging in oil-impregnated paper-insulated cables typically exhibits regional degradation, with fault areas distributed continuously; while water tree aging in cross-linked polyethylene cables may occur at specific points and extend radially along the insulation. These physical characteristics are transformed into constraints on the shape and extent of the neighborhood. Next, a probabilistic decay function is defined in conjunction with the fault type characteristics, as different types of faults exhibit different propagation and decay characteristics in space. Open-circuit faults are typically point-like, requiring a small and symmetrical neighborhood along the cable's cross-section. Short-circuit faults may affect adjacent conductors, necessitating consideration of electric field coupling between conductors in the neighborhood. Insulation aging, on the other hand, can extend along the cable's axial direction, requiring an elliptical neighborhood with its major axis along the cable's direction. Adaptive probability functions are used to describe these characteristics, making the neighborhood boundaries less rigid and more probabilistically variable. Finally, neighborhood parameters are dynamically adjusted, updating the neighborhood definition in real-time based on new information acquired during clustering. For example, when a fault region exhibits strong directionality, the neighborhood shape is adjusted to better accommodate this characteristic; when a steep fault boundary gradient is detected, the neighborhood range is reduced to improve boundary accuracy. This probabilistic neighborhood definition, considering the cable's physical characteristics, makes the clustering process more consistent with the actual physical laws of cable faults, improving the accuracy and physical interpretability of fault region identification.

[0129] The neighborhood growth process, which maintains clear fault boundaries, is the third key step in the DPNG-EPMC algorithm. It addresses the problems of blurred boundaries, overgrowth, or undergrowth in traditional region growing algorithms. Neighborhood growth is the core process of clustering algorithms, determining how clustered regions expand from seed points to form complete fault regions. In cable fault location, accurate boundary identification is crucial for fault repair, and traditional fixed-threshold region growing struggles to adapt to complex and variable fault characteristics. The neighborhood growth process that maintains clear boundaries first employs a multi-scale growth strategy. In the early stages of clustering, a larger growth step size is used to quickly expand to the approximate fault region. As the clustered region approaches the potential boundary (determined by feature gradients), the growth step size is gradually reduced for a refined boundary search. This strategy balances efficiency and accuracy, avoiding premature stopping or overgrowth. Next, adaptive threshold control is implemented. Traditional region growing typically uses a fixed similarity threshold to determine whether to include a point in the clustered region, while the DPNG-EPMC algorithm dynamically adjusts the threshold based on the statistical characteristics of the current clustered region. For example, when the feature distribution within a cluster is relatively uniform, the threshold is increased to strictly control the addition of new points; when the feature distribution within a region is diverse, the threshold is decreased to allow for more variation. This adaptive mechanism enables the algorithm to adapt to different types and severity of faults. Furthermore, confidence-based growth control is implemented. Unlike traditional binary decision-making (inclusion or exclusion), the DPNG-EPMC algorithm assigns a cluster confidence score to each point, representing its probability of belonging to a fault region. During growth, points with high confidence are prioritized, gradually expanding to regions with lower confidence, forming an orderly growth pattern from the inside out. This confidence-based approach makes boundary transitions smoother and more natural while preserving the richness of boundary information. Finally, multi-directional balanced growth control is employed to ensure that the clustered regions expand evenly in all directions, avoiding irregular growth caused by local feature anomalies. For example, when the growth rate in one direction is significantly faster than in others, the growth parameters in that direction are adjusted to maintain overall balance. This multi-directional balancing strategy is particularly suitable for handling irregularly shaped fault regions, such as unevenly distributed insulation aging along the cable axis. By combining these technologies, the neighborhood growth process, which maintains clear boundaries, can accurately identify the boundaries of various cable faults, providing accurate guidance for subsequent repair work.

[0130] The edge-preserving constraint, which prevents clustered regions from crossing actual boundaries, is the fourth key step in the DPNG-EPMC algorithm. It addresses the issues of boundary leakage and region merging in traditional clustering algorithms. In cable fault location, different fault regions may be close to each other yet independent; accurately distinguishing these regions is crucial for formulating repair strategies. The edge-preserving constraint is a specially designed mechanism that ensures the clustering process respects the natural boundaries in the data, preventing the erroneous merging of different fault regions or the inclusion of non-faulty regions within faulty regions. The edge-preserving constraint first implements a gradient constraint mechanism, calculating the gradient field in the feature space and identifying regions with significant gradient changes as potential boundaries. During cluster expansion, encountering high-gradient regions increases the resistance to expansion in that direction, slowing or stopping growth in that direction. For example, in time-domain reflectance data, fault points often exhibit abrupt changes in reflectance coefficients, forming high-gradient regions; in frequency-domain response data, different types of faults may create characteristic response changes in specific frequency bands, also generating gradient boundaries. By monitoring these gradient changes, clustered regions are prevented from crossing natural boundaries. Next, a topology constraint mechanism is implemented to maintain the topological structure of the clustered regions, preventing unreasonable shapes. For example, in the case of insulation aging faults in cables, the affected areas are typically continuously distributed and should not exhibit shapes that violate physical laws, such as "islands" or "bridging." By examining the connectivity and shape characteristics of clustered regions, potential topological errors are identified and corrected to ensure that the clustering results conform to the physical characteristics of cable faults. Furthermore, a competition suppression mechanism is implemented, where multiple clustered regions compete with each other when they grow simultaneously. Based on the characteristic consistency, growth potential, and boundary gradient of each region, their growth parameters are dynamically adjusted, allowing regions with more defined features and clearer boundaries to gain a competitive advantage. This competition mechanism effectively prevents erroneous merging at weak boundaries and is particularly suitable for handling multi-point fault situations. Finally, a post-processing optimization mechanism is implemented to refine the results after the initial clustering. The characteristics of the cluster boundaries, such as boundary length, curvature, and gradient strength, are analyzed to identify and correct potential boundary errors. For example, for abnormally prominent boundary portions, their characteristic consistency is checked to determine whether to retain (potentially a genuine fault feature) or prune (potentially an erroneous growth); for boundary gaps, their rationality is assessed to determine whether to fill (potentially a missed area) or retain (potentially a genuine non-faulty area). This post-processing optimization ensures the smoothness and physical plausibility of the final cluster boundaries. Through these edge-preserving constraint mechanisms, the DPNG-EPMC algorithm can accurately identify the actual boundaries of cable faults, maintaining high-precision fault location and boundary identification capabilities even in the case of multi-point faults or complex background interference, providing precise location guidance for cable repair.

[0131] By organically combining four core steps—dynamic seed point selection based on fault probability distribution, probabilistic neighborhood definition considering cable physical characteristics, a neighborhood growth process that maintains clear fault boundaries, and edge-preserving constraints to prevent clustered regions from crossing actual boundaries—the Dynamic Probabilistic Neighborhood Growth and Edge-Preserving Maximum Clustering algorithm achieves precise fault location and complete boundary identification for cables. Compared with traditional clustering methods, this algorithm exhibits significant advantages in cable fault location: First, it fully considers the physical characteristics and fault mechanisms of cables, making the clustering process more consistent with the formation and distribution patterns of actual faults; second, it employs a dynamic adaptive parameter adjustment strategy, adapting to different types and severity of faults; third, it pays particular attention to accurate boundary identification, ensuring the accuracy and reliability of the location results; and finally, it can effectively handle multi-point faults and complex background interference, improving its application capability in complex real-world environments. Through this high-precision fault location and boundary identification, maintenance personnel can accurately understand the location, extent, and boundary characteristics of faults, formulate more targeted repair plans, greatly improve the efficiency and quality of cable maintenance, reduce unnecessary excavation and repair costs, and extend the service life of cables.

[0132] Example 8:

[0133] In this embodiment, the step of evaluating the contribution of each modality's data through information entropy and mutual information analysis, optimizing the model using adaptive weight allocation and lightweight neural network technology, and obtaining the final cable fault detection report includes:

[0134] Based on the fault type judgment result and the fault location coordinates, the contribution of each mode to fault detection and the complementarity between modes are quantified by calculating the information entropy, mutual information and conditional mutual information of each modality data, and a modality contribution score is obtained.

[0135] Based on the modal contribution score, an adaptive weight allocation strategy is designed in combination with cable type, environmental conditions and fault characteristics. The fusion weights of each modal data of time domain reflection, frequency domain analysis, impedance measurement and insulation resistance are dynamically adjusted to obtain an optimized weight allocation scheme.

[0136] Based on the optimized weight allocation scheme, a lightweight neural network model suitable for small handheld devices is constructed through network pruning, knowledge distillation, and quantization compression techniques. Combined with augmented reality technology, fault visualization and intelligent report generation are achieved to obtain the final cable fault detection report.

[0137] In this embodiment, based on the fault type determination result and fault location coordinates, the final cable fault detection is constructed using information theory and model optimization methods. This approach considers both the contribution of each modal data and the resource limitations of small handheld devices, achieving efficient and reliable fault detection.

[0138] Specifically, based on the fault type judgment result and the fault location coordinates, the contribution of each mode to fault detection and the complementarity between modes are quantified by calculating the information entropy, mutual information, and conditional mutual information of each modal data, resulting in a modal contribution score. This step is the foundation for optimizing the modal fusion strategy, using information theory to objectively evaluate each modal data. Information entropy is a fundamental indicator for quantifying information uncertainty, first proposed when information theory was established. In cable fault detection, information entropy is used to measure the average amount of information contained in each modal data. The calculation process first normalizes and discretizes each modal data, mapping continuous data to a finite state space, then estimates the probability distribution of each state, and finally calculates the information entropy value according to the Shannon entropy formula. A higher entropy value indicates a more uniform data distribution, containing more uncertainty and potentially richer information; a lower entropy value indicates a more concentrated data distribution and lower uncertainty. For example, in a normal cable, the time-domain reflection waveform may be relatively smooth with a lower entropy value; while in a faulty cable, the reflection waveform may contain multiple abrupt changes with a higher entropy value. By comparing the entropy changes of various modal data under fault conditions, their information value is initially assessed. Mutual information is an indicator that measures the degree of interdependence between two random variables and has a dual application in modal contribution assessment: first, to assess the correlation between each modal data and the fault type and location; and second, to assess the redundancy between different modal data. First, the mutual information between each modal data and the fault type judgment result is calculated to quantify the contribution of that modality to fault classification. Then, the mutual information between each modal data and the fault location coordinates is calculated to quantify the contribution of that modality to fault location. Finally, the mutual information matrix between different modal data is calculated to assess the degree of information overlap between them. High mutual information values ​​indicate strong correlation, potentially indicating high contribution (when related to fault judgment) or high redundancy (when related to other modalities). Conditional mutual information further considers the additional contribution of a specific modality when other modal data are known. It measures the amount of additional information X provides to Y given variable Z. In cable fault detection, conditional mutual information is used to assess the marginal benefit of adding other modal data to existing modal data. The conditional mutual information of each modality with the fault diagnosis result is calculated under the condition that all other modalities are known, quantifying its unique contribution. This analysis is particularly important because it reveals the complementarity between modalities, guiding the priority selection of highly complementary modality combinations when resources are limited. Based on the comprehensive analysis of information entropy, mutual information, and conditional mutual information, a contribution score is generated for each modality's data, including multiple dimensions such as overall contribution, classification contribution, localization contribution, and complementarity score, providing an objective basis for subsequent weight optimization.

[0139] Based on the modal contribution scores, an adaptive weight allocation strategy is designed by considering cable type, environmental conditions, and fault characteristics. This dynamically adjusts the fusion weights of each modal data point (time-domain reflection, frequency-domain analysis, impedance measurement, and insulation resistance) to obtain an optimized weight allocation scheme. This step overcomes the limitations of traditional fixed-weight fusion, achieving the goal of intelligently adjusting the fusion strategy according to specific scenarios. The adaptive weight allocation strategy first establishes a comprehensive decision-making model, using modal contribution scores, cable type, environmental conditions, and fault characteristics as input variables, and outputting the optimal combination of modal weights. Cable type is a key factor affecting the detection strategy; different types of cables have different electrical characteristics and fault modes. For example, oil-paper insulated cables are susceptible to moisture, making insulation resistance measurement more effective for fault detection; while water tree aging in cross-linked polyethylene cables is more pronounced in frequency-domain analysis. A cable type knowledge base is maintained, recording applicable detection methods and typical fault characteristics for each type of cable, serving as a reference for weight adjustment. Environmental conditions include external factors such as temperature, humidity, and electromagnetic interference, which affect the quality and reliability of measurement data. For example, in environments with strong electromagnetic interference, time-domain reflectometry may be significantly affected, and its weight should be reduced accordingly; high humidity environments may affect the accuracy of insulation resistance measurements, and its weight should also be adjusted. The test environment is monitored in real time using integrated environmental sensors, or environmental information is obtained from user input to dynamically assess the reliability of each modal data and adjust the weights accordingly. Fault characteristics refer to features such as fault type, severity, and development stage, which determine which modal data are more diagnostically valuable. For example, time-domain reflectometry is usually most effective for open-circuit faults; for insulation aging, frequency domain analysis and insulation resistance may provide more information. Based on the preliminary fault judgment results, possible fault characteristics are identified, and the best detection methods for similar faults in historical case databases are referenced to adjust the weights of each modality accordingly. Based on the above multi-dimensional information, multi-objective optimization algorithms (such as genetic algorithms, particle swarm optimization, or Bayesian optimization) are applied to solve for the optimal weight combination. Optimization objectives include maximizing fault detection accuracy, minimizing resource consumption (such as computational complexity), and maximizing information utilization efficiency. The optimization process considers the complementarity and redundancy between modes to ensure that the weight allocation can fully utilize the advantages of each modal data while avoiding redundant calculations. The final optimized weight allocation scheme not only adapts to the current specific detection scenario, but also updates dynamically as the detection process progresses and new information is acquired, achieving true adaptive fusion.

[0140] Based on the optimized weight allocation scheme, a lightweight neural network model suitable for small handheld devices is constructed using network pruning, knowledge distillation, and quantization compression techniques. Combined with augmented reality technology, this enables fault visualization and intelligent report generation, resulting in the final cable fault detection report. This step aims to transform complex fault detection algorithms into a practical, lightweight solution for the field, while providing intuitive visualization and professional report output. Network pruning is the first key technology in lightweight neural networks, reducing network size by identifying and removing neurons and connections with minimal impact on the output. First, based on the optimized weight allocation scheme, different pruning strategies are determined for each modality of data processing—modalities with high contribution retain more parameters, while those with low contribution undergo more aggressive pruning. The pruning process employs sensitivity analysis, identifying parameters with low importance by calculating the impact of each parameter on the loss function. For important modalities such as time-domain reflection, a more conservative pruning strategy may be adopted, removing only 10-20% of unimportant parameters; while for modalities with low contribution in the current scenario, a more aggressive strategy may be adopted, removing more than 50% of parameters. After pruning, fine-tuning training is performed to recover some performance loss, ensuring that the simplified network still maintains high accuracy. Knowledge distillation is the second key technique for lightweighting, achieving model compression by transferring knowledge from a complex "teacher" model to a simple "student" model. In cable fault detection, a complete multimodal fusion model acts as the teacher, while a smaller model designed for resource-constrained environments acts as the student. During distillation, the student model not only learns to match true labels but also learns to mimic the output distribution of the teacher model. This method is particularly suitable for handling situations with high uncertainty, such as minor faults or mixed faults, because soft labels contain similarity information between categories, which helps improve the model's performance in boundary cases. Based on modality contribution scores, targeted distillation strategies are designed to give more attention to high-contribution modalities, ensuring that the most critical fault features are preserved during the distillation process. Quantization compression is the third key technique for lightweighting, converting model parameters from high-precision floating-point numbers to low-bit integers, reducing storage and computational requirements. A mixed-precision quantization strategy is adopted, maintaining high precision (e.g., 8 bits) for parameters on critical paths of fault detection and lower precision (e.g., 4 bits or 2 bits) for secondary paths. Quantization strategies are closely integrated with modal contribution levels. The processing network maintains high accuracy for important modes, while less important modes employ more aggressive quantization, balancing accuracy and efficiency. Augmented reality (AR) technology overlays virtual information onto the real world, providing a revolutionary visualization method for cable fault detection. It captures on-site cable images using a mobile phone or dedicated AR glasses camera, then overlays fault-related information such as fault location, type, and severity. Visualization methods include color coding (e.g., red for severe fault areas, yellow for potential risk areas), 3D markings (indicating precise fault points), and dynamic flow graphs (showing current flow or signal propagation paths).AR visualization not only intuitively displays fault conditions but also provides context-sensitive repair guidance, such as marking entry points, displaying operation steps, or suggesting necessary tools. Intelligent report generation is the final output, transforming complex technical analysis into a structured, professional report. The report includes a fault summary (type, location, severity), detailed analysis (key findings from each modality), fault diagnosis (causal analysis and development trends), repair recommendations (methods, materials, and time estimates), and preventative measures (suggestions to avoid similar faults). The report's technical depth and detail are automatically adjusted based on the user's role (e.g., field technician, maintenance engineer, or management decision-maker) to ensure information applicability. Furthermore, report generation incorporates knowledge accumulation, providing experience-based diagnostic and repair recommendations by comparing current faults with historical cases, enabling continuous learning and evolution. Through the integrated application of these lightweight and visualization technologies, complex cable fault detection algorithms can operate efficiently in resource-constrained field environments, providing maintenance personnel with intuitive and accurate fault diagnosis and repair guidance, ultimately improving the efficiency and quality of cable maintenance.

[0141] Modal contribution score is a quantitative indicator used to comprehensively evaluate the value and role of various modal data in cable fault detection. It includes scores across multiple dimensions, reflecting the different characteristics and functions of the modal data. The overall contribution score is a comprehensive evaluation considering multiple aspects such as fault classification, location, and feature extraction, reflecting the importance of the modality in the entire detection process. The calculation method involves weighting the scores of each sub-item according to specific weights, or using dimensionality reduction techniques such as principal component analysis to extract key features from the multi-dimensional evaluation. The classification contribution score focuses on evaluating the degree to which modal data helps in fault type identification, primarily based on the mutual information between modal data and fault type, and the improvement in accuracy in detecting specific types of faults. For example, time-domain reflectometry data may achieve a high classification contribution score in open-circuit fault detection, but a lower score in insulation aging detection. The location contribution score evaluates the degree to which modal data helps in determining fault location, primarily based on the mutual information between modal data and fault location, and the improvement in location prediction accuracy. For example, time-domain reflectometry data typically has a high location contribution score because the time delay of the reflected wave is directly related to the fault distance. The complementarity score assesses the degree of information complementarity between this modality and other modalities, primarily based on conditional mutual information analysis. High complementarity means that this modality provides unique information not included in other modalities, and even if its individual contribution is not high, it is still of significant value in multimodal fusion. The stability score assesses the performance stability of modal data under different environmental conditions and fault scenarios, based on variance analysis or sensitivity analysis of historical test data. Highly stable modalities provide reliable information under various conditions and are suitable as foundational modalities. The resource efficiency score assesses the time, equipment, and computing resources required to acquire and process the modal data, reflecting the cost-effectiveness ratio. Some modalities may have high contribution but also high resource requirements, requiring a balance between benefits and costs. The interpretability score assesses the degree to which the information provided by the modal data supports the explanation of fault mechanisms, based on consistency analysis of expert knowledge and physical models. Highly interpretable modalities can not only identify faults but also explain their causes, which is particularly valuable for guiding repairs. Based on the above multi-dimensional scores, a multi-objective decision-making method (such as TOPSIS or AHP) is used to generate final modality priority and weight recommendations, providing a scientific basis for adaptive weight allocation. The innovation of modal contribution scoring lies in the fact that it not only considers the direct contribution of modes to fault detection, but also the overall performance under the synergistic effect of multiple modes, making resource allocation more reasonable and the detection effect more optimized.

[0142] Adaptive weight allocation strategy is the core mechanism for achieving intelligent modal fusion in cable fault detection. It dynamically adjusts the importance weights of each modality data according to the specific detection scenario to optimize performance. Unlike traditional fixed-weight methods, adaptive strategies can flexibly adjust according to changes in the situation, achieving "tailor-made" detection optimization. The Analytic Hierarchy Process (AHP) is the fundamental method in adaptive weight allocation, decomposing the modality selection problem into three levels: objective, criteria, and scheme. In cable fault detection, the objective is the most accurate fault diagnosis; criteria include modal contribution scores, cable type matching degree, environmental adaptability, and fault characteristic matching degree; schemes are various possible combinations of modal weights. A judgment matrix is ​​constructed, consistency checks are performed, and finally, the weights of each criterion and the scores of each modality under each criterion are calculated to generate a preliminary weight allocation. Fuzzy logic reasoning handles the uncertainty in the detection environment. Through three steps—fuzzification, reasoning, and defuzzification—expert experience is transformed into computable weight adjustment rules. For example, a fuzzy rule such as "if electromagnetic interference is strong and the contribution of time-domain reflection is moderate, then reduce the weight of time-domain reflection" can handle various complex situations in reality. Fuzzy logic is particularly well-suited for handling environmental conditions and fault characteristics described by language, compensating for the shortcomings of the analytic hierarchy process (AHP) in handling uncertainty. Reinforcement learning frameworks continuously optimize decision-making strategies through interaction with the environment, representing a high-level implementation of adaptive weight allocation. Weight allocation is viewed as a multi-armed gambling problem, where different weight combinations are different "arms," ​​and detection accuracy is the "reward." By balancing exploration and exploitation, the optimal weight allocation strategy is gradually learned, making it particularly suitable for handling dynamically changing detection environments. Reinforcement learning can continuously learn from historical detection experience, and its performance improves over time. Factor graph inference models represent various influencing factors (such as modal contribution, cable type, environmental conditions, etc.) and their interrelationships as factor graphs, calculating the optimal weights through a message passing algorithm. Factor graphs can simultaneously consider the complex interactions between multiple factors; for example, certain environmental conditions may particularly affect the data quality of specific modalities for a particular cable type. This model is particularly suitable for handling complex scenarios with multi-factor interactions, capturing more granular dependencies. Dynamic optimization protocols integrate the above methods and add a real-time feedback mechanism to form a complete adaptive weight allocation strategy. First, initial weights are generated using the Analytic Hierarchy Process (AHP) and fuzzy logic. Then, during the detection process, the weights are dynamically adjusted based on the confidence and consistency of intermediate results. Finally, the experience gained from this detection is incorporated into reinforcement learning to optimize the long-term strategy. This multi-level adaptive mechanism can optimize weight allocation at different time scales, from immediate response to long-term learning, comprehensively improving adaptability. Through this advanced adaptive weight allocation strategy, cable fault detection can intelligently optimize resource allocation according to specific detection scenarios, maximizing detection effectiveness while minimizing resource consumption, achieving truly intelligent detection.

[0143] Example 9:

[0144] like Figure 2 As shown, a small handheld cable tester includes:

[0145] The data acquisition module 10 is used to acquire raw cable data in multiple modes, perform multi-dimensional data acquisition on the cable, and obtain raw multi-mode cable data stream including reflection waveform, frequency response characteristics, impedance change curve and insulation resistance value.

[0146] The data preprocessing module 20 is used to perform noise reduction and standardization operations on the original multimode cable data stream. Through wavelet transform noise reduction, Z-score standardization and time series dynamic normalization, a preprocessed multimode cable test dataset is obtained.

[0147] Feature extraction module 30 is used to construct a geometric deep learning network based on the preprocessed multimodal cable test dataset using the dispersed self-organizing structure of the rotating cube set, and extract the deep features of different modal data through geometric convolution operations and self-organizing learning mechanisms to obtain a multidimensional cable feature vector set;

[0148] The feature fusion module 40 is used to construct a spike neural network gating mechanism based on the multidimensional cable feature vector set, and fuse different modal features through a cross-modal attention mechanism to obtain a unified feature representation after fusion.

[0149] The fault identification module 50 is used to classify faults based on the fused unified feature representation, using a differential evolution optimized backpropagation neural network model, and to locate faults by combining temporal generative adversarial network enhanced samples and dynamic probabilistic neighborhood growth clustering algorithm, so as to obtain fault type judgment results and fault location coordinates.

[0150] The report generation module 60 is used to evaluate the contribution of each modality data based on the fault type judgment result and the fault location coordinates, through information entropy and mutual information analysis, and to optimize the model using adaptive weight allocation and lightweight neural network technology to obtain the final cable fault detection report.

[0151] This embodiment describes the overall structure and working principle of each functional module of a small handheld cable tester. This tester integrates multimodal data acquisition, processing, and analysis functions, achieving high-precision cable fault detection and location.

[0152] The data acquisition module 10 is responsible for acquiring raw cable data in multiple modes. This module integrates various detection units such as a time-domain reflectometer, a frequency-domain analyzer, an impedance measurement unit, and an insulation resistance testing unit, enabling multi-dimensional data acquisition of the cable and outputting raw multi-mode cable data streams including reflection waveforms, frequency response characteristics, impedance change curves, and insulation resistance values.

[0153] The data preprocessing module 20 performs noise reduction and standardization on the original multimodal cable data stream. This module uses wavelet transform to remove environmental noise and measurement interference, Z-score standardization to handle the dimensional differences between different modal data, and a time series dynamic warping algorithm to solve the problem of asynchronous sampling of multimodal data. Finally, it outputs the preprocessed multimodal cable test dataset.

[0154] The feature extraction module 30, based on the preprocessed multimodal cable test dataset, constructs a geometric deep learning network using the distributed self-organizing structure of a rotating cube set. This module establishes a structured representation of different modalities in a high-dimensional feature space through geometric convolution operations, dynamically adjusts the distribution and connectivity of cube units using a distributed self-organizing learning mechanism, designs dedicated feature extraction networks for different modalities, and ultimately outputs a multidimensional cable feature vector set.

[0155] The feature fusion module 40 constructs a spike neural network gating mechanism based on a multi-dimensional cable feature vector set. This module converts feature values ​​into pulse sequences by simulating the firing mechanism of neurons, and designs update gates, reset gates, and memory gates to achieve dynamic information filtering. It fuses features from different modalities through a cross-modal attention mechanism and outputs a unified feature representation after fusion.

[0156] The fault identification module 50 performs fault classification and localization based on the fused unified feature representation. This module utilizes a differential evolution optimized backpropagation neural network model to improve the ability to identify different fault types, enhances the sample data through a temporal generative adversarial network model, and applies dynamic probabilistic neighborhood growth and edge-preserving maximum clustering algorithms for accurate fault localization. Finally, it outputs the fault type judgment result and fault location coordinates.

[0157] The report generation module 60 generates the final cable fault detection report based on the fault type judgment results and fault location coordinates. This module evaluates the contribution of each modality of data through information entropy and mutual information analysis, designs an adaptive weight allocation strategy to dynamically adjust the fusion weight of each modality of data, constructs a lightweight neural network model through network pruning, knowledge distillation and quantization compression techniques, and combines augmented reality technology to realize fault visualization and intelligent report generation.

[0158] Through the coordinated operation of these six functional modules, the small handheld cable tester can efficiently and accurately detect and locate various cable faults, providing strong technical support for cable maintenance.

[0159] Example 10:

[0160] In this embodiment, the data acquisition module 10 includes:

[0161] The time-domain reflectometer unit is used to emit electromagnetic pulses and receive reflected signals to detect cable impedance discontinuities and fault locations.

[0162] The frequency domain analyzer unit is used to measure the transmission characteristics and frequency response of cables at different frequencies.

[0163] Impedance measurement unit, used to measure the characteristic impedance and impedance change curve of a cable;

[0164] Insulation resistance testing unit is used to measure the insulation resistance and insulation performance parameters of cables.

[0165] This embodiment further elaborates on the composition of the data acquisition module and the function of each unit. These units together constitute a multimodal data acquisition system, capable of comprehensively evaluating cable conditions from different perspectives.

[0166] The time-domain reflectometry (TD-RS) unit is one of the core components of the data acquisition module, responsible for transmitting electromagnetic pulse signals and receiving reflected signals. When an electromagnetic pulse propagates in a cable, it will be reflected when it encounters a point of impedance discontinuity (such as a fault point). By analyzing the characteristics of the reflected waveform (such as amplitude, shape, and time delay), the cable impedance discontinuity and fault location can be detected. The time-domain reflectometry method performs excellently in detecting faults with significant impedance changes, such as open circuits and short circuits, and is particularly accurate in fault location.

[0167] The frequency domain analyzer unit is responsible for measuring the transmission characteristics and frequency response of the cable at different frequencies. This unit measures parameters such as attenuation and phase change by inputting signals of different frequencies into the cable, generating a frequency response curve. Frequency domain analysis provides a comprehensive assessment of the cable's condition and is particularly effective for detecting gradual faults such as insulation aging and water immersion. Different types of faults exhibit different characteristic patterns in the frequency domain; for example, insulation aging typically manifests as increased energy attenuation in the high-frequency band, while water immersion may cause a shift in the resonant point in certain frequency bands.

[0168] Impedance measurement units are used to measure the characteristic impedance and impedance variation curve of cables. The characteristic impedance of a cable is an important parameter reflecting its electrical characteristics and is affected by factors such as conductor size, insulation material, and structure. The impedance variation curve shows the change in impedance value along the cable length, helping to identify problems such as localized damage and joint quality. For example, poor contact often manifests as localized impedance fluctuations, while flattening or crushing damage can lead to a continuous shift in impedance.

[0169] The insulation resistance testing unit is responsible for measuring the insulation resistance and insulation performance parameters of cables. Insulation resistance is a direct indicator for evaluating the insulation condition of cables and is affected by factors such as insulation material aging, humidity, and temperature. This unit calculates the insulation resistance value by applying a test voltage and measuring leakage current. It can also measure insulation performance parameters such as dielectric loss factor and polarization index. Insulation resistance testing is particularly effective for detecting faults such as insulation damage, aging, and moisture.

[0170] These four test units work together to collect cable data from different perspectives, forming a multimodal data stream. Each mode provides a unique perspective on the cable condition, complementing each other to form a comprehensive cable evaluation system. By integrating this modal data, the system can more accurately identify and locate various cable faults, including open circuits, short circuits, insulation aging, and poor contact.

[0171] Example 11:

[0172] In this embodiment, the small handheld cable tester further includes:

[0173] The display unit is used to display test data, analysis results, and fault diagnosis reports in real time.

[0174] Storage unit, used to store historical test data, model parameters, and test reports;

[0175] The communication unit is used for data transmission and remote diagnostics with external devices.

[0176] The power management unit is used to provide a stable power supply for each functional module.

[0177] In this embodiment, the auxiliary functional units of the small handheld cable tester are further described in detail. These units together constitute a complete testing system, which improves the practicality of the equipment and the user experience.

[0178] The display unit is the main interface for user interaction with the equipment, used to display test data, analysis results, and fault diagnosis reports in real time. This unit typically uses a high-resolution touchscreen and supports multiple display modes, such as waveform display, spectrum graphs, and fault location maps. The display unit also integrates augmented reality functionality, which can intuitively overlay fault information onto the physical image of the cable, helping maintenance personnel quickly locate the fault point. Users can set parameters, control tests, and browse results via the touchscreen, improving operational convenience.

[0179] The storage unit is responsible for storing historical test data, model parameters, and inspection reports. This unit uses high-capacity flash memory, supporting long-term data storage and fast access. The stored historical data can be used for trend analysis, tracking changes in cable conditions, and predicting potential faults; model parameter storage ensures the equipment can start up quickly and apply the latest optimization algorithms; and inspection report storage facilitates subsequent review and comparison. The storage unit also supports data backup and recovery functions to prevent data loss.

[0180] The communication unit is used for data transmission and remote diagnostics with external devices. This unit integrates multiple communication interfaces, including USB, Wi-Fi, Bluetooth, and mobile networks, supporting connections to computers, smartphones, and cloud platforms. Through the communication unit, the tester can upload test data to a cloud server for advanced analysis, download the latest algorithm updates, or conduct real-time collaborative diagnostics with remote experts. This remote collaboration capability is particularly important in remote areas with limited technical resources, enabling on-site personnel to access professional support.

[0181] The power management unit (PMU) is responsible for providing a stable power supply to all functional modules. This unit utilizes high-energy-density lithium batteries and efficient power management chips, supporting long-term operation and fast charging. The PMU also implements multi-level power consumption control, dynamically adjusting the power supply scheme for each module based on operating conditions to maximize battery life. For example, it automatically reduces the processor frequency in standby mode and provides additional power support during high-intensity computing. Furthermore, the unit provides overvoltage, overcurrent, and overtemperature protection to ensure safe and reliable operation of the device.

[0182] These four auxiliary functional units work in conjunction with the core functional modules described earlier to form a fully functional and powerful small handheld cable tester. With the support of these units, the tester can not only efficiently and accurately detect cable faults, but also provides a good user experience and expandable functionality, meeting the cable testing needs in various complex scenarios.

[0183] This invention achieves high-precision identification and location of various fault types such as cable breakage, short circuit, insulation aging, and poor contact by using technologies such as multimodal data acquisition and fusion, geometric deep learning feature extraction, neural network gating mechanism, and adaptive weight optimization, providing strong technical support for cable maintenance in the power, communication, and construction industries.

[0184] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A cable testing method, characterized in that, include: Acquire raw cable data in multiple modes, perform multi-dimensional data acquisition on the cable, and obtain raw multi-mode cable data stream containing reflection waveform, frequency response characteristics, impedance change curve and insulation resistance value; The original multimode cable data stream is subjected to noise reduction and standardization operations. Through wavelet transform noise reduction, Z-score standardization and time series dynamic normalization, a preprocessed multimode cable test dataset is obtained. Based on the preprocessed multimodal cable test dataset, a geometric deep learning network is constructed using the dispersed self-organizing structure of a set of rotating cubes. The deep features of different modal data are extracted through geometric convolution operations and self-organizing learning mechanisms to obtain a multidimensional cable feature vector set. Based on the multidimensional cable feature vector set, a spike neural network gating mechanism is constructed, and different modal features are fused through a cross-modal attention mechanism to obtain a unified feature representation after fusion; Based on the unified feature representation after fusion, fault classification is performed using a backpropagation neural network model optimized by differential evolution, and fault localization is performed by combining temporal generative adversarial network to enhance samples and dynamic probabilistic neighborhood growth clustering algorithm to obtain fault type judgment results and fault location coordinates. Based on the fault type judgment result and the fault location coordinates, the contribution of each modality data is evaluated through information entropy and mutual information analysis. Adaptive weight allocation and lightweight neural network technology are used for model optimization to obtain the final cable fault detection report. This includes: quantifying the contribution of each modality to fault detection and the complementarity between modalities by calculating the information entropy, mutual information, and conditional mutual information of each modality data, thus obtaining a modality contribution score; designing an adaptive weight allocation strategy based on the modality contribution score, combined with cable type, environmental conditions, and fault characteristics, dynamically adjusting the fusion weights of each modality data (time domain reflection, frequency domain analysis, impedance measurement, and insulation resistance) to obtain an optimized weight allocation scheme; and constructing a lightweight neural network model suitable for small handheld devices using network pruning, knowledge distillation, and quantization compression techniques, combined with augmented reality technology to achieve fault visualization and intelligent report generation, thus obtaining the final cable fault detection report.

2. The cable testing method according to claim 1, characterized in that, The original multimode cable data stream is subjected to noise reduction and standardization operations. Through wavelet transform noise reduction, Z-score normalization, and time series dynamic normalization, a preprocessed multimode cable test dataset is obtained, including: Based on the original multimode cable data stream, an adaptive median filtering algorithm is used to remove environmental noise and measurement interference. Wavelet transform is used to decompose the signal and reconstruct key frequency bands to obtain the noise-reduced cable test data. The noise-reduced cable test data is processed using Z-score standardization to address the dimensional differences between different modal data, and the time series dynamic warping algorithm is used to solve the problem of asynchronous sampling of multimodal data, resulting in a standardized aligned dataset. Based on the standardized aligned dataset, fault features are enhanced and extracted using autoencoder and signal decomposition techniques to reconstruct lost or incomplete data information, thus obtaining the preprocessed multimode cable test dataset.

3. The cable testing method according to claim 1, characterized in that, The geometric deep learning network is constructed using a distributed self-organizing structure of a set of rotating cubes. Through geometric convolution operations and a self-organizing learning mechanism, deep features of different modalities are extracted to obtain a multidimensional cable feature vector set, including: Based on the preprocessed multimodal cable test dataset, a cube unit group with rotatability, orientation sensitivity and spatial positioning is designed. Through geometric convolution operation, a structured representation of different modal data is established in a high-dimensional feature space to obtain the initial geometric feature mapping. For the initial geometric feature mapping, a distributed self-organizing learning mechanism is applied. Through competitive learning, lateral inhibition, and topology preservation mapping, the distribution and connection relationship of the cube units are dynamically adjusted to form a self-organizing network that adapts to the characteristics of cable faults, thus obtaining a self-organizing feature representation. Based on the self-organizing feature representation, dedicated feature extraction networks are designed for different modal data of time-domain reflection, frequency-domain response, impedance characteristics, and insulation characteristics. Combined with multi-scale feature pyramids and attention mechanisms, the multi-dimensional cable feature vector set is obtained.

4. The cable testing method according to claim 1, characterized in that, The constructed spike neural network gating mechanism fuses features from different modalities through a cross-modal attention mechanism to obtain a unified feature representation after fusion, including: Based on the multidimensional cable feature vector set, a bio-inspired spike neural network model is constructed. The feature values ​​are converted into pulse sequences using a time-coding method. By simulating the neuron firing mechanism, the ability to capture the instantaneous characteristics of cable faults is enhanced, resulting in a pulse-coded feature representation. For the pulse code feature representation, the design of updating gate control new information to update the memory state, resetting gate control historical information to affect the current state, and memory gate management long-term dependencies are implemented. By working together with update gate, reset gate and memory gate, dynamic filtering of information is achieved to obtain the feature representation after gating. Based on the feature representation after gating, a cross-modal attention mechanism is implemented. By calculating the correlation matrix between features of different modalities and generating attention weights, complementary information is highlighted and redundant features are suppressed, resulting in the fused unified feature representation.

5. The cable testing method according to claim 1, characterized in that, The fault classification is performed using a backpropagation neural network model optimized by differential evolution, and the fault location is performed by combining temporal generative adversarial network to enhance samples and dynamic probabilistic neighborhood growing clustering algorithm, resulting in fault type judgment results and fault location coordinates, including: Based on the unified feature representation after fusion, a backpropagation neural network model optimized by differential evolution is constructed. The network weights and bias parameters are optimized by the global search capability of the differential evolution algorithm, thereby improving the ability to identify fault types such as open circuit, short circuit, insulation aging, and poor contact, and obtaining a preliminary fault type probability distribution. Based on the preliminary fault type probability distribution, a temporal generative adversarial network model is used to learn the temporal features of various faults, generate diverse synthetic fault samples for data augmentation, improve the ability to identify rare faults and boundary cases, and obtain an enhanced fault feature library. Based on the enhanced fault feature library and the preliminary fault type probability distribution, the dynamic probabilistic neighborhood growth and edge-preserving maximum clustering algorithm is applied to accurately locate the fault and identify the boundary through seed point selection, neighborhood growth and edge preservation constraints, so as to obtain the fault type judgment result and the fault location coordinates.

6. The cable testing method according to claim 3, characterized in that, In the decentralized self-organizing structure of the rotating cube set, the rotatability of the cube units enables them to undergo rotational transformations in a high-dimensional feature space to adapt to feature observations from different angles. The orientation sensitivity of the cube units enables the network to capture the directional features of cable faults, and the spatial positioning enables the cube units to establish geometric relationships between cable multimodal data.

7. The cable testing method according to claim 4, characterized in that, In the aforementioned spike neural network gating mechanism, the update gate controls the degree of update of the memory state by the new information at the current moment through the sigmoid function, the reset gate controls the degree of influence of historical information on the current state, and the memory gate manages long-term dependencies through the tanh function. The update gate, reset gate and memory gate work together to achieve adaptive learning of the timing characteristics of cable faults.

8. The cable testing method according to claim 5, characterized in that, The dynamic probabilistic neighborhood growth and edge-preserving maximum clustering algorithm achieves accurate location of cable faults and complete boundary identification through dynamic seed point selection based on fault probability distribution, probabilistic neighborhood definition considering cable physical characteristics, neighborhood growth process that maintains clear fault boundaries, and edge-preserving constraints that prevent clustering regions from crossing actual boundaries.

9. A small handheld cable tester, characterized in that, include: The data acquisition module is used to acquire raw cable data in multiple modes, perform multi-dimensional data acquisition on the cable, and obtain raw multi-mode cable data streams including reflection waveforms, frequency response characteristics, impedance change curves, and insulation resistance values. The data preprocessing module is used to perform noise reduction and standardization operations on the original multimode cable data stream. Through wavelet transform noise reduction, Z-score standardization and time series dynamic normalization, a preprocessed multimode cable test dataset is obtained. The feature extraction module is used to construct a geometric deep learning network based on the preprocessed multimodal cable test dataset using the dispersed self-organizing structure of the rotating cube set, and extract the deep features of different modal data through geometric convolution operations and self-organizing learning mechanisms to obtain a multidimensional cable feature vector set. The feature fusion module is used to construct a spike neural network gating mechanism based on the multidimensional cable feature vector set, and fuse different modal features through a cross-modal attention mechanism to obtain a unified feature representation after fusion; The fault identification module is used to classify faults based on the fused unified feature representation, using a differential evolution optimized backpropagation neural network model, and to locate faults by combining temporal generative adversarial network enhanced samples and dynamic probabilistic neighborhood growth clustering algorithm, so as to obtain fault type judgment results and fault location coordinates. The report generation module is used to evaluate the contribution of each modality data based on the fault type judgment result and the fault location coordinates through information entropy and mutual information analysis, and to optimize the model using adaptive weight allocation and lightweight neural network technology to obtain the final cable fault detection report. This includes: quantifying the contribution of each modality to fault detection and the complementarity between modalities by calculating the information entropy, mutual information, and conditional mutual information of each modality data, thus obtaining a modality contribution score; designing an adaptive weight allocation strategy based on the modality contribution score, combined with cable type, environmental conditions, and fault characteristics, dynamically adjusting the fusion weights of each modality data (time domain reflection, frequency domain analysis, impedance measurement, and insulation resistance) to obtain an optimized weight allocation scheme; and constructing a lightweight neural network model suitable for small handheld devices using network pruning, knowledge distillation, and quantization compression techniques, combined with augmented reality technology to achieve fault visualization and intelligent report generation, thus obtaining the final cable fault detection report.

10. The small handheld cable tester according to claim 9, characterized in that, The data acquisition module includes: The time-domain reflectometer unit is used to emit electromagnetic pulses and receive reflected signals to detect cable impedance discontinuities and fault locations. The frequency domain analyzer unit is used to measure the transmission characteristics and frequency response of cables at different frequencies. Impedance measurement unit, used to measure the characteristic impedance and impedance change curve of a cable; Insulation resistance testing unit is used to measure the insulation resistance and insulation performance parameters of cables.

11. The small handheld cable tester according to claim 9, characterized in that, Also includes: The display unit is used to display test data, analysis results, and fault diagnosis reports in real time. Storage unit, used to store historical test data, model parameters, and test reports; The communication unit is used for data transmission and remote diagnostics with external devices. The power management unit is used to provide a stable power supply for each functional module.

Images