Artificial intelligence-based cable defect detection method and system

Abstract

The application discloses a kind of cable defect detection method and system based on artificial intelligence, obtain cable detection data by from multimodal data acquisition device;Cable detection data is standardized and time dimension alignment operation is carried out, and normalized multimodal data set is obtained;Convolutional neural network is used to carry out feature extraction to multimodal data set, and preliminary feature set is generated by alternately executing three convolution operations and pooling operation;According to preliminary feature set, obtain multimodal feature vector, and generate the dynamic weight distribution of each modal feature by attention weight calculation module;Based on dynamic weight distribution, weighted fusion is carried out to multimodal feature vector, and fusion feature set is obtained;Judge whether working condition parameter feature vector is included in fusion feature set, if working condition parameter feature vector is included in fusion feature set, then fusion feature set is input long short-term memory network and time series modeling is carried out.The application effectively improves the safety and reliability of power system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of cable defect detection, and particularly discloses a cable defect detection method and system based on artificial intelligence. BACKGROUND

[0002] Cable defect detection is an important guarantee for the safe operation of power systems, which directly relates to the stability of the power grid and the reliability of power supply.

[0003] With the deepening of the reform of urban power grids and rural power grids, accurate identification and prediction of cable defects have become key technical problems that need to be solved in the power industry. Traditional detection methods mainly rely on manual inspection or single modal equipment such as infrared thermal imagers or ultrasonic detectors, but these methods often face the limitations of insufficient precision, low efficiency and poor adaptability in complex environments. Manual inspection is time-consuming and easily affected by subjective factors, and single modal equipment is difficult to capture multi-dimensional features of cable defects, resulting in high missed detection rate and false detection rate, especially in hidden defects and complex geological environments.

[0004] The core challenge in the field of cable defect detection is how to effectively integrate multi-source heterogeneous data and achieve accurate defect feature modeling.

[0005] In the cable operating environment, defect features involve multi-dimensional information such as geometric shape, material properties and working condition parameters, and a single data source is difficult to fully represent the overall picture of the defect, resulting in incomplete feature extraction.

[0006] Although the introduction of multi-modal data enriches the information dimension, the heterogeneity between data makes the fusion process complex, and traditional fusion methods are difficult to capture the relevance between deep features.

[0007] The deficiency of feature fusion further limits the dynamic prediction ability of defects, especially when the cable ages or the operating conditions change, existing models are difficult to accurately simulate the expansion trend of defects. This lack of prediction ability directly affects the real-time and reliability of the detection system.

[0008] Therefore, how to efficiently fuse multi-modal data to extract comprehensive defect features and achieve accurate prediction of defect expansion based on dynamic working conditions has become a key problem in the field of cable defect detection. SUMMARY

[0009] The present application provides a cable defect detection method and system based on artificial intelligence, which aims to solve at least one of the defects in the prior art.

[0010] One aspect of the present application relates to a cable defect detection method based on artificial intelligence, comprising the following steps:

[0011] Obtaining cable detection data from a multi-modal data acquisition device, the multi-modal data acquisition device comprising an infrared thermal imaging unit, an ultrasonic detection unit and an optical sensing unit;

[0012] Standardizing and time-dimension aligning the cable detection data to obtain a normalized multi-modal data set;

[0013] Extracting features from the multi-modal data set using a convolutional neural network, and generating a preliminary feature set by alternately performing three convolution operations and pooling operations;

[0014] Obtaining a multi-modal feature vector from the preliminary feature set, and generating a dynamic weight distribution of each modality feature through an attention weight calculation module;

[0015] Weighted fusion of the multi-modal feature vector based on the dynamic weight distribution to obtain a fusion feature set;

[0016] Determining whether the fusion feature set contains a working condition parameter feature vector, and if the fusion feature set contains a working condition parameter feature vector, inputting the fusion feature set into a long short-term memory network for time series modeling;

[0017] Outputting a defect expansion trend prediction sequence through the long short-term memory network, the defect expansion trend prediction sequence containing feature change amounts of future three detection periods;

[0018] Inputting the fusion feature set and the defect expansion trend prediction sequence into a classification module, and performing feature matching degree calculation using a preset defect threshold group;

[0019] Determining the cable defect grade according to the feature matching degree calculation result, and determining that there is a defect when the matching degree exceeds a first threshold value, and determining that the defect reaches a critical state when the matching degree exceeds a second threshold value.

[0020] Further, the step of standardizing and time-dimension aligning the cable detection data to obtain a normalized multi-modal data set comprises:

[0021] Obtaining original cable detection data from a multi-modal data acquisition device, and performing anomaly detection on cable surface temperature distribution data, cable internal defect echo signals and cable appearance image data in the original cable detection data using a data cleaning method to obtain a first data set;

[0022] According to the first data set, converting data of different modalities into a unified structure using a format conversion method, and if the timestamp deviation of the cable surface temperature distribution data and the cable internal defect echo signals exceeds a preset threshold value, aligning the timestamps by interpolation to obtain a second data set;

[0023] For the second data set, the z-score normalization method is used to standardize the cable surface temperature distribution data and the cable internal defect echo signal, and the gray scale normalization is performed on the cable appearance image data to obtain a third data set;

[0024] The third data set is fragmented by a batch processing method, and the fragmented cable surface temperature distribution data, cable internal defect echo signal and cable appearance image data are spliced in the time dimension by using a multi-modal fusion algorithm to obtain a normalized multi-modal data set.

[0025] Further, the convolutional neural network is used to extract features from the multi-modal data set, and the steps of generating a preliminary feature set by alternately performing three convolution operations and pooling operations include:

[0026] The cable surface temperature distribution data, the cable internal defect echo signal and the cable appearance image data are obtained from the multi-modal data set, and the sequence resampling method is used to convert the cable surface temperature distribution data and the cable internal defect echo signal into a two-dimensional matrix matching the resolution of the appearance image to obtain a first intermediate data set;

[0027] For the first intermediate data set, if the timestamps of the cable surface temperature distribution data or the cable internal defect echo signal deviate from the cable appearance image data by more than a preset threshold, the timestamps are aligned by the linear interpolation method, and the aligned multi-channel matrix is merged by the matrix splicing method to obtain a second intermediate data set;

[0028] According to the second intermediate data set, the first convolution operation is performed by using the convolutional neural network, the feature extraction is performed on the multi-channel matrix by using a preset convolution kernel, and the downsampling is performed by using the max-pooling operation to obtain a first feature set;

[0029] For the first feature set, the second and third convolution and max-pooling operations are alternately performed, the multi-convolution pooling results are fused by using the channel weighting method to obtain a preliminary feature set.

[0030] Further, the multi-modal feature vector is obtained according to the preliminary feature set, and the steps of generating a dynamic weight distribution of each modal feature by the attention weight calculation module include:

[0031] According to the preliminary feature set, the multi-modal feature vector is obtained, the multi-modal feature vector is split into independent modal matrices by using the matrix decomposition method, and the independent modal matrices are converted into one-dimensional feature vectors by using the vector mapping method to obtain a first feature vector set;

[0032] For the first feature vector set, the attention weight calculation module is used to weight process each modal feature vector, and the dynamic weight distribution is calculated by using the softmax function to obtain a weighted feature vector set;

[0033] If the weight value of any modality in the weighted feature vector set is lower than the preset threshold, the weight distribution is adjusted by a linear interpolation method, the adjusted weighted feature vectors are merged by a vector splicing method, and a second feature vector set is obtained;

[0034] According to the second feature vector set, a principal component analysis method is used to reduce the dimension of the high-dimensional feature vector, and an optimized multi-modal feature vector is generated by a matrix reorganization method, and a final feature vector set is obtained.

[0035] Further, the step of weighting and fusing the multi-modal feature vector based on the dynamic weight distribution to obtain the fusion feature set comprises:

[0036] An initial feature matrix is obtained from the multi-modal feature vector, the initial feature matrix is split into independent modality sub-matrices by a matrix decomposition method, and the independent modality sub-matrices are converted into one-dimensional vectors by a vector mapping method to obtain a first vector set;

[0037] For the first vector set, a dynamic weight of each modality vector is calculated using an attention mechanism, and the dynamic weight is normalized by a softmax function to obtain a dynamic weight distribution;

[0038] If the weight value of any modality in the dynamic weight distribution is lower than the preset threshold, the dynamic weight distribution is adjusted by a linear interpolation method, and the first vector set is weighted by a vector weighting method to obtain a weighted vector set;

[0039] According to the weighted vector set, the weighted vectors of each modality are merged by a vector splicing method, and the merged vectors are reduced in dimension by a principal component analysis method to obtain a fusion feature set.

[0040] Further, the step of inputting the fusion feature set into a long short-term memory network for time series modeling if the fusion feature set contains a working condition parameter feature vector comprises:

[0041] The working condition parameter feature vector containing the working condition parameter is obtained from the fusion feature set, the working condition parameter feature vector is split into a time dimension sub-vector and a parameter dimension sub-vector by a matrix decomposition method, and the time dimension sub-vector and the parameter dimension sub-vector are aligned by a linear interpolation method to obtain a first feature set;

[0042] If the standard deviation of the time dimension sub-vector in the first feature set is lower than the preset threshold, the time dimension sub-vector is smoothed by a sliding window method, and the smoothed sub-vector is normalized by a vector normalization method to obtain a second feature set;

[0043] According to the second feature set, the time dimension sub-vector is modeled by using a long short-term memory network, and the parameter dimension sub-vector is weighted by using a gated recurrent unit, to obtain a third feature set;

[0044] The third feature set is reduced in dimension by using a principal component analysis method, and the time dimension feature and the parameter dimension feature after the dimension reduction are combined by using a vector splicing method, to obtain a fourth feature set.

[0045] Further, the long short-term memory network is used to output a defect expansion trend prediction sequence, and the step of obtaining the feature change amount sequence of the future three detection cycles includes:

[0046] The original data set containing the defect feature and the working condition parameter is obtained from the historical detection data, the time dimension data is aligned by using a linear interpolation method, and the working condition parameter data is standardized by using a vector normalization method, to obtain a first data set;

[0047] It is judged whether the standard deviation of the time dimension data of the first data set is lower than a preset threshold value, if lower than the preset threshold value, the time dimension data is smoothed by using a sliding window method, and the smoothed data is reduced in dimension by using a principal component analysis method, to obtain a second data set;

[0048] According to the second data set, the working condition parameter data is weighted by using a gated recurrent unit, and the weighted working condition parameter data and the time dimension data are combined by using a feature splicing method, to obtain a third data set;

[0049] The third data set is modeled by using a long short-term memory network, and a defect expansion trend prediction sequence containing the future three detection cycles is output, to obtain a feature change amount sequence.

[0050] Further, the fusion feature set and the defect expansion trend prediction sequence are input into a classification module, and the step of calculating the feature matching degree by using a preset defect threshold group includes:

[0051] A fifth feature set containing multi-dimensional data is obtained from the fusion feature set and the defect expansion trend prediction sequence, the fifth feature set is reduced in dimension by using a principal component analysis method, and the data after the dimension reduction is standardized by using a vector normalization method, to obtain a sixth feature set;

[0052] According to the sixth feature set, each feature dimension in the sixth feature set is compared by using a preset threshold group, if the feature value of the feature dimension exceeds the upper limit of the corresponding dimension in the threshold group, the feature dimension is marked as an abnormal feature, to obtain an abnormal feature set;

[0053] The abnormal feature set is clustered by using the classification module, and the K-means algorithm is used to group the abnormal feature set, to obtain a plurality of feature subsets.

[0054] For the feature subset, the matching degree of the feature subset and the defect threshold group is calculated, and the matching degree is comprehensively evaluated by a weighted average method to obtain a classification result.

[0055] Further, the cable defect level is determined according to the feature matching degree calculation result, and when the matching degree exceeds the first threshold value, it is determined that there is a defect, and when the matching degree exceeds the second threshold value, it is determined that the defect reaches a critical state, and the steps include:

[0056] A multi-dimensional information feature vector containing multi-dimensional information is obtained from the cable state data, principal component analysis is used to reduce the dimension of the multi-dimensional information feature vector, and the reduced data is processed by a vector normalization method to obtain a standardized feature set;

[0057] According to the standardized feature set, a preset first threshold value and a second threshold value are used to compare each feature dimension, if the feature value of the feature dimension exceeds the first threshold value, it is marked as an abnormal feature, and if the feature value of the feature dimension exceeds the second threshold value, it is marked as a critical feature, to obtain an abnormal feature set and a critical feature set;

[0058] For the abnormal feature set and the critical feature set, a K-means algorithm is used to cluster and group the features to obtain a plurality of feature subsets;

[0059] The matching degree value of the feature subset and the defect level is calculated by a weighted evaluation method to determine the defect level and the critical state, and a determination result is obtained.

[0060] Another aspect of the present application relates to a cable defect detection system based on artificial intelligence for implementing the above-mentioned cable defect detection method based on artificial intelligence, the cable defect detection system based on artificial intelligence comprises:

[0061] The first acquisition module is used to acquire cable detection data from a multi-modal data acquisition device, and the multi-modal data acquisition device includes an infrared thermal imaging unit, an ultrasonic detection unit and an optical sensing unit;

[0062] The second acquisition module is used to standardize and time-dimension align the cable detection data to obtain a normalized multi-modal data set;

[0063] The first generation module is used to extract features from the multi-modal data set by using a convolutional neural network, and to generate a preliminary feature set by alternately performing three convolution operations and pooling operations;

[0064] The second generation module is used to obtain a multi-modal feature vector from the preliminary feature set, and to generate a dynamic weight distribution of each modal feature by using an attention weight calculation module;

[0065] The third acquisition module is used to perform weighted fusion on the multimodal feature vectors based on dynamic weight distribution to obtain a fused feature set;

[0066] A judgment module is used to judge whether the fused feature set contains the operating condition parameter feature vector. If the fused feature set contains the operating condition parameter feature vector, the fused feature set is input into the long short-term memory network for time series modeling;

[0067] An output module is used to output a defect expansion trend prediction sequence through a long short-term memory network. The defect expansion trend prediction sequence includes feature changes in the next three detection cycles.

[0068] The calculation module is used to input the fusion feature set and the defect expansion trend prediction sequence into the classification module and calculate the feature matching degree using the preset defect threshold group;

[0069] The judgment module is used to determine the cable defect level based on the feature matching calculation result, and judge that a defect exists when the matching degree exceeds a first threshold, and judge that the defect reaches a critical state when the matching degree exceeds a second threshold.

[0070] The beneficial effects achieved by the present invention are:

[0071] The present invention provides a cable defect detection method and system based on artificial intelligence. Cable detection data is acquired through a multimodal data acquisition device, the data is standardized and time-aligned, a convolutional neural network is used to extract features, and the multimodal features are dynamically weighted and fused through an attention mechanism. If operating parameters are included, a long short-term memory network is used for time series modeling to predict the defect expansion trend in the next three detection cycles. Finally, the fused features and the predicted sequence are input into a classification module, and the defect level is determined based on a preset threshold. The present invention realizes the accurate detection of cable defects and the prediction of future development trends, effectively improves the safety and reliability of the power system, provides a scientific basis for cable maintenance decisions, and has important engineering application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 The figure is a flow chart of an embodiment of a cable defect detection method based on artificial intelligence of the present invention. DETAILED DESCRIPTION

[0073] In order to better understand the above technical solution, the above technical solution will be described in detail below with reference to the accompanying drawings and specific implementation methods.

[0074] like Figure 1 As shown, the first embodiment of the present invention proposes a cable defect detection method based on artificial intelligence, comprising the following steps:

[0075] Step S100, acquiring cable detection data from a multi-modal data acquisition device, the multi-modal data acquisition device comprising an infrared thermal imaging unit, an ultrasonic detection unit, and an optical sensing unit.

[0076] The multi-modal data acquisition device refers to a system device that synchronously acquires different types of data (such as vision, hearing, physiological signals, etc.) by integrating multiple sensors or devices, aiming to comprehensively capture and collaboratively analyze cross-modal information.

[0077] The cable detection data is a structured or unstructured information set reflecting the physical state and performance parameters of the cable, which is collected through multi-modal technical means. The core of the cable detection data lies in quantitatively evaluating the health status and potential risks of the cable.

[0078] Step S200, standardizing the cable detection data and performing time dimension alignment operation to obtain a normalized multi-modal data set.

[0079] Standardization is a data preprocessing method that converts data to a unified dimension and distribution range through mathematical transformation. The core goal of standardization is to eliminate the bias caused by the difference in dimension or numerical range of different features, and to improve the comparability and computational efficiency of data in model training or analysis. Usually, the data distribution is adjusted to make the mean value 0 and the standard deviation 1.

[0080] Temporal alignment refers to eliminating the deviation of multi-modal data in the time axis through technical means, ensuring that the time series data obtained by different sensors or acquisition devices can accurately match the occurrence time of the same physical event. The core goal of temporal alignment is to solve the time series misalignment problem caused by sampling frequency difference, clock reference inconsistency or transmission delay, and to provide a time series consistency basis for subsequent multi-modal fusion and analysis.

[0081] Normalization is a data preprocessing technique that maps data to a specific interval (such as [0, 1] or [-1, 1]) through mathematical transformation, in order to eliminate the dimension difference between features and improve the stability and convergence efficiency of model training. The core of normalization is to adjust the numerical range of data to achieve comparability between different features, and to avoid model bias caused by dimension or scale difference.

[0082] A multimodal dataset refers to a collection containing multiple types (modalities) of data, such as text, images, audio, video, sensor signals (e.g., temperature, accelerometer), etc. The core value of a multimodal dataset lies in the fusion of information from different modalities, enhancing the model's understanding of complex scenarios and compensating for the limitations of single-modal data. The construction and processing of multimodal datasets require addressing key issues such as modality alignment, heterogeneous data fusion, and semantic consistency.

[0083] Step S300: Feature extraction using a convolutional neural network (CNN) on the multimodal dataset, generating an initial feature set by alternately performing three convolution operations and pooling operations.

[0084] A convolutional neural network (CNN) is a deep learning model specifically designed for processing grid-like data (e.g., images, audio, video). The core of a CNN automatically extracts local features and constructs hierarchical representations through convolution, pooling, and nonlinear activation functions (e.g., ReLU), significantly reducing the parameter size while preserving the spatial or temporal correlation of the data.

[0085] Convolution operation is the core operation of a convolutional neural network (CNN), which calculates the weighted sum of local regions in the input data by sliding a filter (convolution kernel) over the data, extracting local features in space or time, and preserving the internal structure of the data (e.g., image edges, textures). The core idea is to reduce the number of model parameters through local connections and parameter sharing, while capturing translation-invariant features.

[0086] Pooling operation is an important component of a convolutional neural network, which reduces the data dimension and preserves key features by downsampling the local regions of the input feature map, thereby reducing the computational load, enhancing the model's robustness to small displacements, and alleviating overfitting. The core idea is to aggregate local information and gradually abstract high-level semantic features.

[0087] Initial feature set refers to the basic feature set extracted or constructed from raw data in machine learning or deep learning tasks, used for the initial training and analysis of models. It usually contains unoptimized original features or features generated through simple rules. The core goal is to quickly build key representations of data, providing a foundation for subsequent feature selection, optimization, or high-order feature engineering.

[0088] Step S400, obtain a multimodal feature vector according to the preliminary feature set, and generate a dynamic weight distribution of each modality feature through an attention weight calculation module.

[0089] A multimodal feature vector is a unified vector representation generated by fusing features from different modalities (such as images, text, audio, sensor data, etc.), aiming to capture cross-modal complementary information and associated patterns to support joint analysis and reasoning of complex tasks. The core goal of the multimodal feature vector is to break down the modality barrier and enhance the model's understanding of complex scenarios through the synergy of multiple data sources.

[0090] The attention weight calculation module is the core component of the attention mechanism, which dynamically quantifies the relevance or dependency strength between different elements in the input sequence, and guides the model to focus on key information by generating a weight distribution. Its essence is to calculate the similarity through the interaction of query (Query) and key (Key), and normalize it into a probability distribution (Softmax), and finally aggregate the value (Value) vector to form a context-aware feature representation.

[0091] Dynamic weight distribution is a mechanism in deep learning models that adaptively adjusts the importance of features by calculating the relevance weights of different elements in the input data in real time, guiding the model to dynamically focus on key information in a specific context or task. The core of dynamic weight distribution is that the weight distribution is not fixed, but is flexibly adjusted with input content, task requirements or time sequence changes to enhance the model's expression ability and robustness.

[0092] Step S500, weight fusion of the multimodal feature vector based on the dynamic weight distribution to obtain a fusion feature set.

[0093] Weighted fusion is a core fusion method in multimodal data processing, which assigns dynamic weights to features or decision results of different modalities to achieve linear combination to generate comprehensive representation. Its core is to adjust the contribution of each modality through weights to adapt to the information complementarity and reliability difference between modalities in different scenarios, thereby improving the robustness and accuracy of the model.

[0094] The fusion feature set is a joint representation generated by integrating features from different modalities in multimodal data processing, aiming to eliminate the heterogeneity between modalities and capture complementary information to improve the model's semantic understanding and reasoning ability. The core of the fusion feature set is to transform feature vectors from visual, linguistic, sensor and other modalities into a unified high-dimensional space representation through feature layer fusion strategies (such as concatenation, weighted combination or nonlinear mapping), supporting joint modeling and decision-making for downstream tasks.

[0095] Step S600, judge whether the working condition parameter feature vector is contained in the fusion feature set, if the working condition parameter feature vector is contained in the fusion feature set, input the fusion feature set into the long short-term memory network for time sequence modeling.

[0096] The working condition parameter feature vector is a numerical representation of the key parameters of the equipment or system in a specific working state. By encoding multi-dimensional monitoring data such as temperature, pressure, and vibration frequency into a structured vector, a machine-readable working condition state expression is achieved. Its core function is to provide a unified high-dimensional input space for machine learning models, supporting device state classification, fault prediction, and performance optimization tasks.

[0097] The long short-term memory network (LSTM) is a special recurrent neural network (RNN) that solves the gradient vanishing / explosion problem of traditional RNNs when processing long sequence data by introducing a gating mechanism and cell state, enabling efficient modeling of long-term dependencies. The core design of the long short-term memory network is to dynamically regulate information flow through the forget gate, input gate, and output gate, balancing the relevance of historical memory and current input, and is widely used in time series data prediction, natural language processing, and other fields.

[0098] Time series modeling is a modeling process that targets time-ordered data sequences (such as stock prices, weather data, and sensor signals) to reveal their internal laws (trends, periodicity, randomness) through mathematical or statistical methods, and to achieve prediction, classification, or anomaly detection. Its core is to capture the time dependency of data (Time Dependency), i.e., the relevance of current observations and historical values, thereby providing a quantitative analysis framework for dynamic systems.

[0099] Step S700, output the defect expansion trend prediction sequence through the long short-term memory network, the defect expansion trend prediction sequence contains the feature change amount of the future three detection periods.

[0100] The defect expansion trend prediction sequence is a quantitative modeling of the development process of defects (such as cracks, wear, corrosion, etc.) in equipment, materials, or systems based on time series analysis methods. By constructing a time series model using historical defect data (such as size, location, frequency, etc.), the rate of future defect expansion, critical state, and failure risk are predicted. Its core goal is to capture the time dependency of defect evolution (such as accelerated expansion phase, periodic fluctuations, etc.) to provide dynamic decision-making basis for maintenance strategy optimization and life assessment.

[0101] Step S800, input the fusion feature set and the defect expansion trend prediction sequence into the classification module, and perform feature matching degree calculation using the preset defect threshold group.

[0102] ‌Classification module‌ is a functional component in machine learning or deep learning systems responsible for mapping input data to pre-defined class labels. Its core role is to achieve data grouping identification (such as text classification, image recognition, fault diagnosis, etc.) based on feature learning or rule judgment, and complete class determination through model output probability distribution or decision boundary. Classification module is usually composed of feature extraction layer, classifier and decision logic, which needs to adapt to task requirements (such as binary classification, multi-classification) and data characteristics (such as structured, time series) to design architecture.

[0103] Defect threshold group‌ is a set of pre-defined key parameters or conditions in defect management or predictive maintenance systems, used to dynamically evaluate the severity, expansion risk or trigger intervention actions (such as early warning, repair) of defects. Its core role is to divide defect state levels through quantitative standards (such as defect number threshold, growth rate threshold, time persistence threshold), providing classification response basis for decision makers.

[0104] ‌Feature matching degree calculation‌ is a key technology in image processing and pattern recognition, used to quantify the similarity or difference between two features (such as key points, regions or templates). The core goal of feature matching degree calculation is to evaluate the consistency of feature pairs through mathematical modeling or algorithms (such as distance metric, similarity function), thereby providing the basis for matching decisions (such as correct matching judgment or false matching elimination).

[0105] Step S900, determine the cable defect level according to the feature matching degree calculation result, when the matching degree exceeds the first threshold, it is determined that there is a defect, when the matching degree exceeds the second threshold, it is determined that the defect reaches the critical state.

[0106] Cable defect level is a classification management mechanism according to the threat degree of defects to the safe operation of power system and the urgency of processing, aiming to guide operation and maintenance personnel to prioritize processing of high-risk defects, and ensure the stability and reliability of cable lines.

[0107] Critical state‌ refers to the critical stage of cable defects between safe operation and potential risk transformation, characterized by defects that have not yet reached the judgment criteria of higher levels (such as critical defects), but there is a possibility of rapid deterioration or triggering a chain of failures in the short term, which requires intervention through dynamic monitoring and predictive maintenance.

[0108] Further, the cable defect detection method based on artificial intelligence provided by the embodiment comprises the following steps:

[0109] Step S110, adopt an infrared thermal imaging unit to collect cable surface temperature distribution data, an ultrasonic detection unit to collect cable internal defect echo signals, and an optical sensing unit to collect cable appearance image data.

[0110] The infrared thermal imaging unit collects cable surface temperature distribution data. The core is to detect the heat distribution on the surface of the cable through infrared radiation, and identify potential overheating points or abnormal areas. For example, poor contact of cable joints may cause local temperature rise, which can be captured by infrared thermal imaging.

[0111] In one possible implementation, a high-resolution infrared thermal imager is used, with a temperature measurement range of -20°C to 150°C and a scanning frequency of 30 frames per second. The cable surface temperature distribution is collected.

[0112] The ultrasonic detection unit collects cable internal defect echo signals. Based on the principle that ultrasonic waves will reflect when encountering defects during propagation in the medium, cracks or bubbles inside the cable can be detected.

[0113] For example, an ultrasonic probe with a frequency of 5MHz is used to scan along the cable axis, emit pulses and receive internal defect echo signals.

[0114] The amplitude and time difference of the internal defect echo signals of the cable can reflect the size and location of the defect. For example, a 2mm bubble is detected inside a certain section of the cable, and the internal defect echo signal of the cable shows a high-amplitude spike, with a time delay corresponding to a defect depth of about 5mm.

[0115] The optical sensing unit collects cable appearance image data, which is used to detect surface physical damage such as wear, cracks or foreign matter attachment. In one embodiment, a high-resolution CCD camera is used in conjunction with a ring-shaped light source to capture cable surface images with a resolution of 1920x1080 pixels.

[0116] Preferably, image acquisition should be carried out under uniform lighting to avoid glare interference.

[0117] Step S120, through signal processing, the cable surface temperature distribution data, the cable internal defect echo signal and the cable appearance image data are denoised and filtered to obtain the processed cable surface temperature distribution data, the cable internal defect echo signal and the cable appearance image data.

[0118] It should be noted that the ambient temperature and wind speed may interfere with the accuracy of the data, so it is necessary to collect in a relatively stable environment and adjust the data in combination with the environmental correction algorithm.

[0119] The processed temperature data can clearly show the temperature gradient of the cable surface, and the temperature of the abnormal point may reach 80°C, while the normal area is 40°C, which helps to quickly locate potential faults and improve the inspection efficiency.

[0120] Specifically, the signal processing needs to denoise the echo, and wavelet transform is used to filter out background noise to ensure clear defect signals. This method can accurately locate internal defects and ensure the long-term reliability of the cable operation.

[0121] In signal processing, median filtering is used to remove image noise, and then edge detection algorithm is used to highlight the crack profile. For example, a 1 cm long crack is detected on the surface of the cable, and the image shows a clear black line with a width of about 0.5 mm.

[0122] It can be understood that the processed image data can intuitively present the degree of surface damage, and the damage type can be automatically classified in combination with a deep learning algorithm, thereby improving the intelligent level of detection.

[0123] In an embodiment, comprehensive analysis of the above three kinds of data can realize overall evaluation of the cable state. For example, the infrared thermal imaging of a certain cable joint shows that the temperature anomaly is 75°C, the ultrasonic wave detection finds that there is a small bubble defect inside, and the optical image shows that the surface is slightly worn.

[0124] The comprehensive analysis shows that the joint may be caused by internal defects and surface damage, and the heating needs to be repaired in priority. This multi-dimensional detection method significantly improves the diagnostic accuracy and reduces the risk of missed detection, providing a reliable basis for cable maintenance.

[0125] Further, the cable defect detection method based on artificial intelligence provided by the embodiment comprises the following steps:

[0126] In step S210, the original cable detection data is obtained from the multi-modal data acquisition device, and the data cleaning method is used to detect the abnormality of the cable surface temperature distribution data, the cable internal defect echo signal and the cable appearance image data in the original cable detection data, to obtain a first data set.

[0127] The cable detection data obtained by the multi-modal data acquisition device includes cable surface temperature distribution data, cable internal defect echo signal and cable appearance image data, and data cleaning is a key step to ensure data quality.

[0128] For example, the cleaning method can detect abnormal values for the cable surface temperature distribution data. For example, the temperature of a certain section of cable suddenly changes to 100°C, which is far beyond the normal range of 40°C to 50°C, and may be caused by sensor failure or external interference.

[0129] The median filtering method is used to eliminate such abnormal points, and the random spikes in the cable internal defect echo signal caused by environmental noise are denoised, such as filtering out noise with an amplitude exceeding 3 times the standard deviation by using a sliding window.

[0130] The cable appearance image data needs to check the image blur or uneven illumination, for example, to eliminate dark spot images caused by light source obstruction.

[0131] The first data set after cleaning retains reliable data, laying a foundation for subsequent processing.

[0132] Step S220, according to the first data set, the data of different modalities are converted into a unified structure by using a format conversion method, if the timestamp deviation of the cable surface temperature distribution data and the cable internal defect echo signal exceeds a preset threshold, the timestamps are aligned by interpolation method to obtain a second data set.

[0133] In a possible implementation, the format conversion unifies the data of different modalities into a structured format, such as JSON format, which contains timestamps and data values.

[0134] It should be noted that the timestamp deviation may be caused by device synchronization error, for example, the timestamp of the cable surface temperature distribution data is 10:00:00, and the timestamp of the cable internal defect echo signal is 10:00:02, which exceeds the threshold of 1 second.

[0135] The linear interpolation method is used to estimate the aligned value based on the data of adjacent time points, such as the echo signal value at 10:00:00. This method ensures the time alignment of the second data set, which is convenient for multi-modal fusion.

[0136] Step S230, for the second data set, the cable surface temperature distribution data and the cable internal defect echo signal are standardized by using the z-score normalization method, and the cable appearance image data is gray normalized to obtain a third data set.

[0137] Specifically, the z-score normalization process temperature data and echo signal, which converts the data into a standard distribution with a mean of 0 and a standard deviation of 1. For example, the temperature data range is 40-80℃, which is mapped to -2-2 after normalization, which is convenient for cross-modal comparison.

[0138] The cable appearance image data is normalized by gray scale, which scales the pixel value to 0-1, such as adjusting the original pixel value 200 to 0.78. This normalized third data set eliminates the dimensional difference and improves the compatibility of the fusion algorithm.

[0139] Step S240, the third data set is divided into batches by using a batch processing method, and the divided cable surface temperature distribution data, cable internal defect echo signal and cable appearance image data are spliced in time dimension by using a multi-modal fusion algorithm to obtain a normalized multi-modal data set.

[0140] Preferably, the batch processing method divides the third data set into batches according to time period, for example, every 5 minutes of data is a batch, generating a 1000-recorded batch data set.

[0141] The multi-modal fusion algorithm splices the fragmented data in the time dimension, for example, at a certain time point, the temperature data indicates 50℃, the cable internal defect echo signal shows 2mm defect, and the image data captures 0.3mm crack, and after fusion, a complete feature vector is formed. This normalized multi-modal data set provides a unified perspective for subsequent analysis.

[0142] It can be understood that each step of the above method is interlocking, cleaning ensures data quality, format conversion and time alignment ensures data consistency, normalization eliminates dimensional differences, and fragmentation and fusion optimize data organization. This progressive processing method provides high-quality input for cable condition evaluation, significantly improving the reliability of subsequent analysis.

[0143] Further, the cable defect detection method based on artificial intelligence provided by the embodiment comprises the following steps:

[0144] In step S310, cable surface temperature distribution data, cable internal defect echo signal and cable appearance image data are obtained from the multi-modal data set, and a sequence resampling method is used to convert the cable surface temperature distribution data and the cable internal defect echo signal into a two-dimensional matrix matching the resolution of the appearance image, to obtain a first intermediate data set.

[0145] For example, when extracting the cable surface temperature distribution data, the cable internal defect echo signal and the cable appearance image data from the multi-modal data set, it is necessary to ensure that the data format is adapted to subsequent processing.

[0146] The cable surface temperature distribution data is usually recorded in the form of time series, for example, temperature values are collected every second to form a one-dimensional array, such as [45℃, 46℃, 47℃…] for a certain cable section within 10 seconds.

[0147] The cable internal defect echo signal is waveform data generated by ultrasonic detection, which contains defect position and intensity, for example, the echo amplitude at a defect depth of 2mm is 0.8V.

[0148] The cable appearance image data is a high-resolution grayscale image, such as a resolution of 512x512 pixels, which records the cable surface crack.

[0149] The sequence resampling method converts one-dimensional temperature and echo signal into a two-dimensional matrix matching the image resolution. For example, 10-second temperature data is resampled into a 512x512 matrix, each row repeats the temperature value and fills in the time axis by interpolation, ensuring alignment with the image pixels.

[0150] The echo signal is processed similarly, mapping the amplitude value to the matrix element to form the first intermediate data set.

[0151] In step S320, for the first intermediate data set, if the time stamp of the cable surface temperature distribution data or the cable internal defect echo signal deviates from the cable appearance image data by more than a preset threshold, the time stamps are aligned by linear interpolation, and the aligned multi-channel matrix is merged by a matrix splicing method to obtain a second intermediate data set.

[0152] In a possible implementation, the time stamp alignment is critical. It is assumed that the temperature data time stamp is 10:00:00, the echo signal is 10:00:01, and the image data is 10:00:02, and the deviation exceeds the threshold of 0.5 seconds.

[0153] The linear interpolation method can fill in the missing time point data, for example, estimating the echo signal value at 10:00:00 based on the amplitude of 0.8V and 0.9V before and after the interpolation as 0.85V.

[0154] The matrix splicing method merges the aligned temperature matrix, echo matrix and image matrix by channel to form a three-channel second intermediate data set, similar to the multi-channel structure of an RGB image, which is convenient for unified processing.

[0155] In step S330, according to the second intermediate data set, a convolutional neural network is used to perform a first convolution operation, a preset convolution kernel is used for feature extraction for the multi-channel matrix, and a maximum pooling operation is used for downsampling to obtain a first feature set.

[0156] Specifically, the first convolution operation of the convolutional neural network extracts features from the second intermediate data set.

[0157] The preset 3x3 convolution kernel scans the temperature, echo and image channels respectively, and captures local patterns such as temperature gradients or crack edges.

[0158] The maximum pooling operation downsamples with a 2x2 window to retain significant features, for example, reducing a 512x512 matrix to 256x256, reducing the amount of calculation, and generating the first feature set.

[0159] It should be noted that the pooling retains high-intensity features such as crack area pixel values of 0.9, while ignoring low-value noise.

[0160] In step S340, for the first feature set, the second and third convolution and maximum pooling operations are alternately performed, the channel weighting method is used to fuse the results of multiple convolution and pooling operations, and a preliminary feature set is obtained.

[0161] Preferably, the second and third convolution and pooling operations are alternately performed to further refine the features. For example, the second convolution uses a 5x5 convolution kernel to focus on larger range defect patterns, and the matrix is reduced to 128x128 after pooling. The third operation is similar to generate a 64x64 matrix.

[0162] The channel weighting method fuses multiple convolution results, for example, assigning a weight of 0.4 to the temperature channel, 0.3 to the echo, and 0.3 to the image, to generate a preliminary feature set. This hierarchical extraction and fusion method ensures comprehensive features and is suitable for subsequent cable state analysis.

[0163] It can be understood that the above method gradually converts heterogeneous data into a unified feature representation through resampling, alignment, convolution, and fusion.

[0164] Each step is linked, sequence resampling ensures consistent dimensions, time alignment eliminates bias, convolution pooling extracts key patterns, and channel weighting optimizes feature representation, providing high-quality input for cable defect detection.

[0165] Further, the cable defect detection method based on artificial intelligence provided by the embodiment comprises the following steps:

[0166] In step S410, a multi-modal feature vector is obtained from the preliminary feature set, a matrix decomposition method is used to split the multi-modal feature vector into independent modal matrices, a vector mapping method is used to convert the independent modal matrices into one-dimensional feature vectors, and a first feature vector set is obtained.

[0167] For example, in the cable defect detection scenario, the preliminary feature set usually contains high-dimensional representations of multi-modal data such as temperature, echo signal, and image.

[0168] The process of obtaining a multi-modal feature vector aims to integrate these features into a unified vector representation.

[0169] In one possible implementation, the preliminary feature set is a 64x64 three-channel matrix containing temperature gradient, echo amplitude, and image texture information.

[0170] Through flattening operation, each channel matrix is converted into a one-dimensional vector, for example, the temperature channel generates a 4096-dimensional vector, forming an initial multi-modal feature vector. This unified representation facilitates subsequent decomposition and processing.

[0171] The matrix decomposition method splits the multi-modal feature vector into independent modal matrices to separate the unique information of each modality.

[0172] Specifically, non-negative matrix factorization can decompose a 4096-dimensional vector into three independent matrices corresponding to temperature, echo, and image modalities. For example, the temperature matrix preserves the periodic changes of the time series, the echo matrix highlights the signal strength of the defect position, and the image matrix focuses on the geometric features of the crack.

[0173] After decomposition, each modality matrix has a size of about 1365 dimensions, maintaining information independence and providing clear input for subsequent mapping.

[0174] The vector mapping method converts the independent modality matrices into one-dimensional feature vectors to form a first feature vector set.

[0175] In an embodiment, the full connection layer mapping compresses the 1365-dimensional temperature matrix into a 512-dimensional vector, and the echo and image modalities are processed similarly to finally generate three 512-dimensional vectors. This compression retains key patterns such as temperature anomaly points or crack edge features, facilitating unified analysis.

[0176] Step S420, for the first feature vector set, the attention weight calculation module is used to weight process each modality feature vector, and a dynamic weight distribution is calculated through a softmax function to obtain a weighted feature vector set.

[0177] The attention weight calculation module weights the first feature vector set to highlight important modalities.

[0178] Preferably, the attention mechanism calculates a dynamic weight through a softmax function. For example, the weight of a certain cable section temperature vector is 0.5, the weight of an echo is 0.3, and the weight of an image is 0.2, reflecting a greater contribution of temperature anomalies to defect detection.

[0179] Step S430, if the weight value of any modality in the weighted feature vector set is lower than a preset threshold, the weight distribution is adjusted through a linear interpolation method, and the adjusted weighted feature vectors are merged through a vector splicing method to obtain a second feature vector set.

[0180] If the image modality weight is lower than the threshold value 0.25, the weight is adjusted by a linear interpolation method, for example, the image weight is increased to 0.25, and the temperature and echo weights are correspondingly reduced to 0.45 and 0.3, ensuring balanced weight distribution. This dynamic adjustment enhances the robustness of feature expression.

[0181] The vector splicing method merges the adjusted weighted feature vectors into a second feature vector set. For example, three 512-dimensional vectors are spliced into a 1536-dimensional vector, integrating multi-modality information. This splicing retains the weighted contribution of each modality, facilitating subsequent dimension reduction processing.

[0182] Step S440, according to the second feature vector set, a principal component analysis method is used to reduce the dimension of the high-dimensional feature vector, and an optimized multi-modality feature vector is generated through a matrix reorganization method to obtain a final feature vector set.

[0183] The principal component analysis method reduces the dimension of the second feature vector set to reduce redundancy.

[0184] It should be noted that the 1536-dimensional vector is reduced to 256-dimensional through principal component analysis, retaining 90% of the variance, to form an optimized multi-modality feature vector.

[0185] It can be understood that the vectors after dimension reduction still retain key information such as temperature anomaly, echo intensity and crack characteristics.

[0186] The matrix reorganization method further adjusts the vector structure, for example, reorganizes a 256-dimensional vector into three sub-vectors according to the modal proportion, and ensures that the final feature vector set is suitable for the cable state classification or defect positioning task. This hierarchical processing method generates an efficient feature representation through decomposition, mapping, weighting and dimension reduction, and provides a high-quality input for subsequent analysis.

[0187] Further, the cable defect detection method based on artificial intelligence provided by the embodiment comprises the following steps:

[0188] In step S510, an initial feature matrix is obtained from the multi-modal feature vector, a matrix decomposition method is used to split the initial feature matrix into independent modal sub-matrices, and a vector mapping method is used to convert the independent modal sub-matrices into one-dimensional vectors to obtain a first vector set.

[0189] For example, in the cable defect detection scenario, the initial feature matrix is obtained from the multi-modal feature vector.

[0190] Suppose the multi-modal feature vector integrates cable surface temperature distribution data, cable internal defect echo signal and cable appearance image data, forming a high-dimensional representation.

[0191] Specifically, the cable surface temperature distribution data reflects the thermal distribution of the cable surface, the cable internal defect echo signal captures the echo characteristics of the internal defects, and the cable appearance image data reveals the thermal anomaly area.

[0192] In one embodiment, the initial feature matrix is a 64x64x3 three-dimensional matrix, and each channel corresponds to one modality. The matrix decomposition method splits this matrix into three independent modal sub-matrices to separate the unique information of each modality.

[0193] Preferably, the singular value decomposition technique can separate the temperature sub-matrix into a 64x64 two-dimensional matrix, retaining the thermal gradient characteristics; the ultrasonic wave sub-matrix highlights the echo peak value; and the infrared image sub-matrix focuses on the hot spot area. After decomposition, each sub-matrix independently represents single modality information, facilitating subsequent processing.

[0194] It should be noted that the vector mapping method converts the independent modal sub-matrices into one-dimensional vectors to form a first vector set.

[0195] In one possible implementation, the temperature sub-matrix generates a 4096-dimensional vector through a flattening operation, and then compresses it into a 512-dimensional vector through a fully connected layer, retaining key thermal anomaly points.

[0196] The ultrasonic and infrared modalities are processed similarly, and finally three 512-dimensional vectors are generated to form the first vector set. This mapping method simplifies the processing complexity of high-dimensional data.

[0197] In step S520, the dynamic weights of the modal vectors are calculated for the first vector set using the attention mechanism, and the dynamic weight distribution is obtained by normalizing the dynamic weights through the softmax function.

[0198] The attention mechanism calculates the dynamic weights for the first vector set to highlight the key modalities. For example, the softmax function assigns weights according to the modal contribution, and the temperature vector may have a weight of 0.6, the ultrasonic wave has a weight of 0.25, and the infrared has a weight of 0.15.

[0199] In step S530, if the weight value of any modality in the dynamic weight distribution is lower than the preset threshold, the dynamic weight distribution is adjusted by the linear interpolation method, and the first vector set is weighted by the vector weighting method to obtain the weighted vector set.

[0200] If the infrared weight is lower than the threshold of 0.2, the linear interpolation method adjusts the weight to 0.2, and the temperature and ultrasonic wave weights are correspondingly reduced to 0.55 and 0.25. This dynamic adjustment ensures balanced contribution of each modality.

[0201] It can be understood that the vector weighting method weights the first vector set according to the dynamic weights to generate the weighted vector set.

[0202] Specifically, the 512-dimensional temperature vector is multiplied by the weight of 0.55, and the ultrasonic and infrared vectors are similarly weighted to generate the weighted vector set, highlighting important features.

[0203] In step S540, according to the weighted vector set, the weighted vectors of each modality are merged by the vector splicing method, and the merged vectors are dimensionally reduced by the principal component analysis method to obtain the fusion feature set.

[0204] The vector splicing method merges the three weighted vectors into a 1536-dimensional vector, integrating multi-modality information.

[0205] The principal component analysis method reduces the dimension of the merged vector to 256 dimensions, retaining 95% of the variance, to form the fusion feature set. For example, the reduced vector still retains the temperature anomaly, echo intensity, and hot spot features, which are suitable for cable defect classification tasks. This hierarchical processing method generates efficient feature representation through decomposition, mapping, weighting, and dimension reduction, providing high-quality input for subsequent analysis.

[0206] Further, the cable defect detection method based on artificial intelligence provided in the embodiment includes the following steps:

[0207] In step S610, a working condition parameter feature vector containing working condition parameters is obtained from the fusion feature set, a matrix decomposition method is used to split the working condition parameter feature vector into a time dimension sub-vector and a parameter dimension sub-vector, a linear interpolation method is used to align the time dimension sub-vector and the parameter dimension sub-vector, and a first feature set is obtained.

[0208] Exemplarily, in the cable state monitoring scene, a working condition parameter feature vector containing working condition parameters is obtained from the fusion feature set, aiming to capture dynamic characteristics in cable operation.

[0209] The fusion feature set integrates multi-modal data such as temperature, current load and vibration signals to form a high-dimensional feature representation.

[0210] Suppose the fusion feature set is a 256-dimensional vector containing time series information and working condition parameters such as temperature change over time, current fluctuation and vibration frequency.

[0211] The matrix decomposition method splits this vector into a time dimension sub-vector and a parameter dimension sub-vector.

[0212] Preferably, a non-negative matrix factorization technique is used to split the 256-dimensional vector into a 128-dimensional time dimension sub-vector reflecting the time series change of temperature and vibration, and a 128-dimensional parameter dimension sub-vector highlighting static working condition characteristics such as current load. This decomposition method facilitates the processing of dynamic and static information respectively.

[0213] In one possible implementation, a linear interpolation method is used to align the time dimension sub-vector and the parameter dimension sub-vector.

[0214] Since the time dimension sub-vector can have different sampling frequencies, such as temperature data sampled every second and current data sampled every 5 seconds, linear interpolation interpolates low-frequency data to a uniform time step to form an aligned first feature set. For example, the interpolated time dimension sub-vector contains temperature and vibration values every second, and the parameter dimension sub-vector synchronously matches the corresponding current values, ensuring time consistency.

[0215] In step S620, if the standard deviation of the time dimension sub-vector in the first feature set is lower than a preset threshold, a sliding window method is used to smooth the time dimension sub-vector, and a vector normalization method is used to normalize the smoothed sub-vector to obtain a second feature set.

[0216] It should be noted that if the standard deviation of the time dimension sub-vector in the first feature set is lower than the threshold, such as 0.1, it indicates that the data fluctuation is small, which can be caused by noise or insufficient sensor accuracy.

[0217] The sliding window method smoothes the time dimension sub-vector. Specifically, a mean filter with a window size of 5 is used to eliminate transient noise and make the temperature change curve smoother.

[0218] The vector normalization method normalizes the smoothed sub-vector, such as normalizing the temperature value to the interval of 0 to 1 to form a second feature set. This processing enhances the stability of the data and facilitates subsequent modeling.

[0219] In step S630, according to the second feature set, a long short-term memory network is used to model the time series of the time dimension sub-vector, and a gating recurrent unit is used to weight the parameter dimension sub-vector to obtain a third feature set.

[0220] Specifically, the long short-term memory network models the time series of the time dimension sub-vector in the second feature set. For example, the long short-term memory network contains 128 hidden units to capture the long-term dependence relationship between temperature and vibration, such as the trend of vibration frequency change after temperature rise.

[0221] The gating recurrent unit weights the parameter dimension sub-vector, preferably by assigning dynamic weights to the current load through 32 units to highlight the key working condition influence to form the third feature set. This modeling method effectively extracts the internal relationship between time series and parameters.

[0222] In step S640, the third feature set is reduced in dimension by a principal component analysis method, and the reduced time dimension features and parameter dimension features are combined by a vector splicing method to obtain a fourth feature set.

[0223] In one embodiment, the principal component analysis method reduces the third feature set in dimension, retains 90% of the variance, and compresses the high-dimensional features to 64 dimensions to reduce the computational complexity.

[0224] The vector splicing method combines the reduced time dimension features and parameter dimension features to form the fourth feature set. For example, the 64-dimensional vector after merging integrates the temperature change trend, vibration pattern and current load characteristics, providing a comprehensive feature representation for cable condition assessment. This hierarchical processing method generates an efficient feature set by decomposition, alignment, smoothing, modeling and dimension reduction, providing reliable input for subsequent analysis.

[0225] Further, the cable defect detection method based on artificial intelligence provided in the embodiment includes the following steps:

[0226] In step S710, an original data set containing defect features and working condition parameters is obtained from historical detection data, time dimension data is aligned by a linear interpolation method, and working condition parameter data is standardized by a vector normalization method to obtain a first data set.

[0227] In the field of cable condition monitoring, obtaining an original dataset containing defect features and working condition parameters from historical detection data is the basis for building a prediction model. For example, the original dataset may contain size data of cable surface defects, such as crack length, depth, and working condition parameters, such as environmental temperature, current load, and mechanical vibration frequency.

[0228] Suppose the dataset covers 1000 time points, the defect feature is crack length over time, and the working condition parameters include temperature and current values recorded every minute.

[0229] The dataset may have inconsistent time steps due to different sensor sampling frequencies, for example, crack length data is recorded every minute, while temperature data is recorded every 5 minutes.

[0230] The linear interpolation method is used to align the time dimension data. Specifically, for the low-frequency sampling of temperature data, the linear interpolation method is used to expand it to one data point per minute. For example, the temperature changes from 20℃ to 25℃ in a certain period, and the temperature values of the intermediate time points are generated by interpolation to form a time step consistent with the crack length data. This alignment ensures the consistency of the data time for subsequent analysis.

[0231] The vector normalization method is used to standardize the working condition parameter data to generate the first dataset. Preferably, the temperature and current values are normalized to the interval of 0 to 1. For example, the temperature range is 10℃ to 50℃, and after normalization, 20℃ is mapped to 0.25, and the current load from 100A to 500A is mapped to the corresponding proportional value. This standardization facilitates the comparability of data with different dimensions in the model.

[0232] In step S720, it is determined whether the standard deviation of the time dimension data of the first dataset is lower than a preset threshold value. If it is lower than the preset threshold value, the sliding window method is used to smooth the time dimension data, and the principal component analysis method is used to reduce the dimension of the smoothed data to obtain the second dataset.

[0233] The determination of whether the standard deviation of the time dimension data of the first dataset is lower than the preset threshold value, for example, 0.1, is to evaluate the volatility of the data.

[0234] It should be noted that if the standard deviation of the crack length data is lower than 0.1, it means that the data changes smoothly and may be affected by noise.

[0235] In one possible implementation, the sliding window method is used to smooth the time dimension data. For example, using a mean filter with a window size of 3, for crack length data of 1.2mm, 1.3mm, and 1.1mm at a certain time point, the smoothed value is 1.2mm, reducing the influence of instantaneous fluctuations.

[0236] The principal component analysis method performs dimension reduction processing on the smoothed data to generate a second data set.

[0237] It can be understood that the principal component analysis compresses the original 128-dimensional features to 32 dimensions by retaining 80% of the variance, retaining the crack length trend and key working condition parameter information. This dimension reduction reduces the computational complexity of subsequent modeling.

[0238] Step S730, according to the second data set, using a gated recurrent unit to weight the working condition parameter data, and merging the weighted working condition parameter data and the time dimension data through a feature splicing method to obtain a third data set.

[0239] The gated recurrent unit weights the working condition parameter data of the second data set.

[0240] Specifically, the gated recurrent unit assigns dynamic weights to the current load, for example, by highlighting the current impact under high temperature and high load through 16 units.

[0241] The feature splicing method merges the weighted working condition parameter data and the time dimension data to generate a third data set. For example, after splicing, a 64-dimensional vector is formed, integrating the crack length trend and the weighted temperature and current features.

[0242] Step S740, performing time series modeling on the third data set through a long short-term memory network to output a defect propagation trend prediction sequence containing the next three detection periods, and obtaining a feature change amount sequence.

[0243] The long short-term memory network performs time series modeling on the third data set to output a defect propagation trend prediction sequence for the next three detection periods.

[0244] In one embodiment, the long short-term memory network contains 64 hidden units to capture the long-term dependence of crack length changes with temperature and current. For example, the prediction result shows that the crack length may expand from 1.2 mm to 1.5 mm in the next three periods. This prediction sequence provides a key reference for cable maintenance.

[0245] Further, the cable defect detection method based on artificial intelligence provided by the embodiment comprises the following steps:

[0246] Step S810, obtaining a fifth feature set containing multi-dimensional data from the fusion feature set and the defect propagation trend prediction sequence, performing dimension reduction processing on the fifth feature set by a principal component analysis method, and normalizing the dimension-reduced data by a vector normalization method to obtain a sixth feature set.

[0247] In the field of cable condition monitoring, the fusion feature set and the prediction sequence provide the basis for constructing the multi-dimensional fifth feature set. For example, the fusion feature set may include cable crack width and surface corrosion degree, and the prediction sequence includes the crack expansion trend in the next three periods, such as the width increasing from 0.8 mm to 1.0 mm.

[0248] The working condition parameters such as humidity and voltage fluctuation are also included to form the fifth feature set, which may have a dimension of up to 100.

[0249] The principal component analysis method reduces the dimension of the first feature set and retains the main information.

[0250] Preferably, 90% of the variance is retained by principal component analysis, which compresses 100 dimensions to 20 dimensions, retains the crack width trend and key working condition parameters, and reduces the complexity of subsequent processing.

[0251] The vector normalization method standardizes the reduced dimension data to generate the sixth feature set.

[0252] Specifically, the crack width from 0.5 mm to 2.0 mm is normalized to the interval of 0 to 1, such as 1.0 mm mapped to 0.33; the humidity from 30% to 80% is mapped to the corresponding proportional value. This standardization ensures that different dimensional features are comparable. It should be noted that the normalized data facilitates subsequent threshold comparison and clustering processing.

[0253] Step S820, according to the sixth feature set, a preset threshold group is used to compare each feature dimension in the sixth feature set, if the feature value of the feature dimension exceeds the upper limit of the corresponding dimension in the threshold group, it is marked as an abnormal feature, and an abnormal feature set is obtained.

[0254] According to the sixth feature set, a preset threshold group is used for feature comparison. If the feature value of a certain feature dimension exceeds the upper limit, it is marked as abnormal.

[0255] In one embodiment, the upper limit of the normalized crack width value is 0.8, and if a data point is 0.9, it is marked as abnormal. The upper limit of the humidity is 0.7, and if the value is 0.75, it is also marked as abnormal. These abnormal features form an abnormal feature set. The threshold setting is based on historical data analysis to ensure the sensitivity of abnormal detection.

[0256] Step S830, the abnormal feature set is clustered by a classification module, and the K-means algorithm is used to group the abnormal feature set to obtain a plurality of feature subsets.

[0257] The abnormal feature set is clustered by a classification module, and the K-means algorithm is used for grouping.

[0258] It can be understood that the K-means algorithm divides the abnormal features into three groups, such as a high crack width group, a high humidity group, and a mixed abnormal group.

[0259] The K value is determined as 3 by the elbow rule according to the Euclidean distance between the characteristics. For example, the high crack width group contains crack data with normalized values of 0.85 to 0.95, reflecting potential serious defects.

[0260] In step S840, the matching degrees of the feature subset and the defect threshold group are calculated, and the matching degrees are comprehensively evaluated by a weighted average method to obtain a classification result.

[0261] The matching degrees of the feature subset and the defect threshold group are calculated. For example, the defect threshold group defines the serious crack width as a normalized value of 0.9 or more, and 80% of the data in the high crack width group exceeds this value, and the matching degree is high.

[0262] The matching degrees are comprehensively evaluated by a weighted average method, and the weights are allocated according to the importance of the characteristics, such as a crack width weight of 0.6 and a humidity weight of 0.4. The subset with a high matching degree suggests that maintenance needs to be prioritized.

[0263] In one possible implementation, the clustering result of the abnormal feature set can be further refined. For example, by adjusting the K value to 4, the case of high crack width accompanied by high humidity is distinguished, and the composite defect is accurately located.

[0264] Preferably, the matching degree evaluation is combined with historical maintenance records to verify the reliability of the classification result. This multi-dimensional analysis improves the defect recognition accuracy and provides a reliable basis for cable maintenance.

[0265] Further, the cable defect detection method based on artificial intelligence provided by the embodiment comprises the following steps:

[0266] In step S910, a multi-dimensional information feature vector containing multi-dimensional information is obtained from the cable state data, principal component analysis is used to reduce the dimension of the multi-dimensional information feature vector, and the data after dimension reduction is processed by a vector normalization method to obtain a standardized feature set.

[0267] In the field of cable state monitoring, obtaining a multi-dimensional information feature vector is the core of analyzing the health state of the cable.

[0268] For example, the multi-dimensional information feature vector can contain dimensions such as cable insulation resistance, partial discharge intensity, temperature distribution, and operating load.

[0269] The insulation resistance reflects the aging degree of the cable, the partial discharge intensity indicates internal defects, and the temperature and load are related to the operating environment. These data are collected in real time from sensors to form an initial high-dimensional feature vector, which can have a dimension of 50.

[0270] It should be noted that high-dimensional data increases the complexity of analysis, and therefore needs to be further processed.

[0271] Principal component analysis is a common method for dimensionality reduction, which is used to retain key information. Specifically, principal component analysis projects the original features to a new coordinate system through linear transformation, and preferentially retains the dimensions with larger variance. For example, 50-dimensional feature vectors are compressed to 15-dimensional through principal component analysis, retaining major information such as insulation resistance and partial discharge intensity. This dimensionality reduction reduces redundant dimensions and facilitates subsequent processing.

[0272] The vector normalization method standardizes the dimensionality reduced data to ensure that different dimensional features are comparable. In one embodiment, the insulation resistance is normalized from 10 MΩ to 100 MΩ to the interval of 0 to 1, such as 50 MΩ mapped to 0.4; the partial discharge intensity is mapped from 100 pC to 500 pC to 0 to 1, such as 300 pC mapped to 0.5.

[0273] Preferably, the normalized standard feature set provides a consistency basis for threshold comparison.

[0274] Step S920, according to the standard feature set, a preset first threshold and a second threshold are used to compare each feature dimension, if the feature value of the feature dimension exceeds the first threshold, it is marked as an abnormal feature; if the feature value of the feature dimension exceeds the second threshold, it is marked as a critical feature, to obtain an abnormal feature set and a critical feature set.

[0275] According to the standard feature set, the first threshold and the second threshold are used for feature comparison. Exemplarily, the first threshold is set to 0.7, and the second threshold is set to 0.9. If the normalized value of insulation resistance reaches 0.75, it is marked as abnormal; if it reaches 0.95, it is marked as critical. If the normalized value of partial discharge intensity is 0.8, it is marked as abnormal; if it is 0.92, it is marked as critical. These markings form an abnormal feature set and a critical feature set. It can be understood that the critical feature indicates a higher risk and needs to be paid attention to first.

[0276] Step S930, for the abnormal feature set and the critical feature set, a K-means algorithm is used to cluster and group the features to obtain a plurality of feature subsets.

[0277] For the abnormal feature set and the critical feature set, the K-means algorithm is used for clustering and grouping. For example, the algorithm divides the abnormal features into three groups: a high insulation resistance abnormal group, a high partial discharge abnormal group, and a mixed abnormal group. The grouping is based on the distance between the features, and the K value is determined to be 3 through analysis. In one possible implementation, the high partial discharge group contains data with normalized values of 0.85 to 0.95, reflecting the risk of serious defects.

[0278] Step S940, the matching degree value of the feature subset and the defect level is calculated by a weighted evaluation method to determine the defect level and the critical state, and a judgment result is obtained.

[0279] The matching degree of the feature subset and the defect level is calculated by a weighted evaluation method. Specifically, the defect level is divided into slight, medium and severe, corresponding to different threshold ranges. For example, a severe defect is defined as a partial discharge normalized value of 0.9 or higher. The weights are assigned according to the importance of the features, such as a partial discharge weight of 0.5, an insulation resistance of 0.3, and a temperature of 0.2. In the high partial discharge group, 90% of the data exceeds 0.9, the matching degree is high, and it is determined as a severe defect. The critical state indicates a potential deterioration trend.

[0280] In one embodiment, the clustering results are further refined in combination with operating load data. For example, the K value is adjusted to 4 to distinguish the case of high partial discharge accompanied by high temperature, and the composite defect is accurately located.

[0281] Preferably, the matching degree is evaluated with reference to historical data to enhance the reliability of the determination. This multi-dimensional analysis provides accurate basis for cable maintenance.

[0282] The present application relates to an artificial intelligence-based cable defect detection system for implementing the above-mentioned artificial intelligence-based cable defect detection method, which comprises a first acquisition module, a second acquisition module, a first generation module, a second generation module, a third acquisition module, a judgment module, an output module, a calculation module and a determination module. The first acquisition module is used to acquire cable detection data from a multi-modal data acquisition device, which includes an infrared thermal imaging unit, an ultrasonic detection unit and an optical sensing unit. The second acquisition module is used to standardize and time-align the cable detection data to obtain a normalized multi-modal data set. The first generation module is used to extract features from the multi-modal data set using a convolutional neural network, and generate a preliminary feature set by alternately performing three convolution operations and pooling operations. The second generation module is used to obtain a multi-modal feature vector from the preliminary feature set, and generate a dynamic weight distribution of each modal feature through an attention weight calculation module. The third acquisition module is used to weight and fuse the multi-modal feature vector based on the dynamic weight distribution to obtain a fused feature set. The judgment module is used to determine whether the fused feature set contains a working condition parameter feature vector. If the fused feature set contains a working condition parameter feature vector, the fused feature set is input into a long short-term memory network for time series modeling. The output module is used to output a defect expansion trend prediction sequence through the long short-term memory network, which contains the feature change amount of the future three detection periods. The calculation module is used to input the fused feature set and the defect expansion trend prediction sequence into a classification module, and perform feature matching degree calculation using a preset defect threshold group. The determination module is used to determine the cable defect level according to the feature matching degree calculation result. When the matching degree exceeds the first threshold, it is determined that there is a defect. When the matching degree exceeds the second threshold, it is determined that the defect reaches a critical state.

[0283] Compared with the prior art, the cable defect detection method and system based on artificial intelligence provided by the embodiment acquires cable detection data through a multi-modal data acquisition device, performs standardization processing and time alignment on the data, extracts features by using a convolutional neural network, and dynamically weights and fuses multi-modal features through an attention mechanism. If working condition parameters are contained, a long short-term memory network is used to perform time series modeling, and a defect expansion trend in the next three detection periods is predicted. Finally, the fused features and the predicted sequence are input into a classification module, and a defect grade is determined based on a preset threshold. The embodiment realizes accurate detection of cable defects and prediction of future development trends, effectively improves the safety and reliability of the power system, provides a scientific basis for cable maintenance decision-making, and has important engineering application value.

[0284] While the preferred embodiments of the application have been described, additional variations and modifications can be made to the embodiments by those skilled in the art once they learn of the basic inventive concepts. Therefore, the appended claims are intended to encompass within their scope all such variations and modifications as are included within the scope of the application. It should be apparent that the application is not limited to the specific embodiments described herein, but can be practiced with modification within the spirit and scope of the present application. Accordingly, though the present application has been described in some detail to provide a thorough understanding thereof, it will be apparent that certain minuscule changes can be made in the application without departing from the spirit and the scope of the application. Therefore, the scope of the present application should be determined not with the foregoing description, but with the appended claims, including all equivalents of the claims as encompassed by the language of the claims.

Claims

1. A cable defect detection method based on artificial intelligence, characterized in that: The following steps are involved: Acquiring cable detection data from a multimodal data acquisition device, wherein the multimodal data acquisition device includes an infrared thermal imaging unit, an ultrasonic detection unit, and an optical sensing unit; Performing standardization and time dimension alignment operations on the cable detection data to obtain a normalized multimodal data set; A convolutional neural network is used to extract features from the multimodal dataset, and a preliminary feature set is generated by alternating three convolution operations and pooling operations. Obtaining a multimodal feature vector based on the preliminary feature set, and generating a dynamic weight distribution of each modal feature through an attention weight calculation module; Performing weighted fusion on the multimodal feature vectors based on the dynamic weight distribution to obtain a fused feature set; Determining whether the fused feature set includes a working condition parameter feature vector, and if the fused feature set includes the working condition parameter feature vector, inputting the fused feature set into a long short-term memory network for time series modeling; Outputting a defect expansion trend prediction sequence through the long short-term memory network, wherein the defect expansion trend prediction sequence includes feature changes in the next three detection cycles; Inputting the fused feature set and the defect expansion trend prediction sequence into a classification module, and calculating feature matching using a preset defect threshold group; The cable defect level is determined based on the feature matching calculation results. When the matching degree exceeds a first threshold, it is determined that a defect exists. When the matching degree exceeds a second threshold, it is determined that the defect has reached a critical state. The step of determining whether the fused feature set includes a working condition parameter feature vector, and if the fused feature set includes the working condition parameter feature vector, inputting the fused feature set into a long short-term memory network for time series modeling includes: Obtaining a feature vector containing operating condition parameters from the fused feature set, splitting the feature vector into a time dimension sub-vector and a parameter dimension sub-vector using a matrix decomposition method, and aligning the time dimension sub-vector and the parameter dimension sub-vector using a linear interpolation method to obtain a first feature set; If the standard deviation of the time dimension sub-vector in the first feature set is lower than a preset threshold, smoothing the time dimension sub-vector using a sliding window method, and normalizing the smoothed sub-vector using a vector normalization method to obtain a second feature set; Based on the second feature set, a long short-term memory network is used to perform time series modeling on the time dimension sub-vector, and a gated recurrent unit is used to perform weighted processing on the parameter dimension sub-vector to obtain a third feature set; Performing dimensionality reduction processing on the third feature set by a principal component analysis method, and merging the time dimension features and parameter dimension features after dimensionality reduction by a vector splicing method to obtain a fourth feature set; The step of outputting a defect expansion trend prediction sequence through the long short-term memory network, wherein the defect expansion trend prediction sequence includes feature changes in the next three detection cycles, comprises: Acquire an original data set including defect characteristics and operating condition parameters from historical detection data, align the time dimension data using a linear interpolation method, and normalize the operating condition parameter data using a vector normalization method to obtain a first data set; Determine whether the standard deviation of the time dimension data of the first data set is lower than a preset threshold. If so, smooth the time dimension data using a sliding window method, and perform dimensionality reduction on the smoothed data using a principal component analysis method to obtain a second data set. According to the second data set, a gated recurrent unit is used to perform weighted processing on the operating condition parameter dimension data, and the weighted parameter dimension data is combined with the time dimension data by a feature splicing method to obtain a third data set; The third data set is subjected to time series modeling by the long short-term memory network, and a defect expansion trend prediction sequence including the next three detection cycles is output to obtain a feature variation sequence.

2. The cable defect detection method based on artificial intelligence according to claim 1, characterized in that: The steps of performing standardization processing and time dimension alignment operations on the cable detection data to obtain a normalized multimodal data set include: Acquire original cable detection data from a multimodal data acquisition device, and perform anomaly detection on cable surface temperature distribution data, cable internal defect echo signals, and cable appearance image data in the original cable detection data using a data cleaning method to obtain a first data set; Based on the first data set, a format conversion method is used to convert data of different modes into a unified structure. If the timestamp deviation between the cable surface temperature distribution data and the cable internal defect echo signal exceeds a preset threshold, the timestamps are aligned by interpolation to obtain a second data set. For the second data set, the cable surface temperature distribution data and the cable internal defect echo signal are normalized using the z-score normalization method, and the cable appearance image data is grayscale normalized to obtain a third data set; The third data set is segmented by a batch processing method, and the segmented cable surface temperature distribution data, cable internal defect echo signals and cable appearance image data are spliced ​​according to the time dimension using a multimodal fusion algorithm to obtain a normalized multimodal data set.

3. The cable defect detection method based on artificial intelligence according to claim 1, characterized in that: The step of extracting features from the multimodal dataset using a convolutional neural network and generating a preliminary feature set by alternately performing three convolution operations and pooling operations includes: Acquiring cable surface temperature distribution data, internal defect echo signals, and cable appearance image data from a multimodal data set, and converting the cable surface temperature distribution data and internal defect echo signals into a two-dimensional matrix matching the resolution of the appearance image using a sequence resampling method to obtain a first intermediate data set; For the first intermediate data set, if the time stamp of the cable surface temperature distribution data or the internal defect echo signal deviates from the cable appearance image data by more than a preset threshold, the time stamps are aligned by linear interpolation, and the aligned multi-channel matrices are merged by matrix splicing to obtain a second intermediate data set; Performing a first convolution operation on the second intermediate dataset using a convolutional neural network, extracting features using a preset convolution kernel for the multi-channel matrix, and downsampling the data using a maximum pooling operation to obtain a first feature set; For the first feature set, the second and third convolution and maximum pooling operations are performed alternately, and the channel weighted method is used to fuse the results of multiple convolution and pooling to obtain a preliminary feature set.

4. The cable defect detection method based on artificial intelligence according to claim 1, characterized in that: The steps of obtaining a multimodal feature vector according to the preliminary feature set and generating a dynamic weight distribution of each modal feature through an attention weight calculation module include: Obtaining multimodal feature vectors according to the preliminary feature set, splitting the multimodal feature vectors into independent modal matrices using a matrix decomposition method, and converting the independent modal matrices into one-dimensional feature vectors using a vector mapping method to obtain a first feature vector set; For the first feature vector set, the attention weight calculation module is used to perform weighted processing on the feature vectors of each modality, and the dynamic weight distribution is calculated by the softmax function to obtain a weighted feature vector set; If the weight value of any mode in the weighted feature vector set is lower than a preset threshold, the weight distribution is adjusted by a linear interpolation method, and the adjusted weighted feature vectors are merged by a vector splicing method to obtain a second feature vector set; According to the second feature vector set, the principal component analysis method is used to reduce the dimensionality of the high-dimensional feature vectors, and the optimized multimodal feature vectors are generated by the matrix recombination method to obtain the final feature vector set.

5. The cable defect detection method based on artificial intelligence according to claim 1, characterized in that: The step of performing weighted fusion on the multimodal feature vectors based on the dynamic weight distribution to obtain a fused feature set includes: Obtaining an initial feature matrix from the multimodal feature vector, splitting the initial feature matrix into independent modal sub-matrices using a matrix decomposition method, and converting the independent modal sub-matrices into one-dimensional vectors using a vector mapping method to obtain a first vector set; For the first vector set, an attention mechanism is used to calculate the dynamic weight of each modal vector, and the dynamic weight is normalized by a softmax function to obtain a dynamic weight distribution; If the weight value of any mode in the dynamic weight distribution is lower than a preset threshold, the dynamic weight distribution is adjusted by a linear interpolation method, and the first vector set is weighted by a vector weighting method to obtain a weighted vector set; According to the weighted vector set, the weighted vectors of each mode are merged using a vector splicing method, and the merged vectors are subjected to dimensionality reduction processing using a principal component analysis method to obtain a fusion feature set.

6. The cable defect detection method based on artificial intelligence according to claim 1, characterized in that: The steps of inputting the fused feature set and the defect expansion trend prediction sequence into a classification module and calculating the feature matching degree using a preset defect threshold group include: Obtaining a fifth feature set containing multi-dimensional data from the fused feature set and the defect expansion trend prediction sequence, performing dimensionality reduction processing on the fifth feature set using a principal component analysis method, and normalizing the reduced dimensionality data using a vector normalization method to obtain a sixth feature set; According to the sixth feature set, each feature dimension in the sixth feature set is compared using a preset threshold group, and if a feature value exceeds an upper limit of the corresponding dimension in the threshold group, it is marked as an abnormal feature, thereby obtaining an abnormal feature set; Clustering the abnormal feature set using a classification module, and grouping the abnormal feature set using a K-means algorithm to obtain multiple feature subsets; For the feature subset, the matching degree between the feature subset and the defect threshold group is calculated, and the matching degree is comprehensively evaluated by a weighted average method to obtain a classification result.

7. The cable defect detection method based on artificial intelligence according to claim 1, characterized in that: The step of determining the cable defect level based on the feature matching calculation result, determining that a defect exists when the matching degree exceeds a first threshold, and determining that the defect reaches a critical state when the matching degree exceeds a second threshold includes: Obtaining a feature vector containing multi-dimensional information from the cable status data, performing dimensionality reduction processing on the feature vector using principal component analysis, and processing the reduced-dimensional data using a vector normalization method to obtain a standardized feature set; According to the standardized feature set, each feature dimension is compared using a preset first threshold and a second threshold. If the feature value exceeds the first threshold, it is marked as an abnormal feature; if the feature value exceeds the second threshold, it is marked as a critical feature, thereby obtaining an abnormal feature set and a critical feature set; For the abnormal feature set and the critical feature set, the K-means algorithm is used to cluster and group the features to obtain multiple feature subsets; The matching value between the feature subset and the defect level is calculated by a weighted evaluation method to determine the defect level and critical state, and obtain a judgment result.

8. An artificial intelligence-based cable defect detection system, used to implement the artificial intelligence-based cable defect detection method according to any one of claims 1 to 7, characterized in that: The artificial intelligence-based cable defect detection system includes: A first acquisition module is used to acquire cable detection data from a multimodal data acquisition device, wherein the multimodal data acquisition device includes an infrared thermal imaging unit, an ultrasonic detection unit, and an optical sensing unit; a second acquisition module, configured to perform standardization processing and time dimension alignment operations on the cable detection data to obtain a normalized multimodal data set; A first generation module is used to extract features from the multimodal dataset using a convolutional neural network, and to generate a preliminary feature set by alternately performing three convolution operations and pooling operations; A second generation module is used to obtain a multimodal feature vector based on the preliminary feature set, and generate a dynamic weight distribution of each modal feature through an attention weight calculation module; a third acquisition module, configured to perform weighted fusion on the multimodal feature vectors based on the dynamic weight distribution to obtain a fused feature set; a judgment module, configured to judge whether the fused feature set contains a working condition parameter feature vector, and if the fused feature set contains the working condition parameter feature vector, inputting the fused feature set into a long short-term memory network for time series modeling; An output module is configured to output a defect expansion trend prediction sequence through the long short-term memory network, wherein the defect expansion trend prediction sequence includes feature changes in the next three detection cycles; A calculation module, configured to input the fused feature set and the defect expansion trend prediction sequence into a classification module, and perform feature matching calculation using a preset defect threshold group; The judgment module is used to determine the cable defect level based on the feature matching calculation result, and judge that a defect exists when the matching degree exceeds a first threshold, and judge that the defect reaches a critical state when the matching degree exceeds a second threshold.

Images