Method and apparatus for determining cable external force damage results based on edge computing

The multimodal recognition model optimized by edge computing and gradient vibration particle swarm optimization solves the problem of low accuracy in monitoring cable external force damage, and achieves efficient and accurate identification of cable external force damage.

CN123115759APending Publication Date: 2026-11-17STATE GRID BEIJING ELECTRIC POWER CO +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611289219.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-24
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies for monitoring cable damage by external forces have low accuracy, high false alarm rate, poor environmental adaptability, and lack effective solutions.

Method used

A multimodal recognition model based on edge computing is adopted. The gradient vibration particle swarm algorithm is used to train the deep learning model for hyperparameter optimization. The social learning mechanism, neighborhood gradient search mechanism and information fusion vibration search mechanism are combined to optimize the feature extraction and recognition model, and realize the fusion recognition of multimodal data.

Benefits of technology

It improves the accuracy of identifying cable damage caused by external forces, enables real-time and accurate monitoring at edge computing nodes, reduces false alarm rates, and enhances environmental adaptability.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115759A_ABST
    Figure CN123115759A_ABST
Patent Text Reader

Abstract

This invention discloses a method and apparatus for determining cable external force damage results based on edge computing. The method includes: acquiring a multimodal recognition model set deployed on edge computing nodes, wherein multiple feature extraction models are trained on corresponding initial deep learning models using a gradient vibration particle swarm optimization algorithm based on corresponding sample modal data. The gradient vibration particle swarm optimization algorithm includes a social learning mechanism, a neighborhood gradient search mechanism, and an information fusion vibration search mechanism; determining real-time modal data corresponding to various modal types; scheduling the corresponding feature extraction models to extract features from the corresponding real-time modal data to obtain multiple real-time modal features; fusing the multiple real-time features to obtain a fused feature to be identified; and scheduling the feature recognition model to identify the fused feature to obtain the cable external force damage identification result. This invention solves the technical problem of low accuracy in monitoring cable external force damage in related technologies.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of power system safety monitoring technology, and more specifically, to a method and apparatus for determining the results of cable external force damage based on edge computing. Background Technology

[0002] With the acceleration of urbanization, the scale of underground cable networks is growing rapidly. Construction activities around cable ducts, such as mechanical excavation, have become a major cause of cable damage or breakage, seriously threatening the safe and stable operation of the power grid. Traditional methods for monitoring cable damage by external forces often rely on single sensors (such as vibration sensors or video surveillance), which suffer from high false alarm rates and poor environmental adaptability.

[0003] There is currently no effective solution to the above problems. Summary of the Invention

[0004] This invention provides a method and apparatus for determining the result of cable external force damage based on edge computing, so as to at least solve the technical problem of low accuracy in monitoring cable external force damage in related technologies.

[0005] According to one aspect of the present invention, a method for determining the result of cable external force damage based on edge computing is provided, applied to an edge computing node, comprising: acquiring a multimodal recognition model set deployed on the edge computing node, wherein the multimodal recognition model set includes multiple feature extraction models and feature recognition models, the multiple feature extraction models are used to extract features from different types of modal data, the multiple feature extraction models are obtained by using a gradient vibration particle swarm optimization algorithm to train the corresponding initial deep learning model through hyperparameter optimization of the corresponding sample modal data, the sample modal data includes first modal data under normal cable operation and second modal data under cable damage by external force, the feature recognition model is trained based on sample fusion data, the gradient vibration particle swarm optimization algorithm includes a social learning mechanism, a neighborhood gradient search mechanism and an information fusion vibration search mechanism, the social learning mechanism is used to control the balance between global search and local development of the particle swarm, and so on. The neighborhood gradient search mechanism is used to search within the neighborhood of the current optimal particle to improve optimization accuracy. The information fusion vibration search mechanism is used to enable the particle swarm to jump out of the local optimum region to explore more hyperparameter space. In each iteration of the gradient vibration particle swarm algorithm, the social learning mechanism, the neighborhood gradient search mechanism, and the information fusion vibration search mechanism are executed sequentially to complete a complete update of the particle swarm until the preset maximum number of iterations is reached. The hyperparameters corresponding to the current optimal particle are used as the target hyperparameters of the corresponding deep learning model. Each particle in the particle swarm represents a set of candidate hyperparameters. Real-time modal data corresponding to multiple modal types are determined. The corresponding feature extraction model is scheduled to extract features from the corresponding real-time modal data to obtain multiple real-time modal features. The multiple real-time features are fused to obtain the fused features to be identified. The feature recognition model is scheduled to identify the fused features to be identified to obtain the cable external force damage identification result.

[0006] Optionally, the social learning mechanism is used to update the position of the particle swarm through adaptive weights and sine and cosine operators to maintain population diversity. Maintaining population diversity is used to prevent the particle swarm from clustering to a local optimum less than a predetermined number of times during the optimization process, so as to maintain global exploration capability. The neighborhood gradient search mechanism is used to guide particles to search within the neighborhood to approach the local optimum through pseudo gradient information. The pseudo gradient information is an alternative gradient vector constructed using the positional difference between the optimal particle and random particles. The neighborhood is the hyperparameter space region within a predetermined distance range of the current optimal particle. The local optimum is the hyperparameter combination that minimizes the loss function value of the deep learning model within the neighborhood. The loss function value is a quantification of the degree of difference between the predicted output and the true output of the deep learning model on the training data. The information fusion vibration search mechanism is used to generate vibration perturbations through the cooperative information between particles to help particles jump out of the local optimum.

[0007] Optionally, before obtaining the set of multimodal recognition models deployed on edge computing nodes, the method further includes: for any feature extraction model, using the corresponding sample modal data as the expected input data and the corresponding training label as the expected output, training the corresponding initial deep learning model using the gradient vibration particle swarm algorithm to obtain the corresponding feature extraction model.

[0008] Optionally, using the corresponding sample modal data as the expected input data and the corresponding training label as the expected output, the gradient vibration particle swarm optimization algorithm is used to train the corresponding initial deep learning model to obtain the corresponding feature extraction model. This includes: determining an initial particle swarm based on the initial hyperparameters of the initial deep learning model; for each particle in the initial particle swarm, using the sample modal data as the expected input data and the training label as the expected output, determining the fitness of each particle; determining the historical optimal position of each particle and the current optimal particle of the particle swarm during the current training process based on the fitness of each particle; training the corresponding initial deep learning model using the gradient vibration particle swarm optimization algorithm based on the historical optimal position of each particle and the current optimal particle of the particle swarm to output the globally optimal particle; and using the hyperparameters in the globally optimal particle as the target hyperparameters of the initial deep learning model to obtain the trained feature extraction model.

[0009] Optionally, the operation of the social learning mechanism includes: obtaining a first updated particle swarm; obtaining adaptive weights corresponding to the first updated particle swarm at the current training iteration, wherein the adaptive weights adopt an exponential decay form and are superimposed with cosine fluctuations on the decay basis; for each first updated particle in the first updated particle swarm, based on the adaptive weights, the current position of the first updated particle at the current training iteration, the position of the current best particle in the current training process, and the historical best position of the first updated particle, the updated position of the first updated particle in the next training iteration is determined using a sine and cosine algorithm; the calculated updated position is used as the position of the first updated particle after social learning, and after all the first updated particles are updated, a second updated particle swarm including multiple second updated particles is obtained.

[0010] Optionally, the operation of the neighborhood gradient search mechanism includes: obtaining pseudo-gradient update information corresponding to the second updated particle swarm, and using the positional difference between the current optimal particle in the current training iteration and the second updated particle randomly selected from the second updated particle swarm, combined with the step size parameter, generating a random vector with the same dimension as the hyperparameter, as the pseudo-gradient update information for the next iteration, wherein the step size parameter is determined based on the distance between the current optimal particle and the randomly selected second updated particle and the previous pseudo-gradient update information; based on the current training iteration, obtaining the first neighborhood gradient search coefficient and the second neighborhood gradient search coefficient, wherein the first neighborhood gradient search coefficient and the... The second neighborhood gradient search coefficients all change non-linearly with time. The first neighborhood gradient search coefficient is used to control the degree of repulsion of the second updated particle away from the worst particle, and the second neighborhood gradient search coefficient is used to control the degree of step size adjustment of the second updated particle moving towards the optimal particle. For each second updated particle in the second updated particle group, the position of the second updated particle is updated by neighborhood search according to the pseudo gradient update information, the first neighborhood gradient search coefficient and the second neighborhood gradient search coefficient, so as to obtain the position of the second updated particle after neighborhood gradient search. After all the second updated particles are updated, a third updated particle group including multiple third updated particles is obtained.

[0011] Optionally, the operation of the information fusion vibration search mechanism includes: randomly matching one other third update particle as a cooperative particle for each third update particle in the third update particle swarm; for each third update particle, determining the cooperative information factor of the third update particle based on the distance between the corresponding third update particle and the corresponding cooperative particle and the current training number, wherein the cooperative information factor is used to control the amplitude of the vibration search; for each third update particle, performing a vibration search update on the position of the third update particle based on the cooperative information factor and the current optimal particle, obtaining the position of the third update particle after the vibration search, and obtaining the fourth update particle swarm after the vibration search after all third update particles are updated, thereby completing one complete iteration update of the particle swarm.

[0012] According to one aspect of the present invention, a device for determining the result of cable external force damage based on edge computing is provided, applied to an edge computing node, comprising: an acquisition module for acquiring a multimodal recognition model set deployed on the edge computing node, wherein the multimodal recognition model set includes multiple feature extraction models and feature recognition models, the multiple feature extraction models being used to extract features from different types of modal data, the multiple feature extraction models being obtained by using a gradient vibration particle swarm optimization algorithm to train the corresponding initial deep learning model through hyperparameter optimization training on the corresponding sample modal data, the sample modal data including first modal data under normal cable operation and second modal data under cable damage by external force, the feature recognition model being trained based on sample fusion data, the gradient vibration particle swarm optimization algorithm including a social learning mechanism, a neighborhood gradient search mechanism, and an information fusion vibration search mechanism, the social learning mechanism being used to control the balance between global search and local development of the particle swarm, the neighborhood gradient search mechanism being used to... To improve optimization accuracy, a search is performed within the neighborhood of the current optimal particle. The information fusion vibration search mechanism is used to enable the particle swarm to escape the local optimum and explore more hyperparameter space. In each iteration of the gradient vibration particle swarm algorithm, the social learning mechanism, the neighborhood gradient search mechanism, and the information fusion vibration search mechanism are executed sequentially to complete a full update of the particle swarm until the preset maximum number of iterations is reached. The hyperparameters corresponding to the current optimal particle are used as the target hyperparameters of the corresponding deep learning model. Each particle in the particle swarm represents a set of candidate hyperparameters. The determination module is used to determine the real-time modal data corresponding to multiple modal types. The extraction module is used to schedule the corresponding feature extraction model to extract features from the corresponding real-time modal data to obtain multiple real-time modal features. The fusion module is used to fuse multiple real-time features to obtain the fused features to be identified. The identification module is used to schedule the feature identification model to identify the fused features to be identified to obtain the cable external force damage identification result.

[0013] According to one aspect of the present invention, an electronic device is provided, comprising: a processor; a memory for storing processor-executable instructions; wherein the processor is configured to execute the instructions to implement the method for determining cable external force damage results based on edge computing as described above.

[0014] According to one aspect of the present invention, a computer-readable storage medium is provided, which, when the instructions in the computer-readable storage medium are executed by a processor of an electronic device, enables the electronic device to perform the cable external force damage determination method based on edge computing as described above.

[0015] In this embodiment of the invention, a multimodal recognition model set deployed on edge computing nodes is obtained. This set includes multiple feature extraction models and feature recognition models. The feature extraction models extract features from different types of modal data. These models are obtained by training corresponding initial deep learning models using a gradient vibration particle swarm optimization algorithm on the corresponding sample modal data. The sample modal data includes first modal data under normal cable operation and second modal data under external force damage. The feature recognition model is trained based on sample fusion data. The gradient vibration particle swarm optimization algorithm includes a social learning mechanism, a neighborhood gradient search mechanism, and an information fusion vibration search mechanism. The social learning mechanism controls the balance between global search and local development of the particle swarm, while the neighborhood gradient search mechanism optimizes the particle swarm within its current optimal neighborhood. A line search is used to improve optimization accuracy, and an information fusion vibration search mechanism is used to enable the particle swarm to jump out of the local optimum region to explore more hyperparameter space. In each iteration of the gradient vibration particle swarm algorithm, the social learning mechanism, the neighborhood gradient search mechanism, and the information fusion vibration search mechanism are executed sequentially to complete a complete update of the particle swarm until the preset maximum number of iterations is reached. The hyperparameters corresponding to the current best particle are used as the target hyperparameters of the corresponding deep learning model, and each particle in the particle swarm represents a set of candidate hyperparameters. Real-time modal data corresponding to multiple modal types are determined; the corresponding feature extraction model is scheduled to extract features from the corresponding real-time modal data to obtain multiple real-time modal features; multiple real-time features are fused to obtain the fused features to be identified; the feature recognition model is scheduled to identify the fused features to be identified to obtain the cable external force damage identification result. This paper employs a gradient vibration particle swarm optimization (PSO) algorithm to train multiple feature extraction models for hyperparameter optimization. In each iteration of the algorithm, a social learning mechanism balances the global search and local exploration of the particle swarm to traverse the hyperparameter space; a neighborhood gradient search mechanism performs a fine-grained search within the neighborhood of the current optimal particle to verify the existence of a better solution; and an information fusion vibration search mechanism allows the particle swarm to escape local optima and explore further regions. These three mechanisms are executed sequentially and complement each other in each iteration, jointly completing a full update of the particle swarm. This allows the particle swarm to gradually approach the globally optimal hyperparameters over multiple iterations, giving the model stronger capabilities. This achieves the technical effect of extracting the most representative modal features from real-time modal data using the optimal feature extraction model, thereby improving the accuracy of cable external force damage identification. This solves the technical problem of low accuracy in cable external force damage monitoring in related technologies. Attached Figure Description

[0016] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0017] Figure 1 This is a flowchart of a method for determining cable external force damage results based on edge computing according to an embodiment of the present invention;

[0018] Figure 2 A flowchart of a cable external force damage monitoring method based on edge computing, provided as an optional embodiment of the present invention;

[0019] Figure 3 A flowchart for training a first deep learning model using a gradient vibration particle swarm optimization algorithm is provided as an optional embodiment of the present invention.

[0020] Figure 4 This is a structural block diagram of a cable external force damage determination device based on edge computing according to an embodiment of the present invention. Detailed Implementation

[0021] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0022] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0023] Example 1

[0024] According to an embodiment of the present invention, an embodiment of a method for determining the result of cable external force damage based on edge computing is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0025] Figure 1 This is a flowchart of a method for determining cable external force damage results based on edge computing, according to an embodiment of the present invention. Figure 1 As shown, the method includes the following steps:

[0026] Step S102: Obtain a multimodal recognition model set deployed on edge computing nodes. This set includes multiple feature extraction models and feature recognition models. The feature extraction models extract features from different types of modal data. These models are trained using the gradient vibration particle swarm optimization algorithm on their respective initial deep learning models based on the corresponding sample modal data. The sample modal data includes first-mode data under normal cable operation and second-mode data under conditions of cable damage from external forces. The feature recognition model is trained based on the sample fusion data. The gradient vibration particle swarm optimization algorithm includes a social learning mechanism, a neighborhood gradient search mechanism, and a signal recognition mechanism. The algorithm employs a fusion vibrational search mechanism, a social learning mechanism to balance global search and local development of the particle swarm, a neighborhood gradient search mechanism to search within the neighborhood of the current best particle to improve optimization accuracy, and an information fusion vibrational search mechanism to enable the particle swarm to escape local optima and explore more hyperparameter spaces. In each iteration of the gradient vibrational particle swarm algorithm, the social learning mechanism, the neighborhood gradient search mechanism, and the information fusion vibrational search mechanism are executed sequentially to complete a full update of the particle swarm until the preset maximum number of iterations is reached. The hyperparameters corresponding to the current best particle are used as the target hyperparameters of the corresponding deep learning model, and each particle in the particle swarm represents a set of candidate hyperparameters.

[0027] Edge computing nodes refer to computing devices deployed on-site in cable tunnels that have the ability to collect, preprocess, and infer data, enabling them to complete data processing and decision-making locally without relying on the cloud. For example, smart gateway devices deployed next to cable tunnels.

[0028] The multimodal recognition model set refers to a collection of multiple deep learning models deployed in edge computing nodes, including multiple feature extraction models and feature recognition models, which can provide edge computing nodes with complete feature extraction and recognition functions. For example, the multimodal recognition model set may include vibration signal feature extraction models, image data feature extraction models, and fusion recognition models.

[0029] Among them, feature extraction models refer to deep learning models used to extract high-dimensional feature vectors from input data of a single modality type. They can transform the original input data into more discriminative feature representations. For example, vibration signal feature extraction models are used to extract vibration features from time-frequency maps, and image data feature extraction models are used to extract visual features from images.

[0030] Among them, the feature recognition model refers to a deep learning model used to classify and identify the fused feature vectors, and can output the classification result of the cable status. For example, after the feature recognition model is input into the fused features, it outputs a binary classification result of whether the cable is in a state of external force damage.

[0031] Among them, sample modal data refers to modal data samples with known state labels used to train the feature extraction model, which can provide supervision signals for model training. For example, the vibration time-frequency diagram of a cable under normal operating conditions and the vibration time-frequency diagram under external force damage conditions.

[0032] Among them, the gradient vibration particle swarm optimization algorithm refers to a hyperparameter optimization algorithm that sequentially executes a social learning mechanism, a neighborhood gradient search mechanism, and an information fusion vibration search mechanism to update the position of the particle swarm in each iteration. It enables the particle swarm to gradually approach the global optimum in the hyperparameter space. For example, by executing the three mechanisms in sequence, the particle swarm finds the hyperparameter combination that makes the model perform best after multiple iterations.

[0033] Deep learning models refer to neural network models with multi-layer nonlinear transformation structures that can automatically learn hierarchical feature representations from data. For example, convolutional neural networks are used to extract spatial features from time-frequency maps.

[0034] Hyperparameter optimization training refers to the process of searching for the optimal hyperparameter values ​​in the hyperparameter space through optimization algorithms to achieve the best model performance. This enables deep learning models to perform at their best on specific tasks, such as searching for the optimal values ​​of hyperparameters such as learning rate, batch size, and number of convolutional kernels in a convolutional neural network.

[0035] In this context, a particle swarm refers to a collection of multiple particles, each representing a set of candidate hyperparameter combinations. The optimal solution can be searched by collective movement in the hyperparameter space. For example, a particle swarm may contain 50 particles, each carrying a set of values ​​for the learning rate and the number of convolutional kernels.

[0036] Among them, the social learning mechanism refers to the mechanism by which particles update their own positions by learning from their own historical best positions and the group's best positions. It can control the balance between global search and local development of the particle swarm. For example, in each iteration, the particles adjust their own positions according to adaptive weights and sine and cosine operators.

[0037] Among them, the neighborhood gradient search mechanism refers to the mechanism by which a particle performs a fine search in the local hyperparameter space region near the current best particle, which can improve the accuracy of local optimization. For example, the particle finely adjusts its position along the pseudo gradient direction in the neighborhood of the current best particle.

[0038] Among them, the information fusion vibration search mechanism refers to the mechanism by which particles update their own positions by cooperating with other particles and generating vibrational perturbations. This mechanism enables particles trapped in a local optimum to jump out of the current region to explore a more distant hyperparameter space. For example, particles generate vibrational amplitude based on the distance between themselves and randomly matched cooperating particles.

[0039] The optimal particle refers to the particle with the highest fitness at the current iteration number. It represents the currently discovered optimal combination of hyperparameters and can serve as a target for other particles to learn from and gravitate towards. For example, the hyperparameter combination corresponding to the current optimal particle makes the model achieve the highest accuracy on the validation set.

[0040] The hyperparameter space refers to the high-dimensional space consisting of the range of values ​​of all hyperparameters to be optimized, which can provide a search domain for the particle swarm. For example, the search range for the learning rate is 0.001 to 0.1, and the search range for the number of convolution kernels is 16 to 256.

[0041] The maximum number of iterations refers to the preset iteration termination condition of the particle swarm optimization algorithm, which can control the computational cost of hyperparameter search. For example, the preset maximum number of iterations is 200, and the globally optimal particle is output after reaching this number.

[0042] In this step, a multimodal recognition model set deployed on the edge computing node is acquired. Each model in this set has undergone hyperparameter optimization training using the gradient vibration particle swarm optimization algorithm, which sequentially updates the particle swarm in each iteration through three mechanisms. Through this step, the edge computing node obtains a complete model set capable of extracting features from multiple modal data and performing fusion recognition, providing a callable inference engine for subsequent real-time monitoring.

[0043] Step S104: Determine the real-time modal data corresponding to each of the multiple modal types;

[0044] Real-time modal data refers to the field data corresponding to each modal type collected in real time by edge computing nodes. It can reflect the actual state of the cable channel at the current moment. For example, the time-frequency diagram data obtained by preprocessing the real-time collected cable vibration signal, and the image data obtained by preprocessing the real-time collected cable image.

[0045] In this step, the edge computing nodes determine the real-time modal data corresponding to various modalities. This involves performing the same preprocessing operations as in the training phase on the raw, real-time data, ensuring that the preprocessed data format matches the input format of the feature extraction models. This step transforms the raw data into an input format directly acceptable to each feature extraction model, preparing it for subsequent feature extraction.

[0046] Step S106: Schedule the corresponding feature extraction model to extract features from the corresponding real-time modal data to obtain multiple real-time modal features;

[0047] Among them, real-time modal features refer to the high-dimensional feature vectors extracted by the feature extraction model from real-time modal data, which can represent the discriminative information of the original data in a compact vector form. For example, the vibration signal feature extraction model outputs a 512-dimensional feature vector.

[0048] In this step, edge computing nodes schedule multiple feature extraction models, each extracting features from real-time modal data of its corresponding modality type, resulting in multiple real-time modal features. This step transforms the original multimodal data into a high-level semantic feature representation, achieving data dimensionality reduction and feature abstraction, and providing high-quality feature input for subsequent fusion and recognition.

[0049] Step S108: Fuse multiple real-time features to obtain the fused features to be identified;

[0050] Among them, multiple real-time features refer to the same set of feature vectors as the multiple real-time modal features extracted in step S106, which can jointly describe the field condition of the cable channel from multiple modal dimensions. For example, vibration features and image features describe the field condition from the two dimensions of vibration energy and visual information, respectively.

[0051] Among them, the fusion feature to be identified refers to the joint feature vector obtained by splicing or fusing multiple real-time modal features. It can integrate complementary information from different modalities for unified identification. For example, 512-dimensional vibration features and 512-dimensional image features are spliced ​​together to form a 1024-dimensional fusion feature vector.

[0052] In this step, edge computing nodes fuse multiple real-time features to obtain a single fused feature to be identified. This step integrates feature information from multiple modalities into a unified feature representation, enabling complementary information between different modalities to be fully utilized by the feature recognition model within the same feature space.

[0053] Step S110: The scheduling feature recognition model identifies the features to be identified and fused, and obtains the cable external force damage identification result.

[0054] Among them, the cable external force damage identification result refers to the cable status judgment result output by the feature recognition model after classifying the fused features. It can indicate whether the current cable channel is being damaged by external force. For example, the identification result is "there is a risk of external force damage" or "normal status".

[0055] In this step, the edge computing node schedules the feature recognition model to classify and identify the fused features, outputting the cable external force damage identification result. Through this step, the edge computing node completes the entire inference chain from data acquisition to status recognition locally, realizing on-site determination of external force damage events, avoiding data transmission delays, and providing technical support for timely early warning.

[0056] Through steps S102-S110 above, a multimodal recognition model set deployed on edge computing nodes is obtained. This set includes multiple feature extraction models and feature recognition models. The feature extraction models extract features from different types of modal data. These models are obtained by training the corresponding initial deep learning models using the gradient vibration particle swarm optimization algorithm on the corresponding sample modal data. The sample modal data includes first-mode data under normal cable operation and second-mode data under external force damage. The feature recognition model is trained based on sample fusion data. The gradient vibration particle swarm optimization algorithm includes a social learning mechanism, a neighborhood gradient search mechanism, and an information fusion vibration search mechanism. The social learning mechanism controls the balance between global search and local exploitation of the particle swarm, while the neighborhood gradient search mechanism is used to find the optimal particle swarm in the current optimal region. The algorithm employs a neighborhood search to improve optimization accuracy, and an information fusion vibration search mechanism to enable the particle swarm to escape local optima and explore more hyperparameter spaces. In each iteration of the gradient vibration particle swarm algorithm, the social learning mechanism, neighborhood gradient search mechanism, and information fusion vibration search mechanism are executed sequentially to complete a full update of the particle swarm until the preset maximum number of iterations is reached. The hyperparameters corresponding to the current optimal particle are used as the target hyperparameters of the corresponding deep learning model, and each particle in the swarm represents a set of candidate hyperparameters. Real-time modal data corresponding to various modal types are determined; the corresponding feature extraction model is scheduled to extract features from the real-time modal data, resulting in multiple real-time modal features; these features are fused to obtain the fused feature to be identified; and the feature recognition model is scheduled to identify the fused feature to obtain the cable external force damage identification result. This paper employs a gradient vibration particle swarm optimization (PSO) algorithm to train multiple feature extraction models for hyperparameter optimization. In each iteration of the algorithm, a social learning mechanism balances the global search and local exploration of the particle swarm to traverse the hyperparameter space; a neighborhood gradient search mechanism performs a fine-grained search within the neighborhood of the current optimal particle to verify the existence of a better solution; and an information fusion vibration search mechanism allows the particle swarm to escape local optima and explore further regions. These three mechanisms are executed sequentially and complement each other in each iteration, jointly completing a full update of the particle swarm. This allows the particle swarm to gradually approach the globally optimal hyperparameters over multiple iterations, giving the model stronger capabilities. This achieves the technical effect of extracting the most representative modal features from real-time modal data using the optimal feature extraction model, thereby improving the accuracy of cable external force damage identification. This solves the technical problem of low accuracy in cable external force damage monitoring in related technologies.

[0057] It should be noted that the feature recognition model can also be obtained by using the gradient vibration particle swarm optimization algorithm described above to train the corresponding initial deep learning model after hyperparameter optimization.

[0058] As an optional implementation, a social learning mechanism is used to update the position of the particle swarm through adaptive weights and sine and cosine operators to maintain population diversity. Maintaining population diversity is used to prevent the particle swarm from clustering to a local optimum less than a predetermined number of times during the optimization process, so as to maintain global exploration capability. A neighborhood gradient search mechanism is used to guide particles to search within the neighborhood to approach the local optimum through pseudo-gradient information. Pseudo-gradient information is an alternative gradient vector constructed using the positional difference between the optimal particle and random particles. The neighborhood is the hyperparameter space region within a predetermined distance range of the current optimal particle. The local optimum is the hyperparameter combination that minimizes the loss function value of the deep learning model within the neighborhood. The loss function value is a quantification of the degree of difference between the predicted output and the true output of the deep learning model on the training data. An information fusion vibration search mechanism is used to generate vibration perturbations through the cooperative information between particles to help particles jump out of the local optimum.

[0059] Among them, the social learning mechanism refers to the rule in the particle swarm algorithm that particles update their own positions by learning from their own historical best positions and the group's best positions. It can control the balance between global search and local development of the particle swarm. For example, in each iteration, particles adjust their own positions according to adaptive weights and sine and cosine operators.

[0060] Among them, adaptive weight refers to the inertial weight coefficient that is dynamically adjusted with the number of training sessions. It can adjust the degree to which the particle inherits its previous motion state. For example, when the weight value is large in the early stage of training, the particle motion amplitude is large, and when the weight value is reduced in the later stage of training, the particle motion amplitude is reduced.

[0061] Among them, the sine and cosine operators refer to mathematical operators that introduce sine or cosine functions as fluctuation terms in the particle position update formula, which can make the particle update trajectory exhibit nonlinear fluctuations. For example, randomly selecting sine or cosine operators can cause the particle to produce fluctuations in direction within different iteration periods.

[0062] Population diversity refers to the degree of dispersion of individual particles in the hyperparameter space. It can reflect the ability of the particle swarm to cover different regions of the hyperparameter space. For example, population diversity is high when particles are dispersed in different regions, and low when particles are gathered in the same region.

[0063] Among them, the local optimum refers to the combination of hyperparameters that minimizes the loss function of the deep learning model within the neighborhood of the current best particle. It can serve as the target of the neighborhood gradient search mechanism, such as a better combination of hyperparameters found after a fine search within the neighborhood of the current best particle.

[0064] Global exploration capability refers to the ability of a particle swarm to search for new, unvisited regions in the hyperparameter space. This can help discover potentially better combinations of hyperparameters. For example, particles can explore regions far from the current best particle by moving significantly or subjecting themselves to random perturbations.

[0065] Among them, the neighborhood gradient search mechanism refers to the rule in the particle swarm algorithm that the particle performs a fine search in the local hyperparameter space region near the current best particle. It can improve the accuracy of local optimization. For example, the particle finely adjusts its position along the pseudo gradient direction in the neighborhood of the current best particle to verify whether there is a better solution.

[0066] Among them, pseudo gradient information refers to an alternative gradient vector constructed using the positional difference between the optimal particle and the random particle. It can indicate the search direction when the real gradient cannot be obtained. For example, the reference direction of the local search can be determined by the difference vector between the optimal particle and the random particle.

[0067] The alternative gradient vector refers to an approximate gradient vector constructed using the positional differences between particles rather than the partial derivatives of the loss function with respect to hyperparameters. It can provide search direction guidance without relying on backpropagation. For example, the positional difference between the optimal particle and the random particle can be used as the alternative gradient vector.

[0068] The neighborhood refers to the hyperparameter space region within a predetermined distance range of the current best particle, which can limit the search range of the neighborhood gradient search mechanism. For example, it is the region centered on the current best particle where the Euclidean distance is less than a set threshold.

[0069] The hyperparameter space region refers to a subset of the hyperparameter space with a specific boundary range, which can serve as a constraint domain for local search. For example, a spherical region with a predetermined radius around the current optimal particle.

[0070] The loss function value refers to the quantitative value of the difference between the predicted output and the actual output of the deep learning model on the training data. It can be used as an indicator to measure the performance of the model. For example, the smaller the cross-entropy loss function value, the higher the classification accuracy of the model.

[0071] The predicted output refers to the output result calculated by the deep learning model after forward propagating the input data. It can be compared with the actual output to calculate the error. For example, the predicted state label of the model output of the vibration time-frequency diagram.

[0072] The true output refers to the known and correct output result corresponding to the input data, which can be used as a supervision target for model training. For example, the true cable state label corresponding to the vibration time-frequency diagram.

[0073] Among them, the quantified value refers to the transformation of qualitative differences into comparable numerical results through mathematical calculations, which can provide a quantitative basis for model performance evaluation. For example, the difference between the model prediction and the true label can be quantified into a loss value of 0.5.

[0074] Among them, the information fusion vibration search mechanism refers to the rule in the particle swarm algorithm that particles update their own positions by cooperating with other particles and generating vibration perturbations. This enables particles trapped in a local optimum to jump out of the current region to explore a more distant hyperparameter space. For example, particles generate vibration amplitude based on the distance between themselves and randomly matched cooperative particles.

[0075] Among them, the cooperative information refers to the information provided by the relative positional relationship between the current particle and the randomly matched cooperative particles, which can be used as a reference for the amplitude of vibration search. For example, the greater the distance between two particles, the greater the vibration amplitude.

[0076] Vibrational disturbance refers to a random offset with oscillatory characteristics applied to the particle's position, which can cause the particle's position to change non-monotonicly and jump out of the local optimum region. For example, a displacement in the form of a cosine wave is superimposed on the particle's current position.

[0077] In this embodiment, the social learning mechanism, neighborhood gradient search mechanism, and information fusion vibration search mechanism included in the gradient vibration particle swarm optimization algorithm are functionally defined. The social learning mechanism maintains population diversity through adaptive weights and sine / cosine operators to sustain global exploration capabilities; the neighborhood gradient search mechanism guides particles to perform a refined search within their neighborhood using pseudo-gradient information to approach local optima; and the information fusion vibration search mechanism generates vibrational perturbations through cooperative information to help particles escape local optima. In this way, the three mechanisms form a clear division of labor and complementarity at the functional level: the social learning mechanism focuses on global-scale population guidance, preventing the particle swarm from prematurely converging to local optima in the early stages of the search, thus ensuring sufficient coverage of the entire hyperparameter space; the neighborhood gradient search mechanism focuses on refined verification at the local scale, deeply exploring promising regions to improve the quality of the final solution; and the information fusion vibration search mechanism focuses on ensuring the escape mechanism, providing additional exploration momentum when the particle swarm may stagnate, thus preventing premature convergence of the algorithm. The synergy of these three elements enhances the hyperparameter optimization process simultaneously in three dimensions: global exploration capability, local search accuracy, and the ability to escape local optima. Compared to algorithms that rely on only one or two mechanisms, this approach effectively avoids missing better regions due to insufficient global search, failing to find better solutions in the neighborhood due to coarse local search, or prematurely stalling due to getting stuck in local optima. As a result, after multiple iterations, a hyperparameter combination that is closer to the global optimum is obtained, thereby enabling the deep learning model to achieve better performance on the training data.

[0078] As an optional embodiment, before obtaining the multimodal recognition model set deployed on the edge computing node, the method further includes: for any feature extraction model, using the corresponding sample modal data as the expected input data and the corresponding training label as the expected output, training the corresponding initial deep learning model using the gradient vibration particle swarm algorithm to obtain the corresponding feature extraction model.

[0079] Among them, sample modal data refers to modal data samples with known state labels used to train the feature extraction model, which can provide supervision signals for model training. For example, the vibration time-frequency diagram of a cable under normal operating conditions and the vibration time-frequency diagram under external force damage conditions.

[0080] The expected input data refers to the data provided to the model as input when training the deep learning model, which can drive the forward propagation calculation of the model. For example, inputting a vibration time-frequency map into a convolutional neural network.

[0081] The expected output data refers to the target output data used as a supervision signal when training a deep learning model. It can be used to calculate the loss value with the model's predicted output to guide the update of model parameters. For example, the cable state label corresponding to the input vibration time-frequency diagram.

[0082] In this embodiment, it is clarified that before obtaining the multimodal recognition model set deployed on edge computing nodes, any feature extraction model needs to be trained first. Specifically, using the corresponding sample modality data as the expected input and the corresponding training label as the expected output, the gradient vibration particle swarm optimization algorithm is used to optimize the hyperparameters of the initial deep learning model, resulting in a trained feature extraction model. This approach establishes a complete process from data preparation to model training completion, enabling each feature extraction model to be adapted and optimized for the distribution characteristics of its respective modality data, ensuring the model's relevance and effectiveness. Simultaneously, by employing the gradient vibration particle swarm optimization algorithm for hyperparameter optimization, the inefficiencies and susceptibility to local optima inherent in methods relying on manual experience or grid search are avoided. Furthermore, since each feature extraction model deployed on the edge computing nodes has been trained in this way, when real-time data arrives, each model can stably extract high-quality discriminative features from the real-time data of its corresponding modality, providing a feature foundation for accurate final recognition.

[0083] As an optional embodiment, using the corresponding sample modal data as the expected input data and the corresponding training label as the expected output, the gradient vibration particle swarm optimization algorithm is used to train the corresponding initial deep learning model to obtain the corresponding feature extraction model. This includes: determining the initial particle swarm based on the initial hyperparameters of the initial deep learning model; for each particle in the initial particle swarm, determining the fitness of each particle using the sample modal data as the expected input data and the training label as the expected output; determining the historical best position of each particle and the current best particle in the particle swarm during the current training process based on the fitness of each particle; training the corresponding initial deep learning model using the gradient vibration particle swarm optimization algorithm based on the historical best position of each particle and the current best particle in the particle swarm to output the globally optimal particle; and using the hyperparameters of the globally optimal particle as the target hyperparameters of the initial deep learning model to obtain the trained feature extraction model.

[0084] The initial hyperparameters refer to the initial values ​​of the hyperparameters set before the start of training of the initial deep learning model. They can serve as the center of the reference range for the values ​​of each particle when the particle swarm is initialized. For example, the initial value of the learning rate is set to 0.01.

[0085] The initial particle swarm refers to the set of initial particle positions set according to the hyperparameters of the initial deep learning model at the start of hyperparameter optimization. It can serve as the starting point set for the algorithm's iterative search. For example, 50 particles can be randomly generated near the initial hyperparameter values ​​to form the initial particle swarm.

[0086] In this context, a particle refers to the basic unit in a particle swarm, representing a set of candidate hyperparameter combinations. It can search for the optimal solution by moving in the hyperparameter space. For example, a particle carries a hyperparameter combination with a learning rate of 0.008 and a batch size of 64.

[0087] Fitness is a quantitative indicator that measures the performance of a deep learning model by the combination of hyperparameters represented by a particle. It can be used as a basis for evaluating the quality of a particle. For example, a higher fitness value means that the set of hyperparameters makes the model more accurate.

[0088] Among them, the historical best position refers to the position with the highest fitness reached by each particle in each iteration. It can record the best search results of the particle. For example, the position found by a particle in the 10th iteration makes the model accuracy reach 92%.

[0089] The current optimal particle refers to the particle with the highest fitness in the entire particle swarm at the current iteration number. It represents the currently discovered optimal combination of hyperparameters and can serve as a target for other particles to learn from and gravitate towards. For example, the hyperparameter combination corresponding to the current optimal particle makes the model achieve the highest accuracy on the validation set.

[0090] Among them, the globally optimal particle refers to the particle with the highest fitness in the entire particle swarm when the preset maximum number of iterations is reached. It represents the final result of hyperparameter optimization and can serve as the source of target hyperparameters for deep learning models. For example, it can be the hyperparameter combination that optimizes the model performance after 200 iterations.

[0091] Among them, the target hyperparameters refer to the hyperparameter values ​​of the globally optimal particle as the final hyperparameters used by the deep learning model, which enables the model to achieve the best performance after deployment. For example, the learning rate of 0.007 and the batch size of 48 obtained by optimization can be used as the final hyperparameters of the model.

[0092] In this embodiment, the specific steps of training the feature extraction model using the gradient vibration particle swarm algorithm are described in detail: First, the particle swarm is initialized based on the hyperparameters of the initial deep learning model. Then, the fitness of each particle is determined using sample modal data and training labels. In turn, the historical best position and the current best particle of each particle are determined. Finally, the gradient vibration particle swarm algorithm is used to train and output the globally optimal particle based on this information. The hyperparameters of the globally optimal particle are used as the target hyperparameters of the model. This approach leverages the parallel search capability of the particle swarm optimization (PSO) algorithm in the hyperparameter space to iteratively optimize the hyperparameters of the initial deep learning model from their initial values ​​to the target hyperparameters that optimize model performance. Specifically, the PSO is initialized to cover multiple possible values ​​near the initial hyperparameters. A quantitative mapping relationship is established between model performance and hyperparameter values ​​through fitness evaluation. Good intermediate solutions found during the search process are avoided by recording historical optimal positions. The PSO is guided by the current optimal particle to move the entire PSO towards a better region. The coordinated iteration of these three mechanisms—gradient vibration PSO—achieves a dynamic balance between global exploration and local refinement. Finally, the output of the globally optimal particle ensures the theoretical global optimality of the hyperparameter optimization results, thereby guaranteeing that the trained feature extraction model has optimal performance on the feature extraction task.

[0093] As an optional embodiment, the operation of the social learning mechanism includes: obtaining a first updated particle swarm; obtaining adaptive weights corresponding to the first updated particle swarm at the current training iteration, wherein the adaptive weights adopt an exponential decay form and are superimposed with cosine fluctuations on the decay basis; for each first updated particle in the first updated particle swarm, based on the adaptive weights, the current position of the first updated particle at the current training iteration, the position of the current best particle in the current training process, and the historical best position of the first updated particle, the updated position of the first updated particle in the next training iteration is determined using a sine and cosine algorithm; the calculated updated position is used as the position of the first updated particle after social learning, and after all the first updated particles are updated, a second updated particle swarm including multiple second updated particles is obtained.

[0094] The first update particle swarm refers to the particle swarm before the social learning mechanism is executed, that is, the set of particles that have not yet been updated by social learning in the current iteration, which can be used as the input particle swarm of the social learning mechanism, for example, the particle swarm obtained after the previous iteration.

[0095] The first update particle refers to a single particle in the first update particle swarm, representing a set of candidate hyperparameter combinations, which can serve as the basic operational unit of the social learning mechanism. For example, the i-th particle in the first update particle swarm carries a set of values ​​for the learning rate and the number of convolutional kernels.

[0096] The current position refers to the position coordinates of the first updated particle in the hyperparameter space under the current training iteration. It can be used as the starting point for calculating the update position. For example, the particle is located at the position with a learning rate of 0.008 and a convolution kernel number of 64 in the current iteration.

[0097] The position of the current best particle refers to the position coordinates of the particle with the highest fitness in the entire particle swarm under the current training iterations in the hyperparameter space. It can serve as one of the targets for particles to approach in social learning. For example, the current best particle is located at a learning rate of 0.007 and a convolution kernel count of 72.

[0098] The updated position refers to the new position of the first updated particle in the next training after being calculated by the social learning mechanism. It can be used as the output of the social learning mechanism, for example, the particle moves from its current position to a new coordinate point in the hyperparameter space.

[0099] The second update particle refers to the particle obtained after the first update particle is updated by the social learning mechanism. It represents the updated candidate hyperparameter combination and can form the basic unit of the second update particle swarm, such as the particle whose position changes after social learning update.

[0100] The second update particle swarm refers to the particle swarm obtained after all the first update particles have completed social learning updates. It can be used as the input for the next mechanism, namely the neighborhood gradient search mechanism. For example, a new particle swarm composed of multiple second update particles.

[0101] This embodiment details the specific operational steps of the social learning mechanism: First, a first updated particle swarm and corresponding adaptive weights are obtained. Then, for each first updated particle, the updated position is calculated using a sine and cosine algorithm based on the adaptive weights, current position, current best particle position, and historical best position. After all particles are updated, a second updated particle swarm is obtained. This method ensures that each particle in the swarm maintains the nonlinear fluctuation characteristics of its update trajectory while moving towards its own historical best and the swarm's best. Specifically, the exponential decay in the adaptive weights allows particles to maintain a large inertia for wide-range searching in the early stages of training and a smaller amplitude for fine-tuning in the later stages. Cosine fluctuations continuously provide periodic amplitude changes throughout the training process to prevent particle motion from falling into a monotonous mode. The random selection of sine or cosine operators in the sine and cosine algorithms causes fluctuations in the particle's movement direction within different iteration periods, resulting in a tortuous trajectory for particles that might otherwise move along a straight path, thereby increasing the frequency of the particle swarm's access to different regions of the hyperparameter space. Through the combined effect of these mechanisms, the social learning mechanism achieves the orderly movement of particle swarms in the hyperparameter space and the maintenance of population diversity at the algorithmic level, providing a reasonably distributed particle swarm that has initially approached the optimal region for the subsequent neighborhood gradient search mechanism.

[0102] As an optional embodiment, the neighborhood gradient search mechanism includes: obtaining pseudo-gradient update information corresponding to the second updated particle swarm, and using the positional difference between the current optimal particle in the current training iteration and the second updated particle randomly selected from the second updated particle swarm, combined with the step size parameter, generating a random vector with the same dimension as the hyperparameter, as the pseudo-gradient update information for the next iteration, wherein the step size parameter is determined based on the distance between the current optimal particle and the randomly selected second updated particle and the previous pseudo-gradient update information; obtaining the first neighborhood gradient search coefficient and the second neighborhood gradient search coefficient based on the current training iteration, wherein the first neighborhood gradient search coefficient is determined based on the distance between the current optimal particle and the randomly selected second updated particle and the previous pseudo-gradient update information; and obtaining the first neighborhood gradient search coefficient and the second neighborhood gradient search coefficient based on the current training iteration. Both the domain gradient search coefficient and the second neighborhood gradient search coefficient change nonlinearly with time. The first neighborhood gradient search coefficient is used to control the degree of repulsion of the second updated particle away from the worst particle, and the second neighborhood gradient search coefficient is used to control the degree of step size adjustment of the second updated particle toward the best particle. For each second updated particle in the second updated particle swarm, the position of the second updated particle is updated by neighborhood search based on the pseudo gradient update information, the first neighborhood gradient search coefficient and the second neighborhood gradient search coefficient, to obtain the position of the second updated particle after neighborhood gradient search. After all the second updated particles are updated, a third updated particle swarm including multiple third updated particles is obtained.

[0103] Among them, pseudo gradient update information refers to the alternative gradient vector calculated in the current iteration for use in the next iteration. It can provide direction guidance for local search of particles when the real gradient cannot be obtained. For example, it can be a random vector generated by combining the position difference between the optimal particle and the random particle with the step size parameter.

[0104] The step size parameter refers to the scaling factor that controls the magnitude of the pseudo-gradient update information. It can adjust the step size of the particle in the neighborhood gradient search. For example, when the step size parameter is 2.5, the particle's movement magnitude is the step size parameter multiplied by the random vector.

[0105] The hyperparameter dimension refers to the total number of hyperparameters to be optimized, which determines the dimension of the hyperparameter space and particles. For example, when optimizing three hyperparameters, namely learning rate, batch size, and number of convolutional kernels, the hyperparameter dimension is 3.

[0106] Among them, a random vector refers to a multidimensional vector whose element values ​​are randomly selected within a predetermined range. It can provide random perturbation components for pseudo gradient update information. For example, a random vector with dimension 3 takes values ​​of [0.3, 0.7, 0.5].

[0107] Among them, the first neighborhood gradient search coefficient refers to the coefficient that changes nonlinearly with time and is used to control the degree of repulsion of the second updated particle away from the worst particle. It can adjust the repulsion of the particle towards the worst particle. For example, when the coefficient is positive, the particle moves away from the worst particle.

[0108] Among them, the second neighborhood gradient search coefficient refers to the coefficient that changes nonlinearly with time and is used to control the step size adjustment of the second updated particle toward the optimal particle. It can adjust the step size of the particle toward the optimal particle. For example, when the coefficient is positive, the particle moves toward the direction of the optimal particle.

[0109] Among them, the worst particle refers to the particle with the lowest fitness in the current particle swarm. It represents the worst combination of hyperparameters that has been found so far and can act as a repulsion term to guide other particles away. For example, the hyperparameter combination corresponding to the particle with the lowest fitness currently results in the lowest model accuracy.

[0110] Among them, the degree of repulsion refers to the strength with which a particle moves away from the worst particle during the renewal process. It can help the particle avoid the hyperparameter region with poor performance. For example, the greater the degree of repulsion, the greater the distance the particle moves away from the worst particle.

[0111] Among them, the step size adjustment degree refers to the adjustment range of the step size by which the particle approaches the optimal particle during the update process. It can control the speed at which the particle approaches the optimal particle. For example, the larger the step size adjustment degree, the greater the magnitude of the particle's approach to the optimal particle.

[0112] The third update particle refers to the particle obtained after the second update particle is updated by the neighborhood gradient search mechanism. It represents the candidate hyperparameter combination after local fine search and can form the basic unit of the third update particle swarm, such as the particle after fine-tuning its position in the neighborhood of the optimal particle.

[0113] Among them, the third update particle swarm refers to the particle swarm obtained after all the second update particles have completed the neighborhood gradient search update. It can be used as the input of the next mechanism, namely the information fusion vibration search mechanism. For example, a new particle swarm composed of multiple third update particles.

[0114] This embodiment details the specific operational steps of the neighborhood gradient search mechanism: First, obtain the pseudo-gradient update information corresponding to the second updated particle swarm and calculate the pseudo-gradient update information for the next iteration. Then, obtain the first and second neighborhood gradient search coefficients. Finally, update the position of each second updated particle using the pseudo-gradient update information and the two coefficients. After all particles are updated, a third updated particle swarm is obtained. In this way, particles perform a fine search within the local region near the currently discovered optimal particle. Specifically, the pseudo-gradient update information uses the positional difference between the optimal particle and random particles to construct a substitute gradient vector, allowing particles to obtain a local search direction reference even without a real gradient. The step size parameter is dynamically adjusted based on the distance between the optimal particle and random particles, enabling the search step size to adapt to the local spatial scale of different iteration stages. The first neighborhood gradient search coefficient changes non-linearly with time, guiding particles to gradually increase their tendency to move away from the worst particle during iteration, thereby avoiding regions with poor performance hyperparameters. The second neighborhood gradient search coefficient changes non-linearly with time, adjusting the step size at which particles approach the optimal particle, allowing particles to gradually approach the local optimal position during the neighborhood search process. Through the synergistic effect of these coefficients, the particle swarm completed a systematic and refined search within the optimal particle neighborhood, verifying whether there is a better combination of hyperparameters within the optimal particle neighborhood, thereby improving the local accuracy of hyperparameter optimization.

[0115] As an optional embodiment, the operation of the information fusion vibration search mechanism includes: randomly matching one other third update particle as a cooperative particle for each third update particle in the third update particle swarm; for each third update particle, determining the cooperative information factor of the third update particle based on the distance between the corresponding third update particle and the corresponding cooperative particle and the current training number, wherein the cooperative information factor is used to control the amplitude of the vibration search; for each third update particle, performing vibration search update on the position of the third update particle based on the cooperative information factor and the current best particle, obtaining the position of the third update particle after vibration search, and obtaining the fourth update particle swarm after vibration search after all third update particles are updated, thereby completing one complete iteration update of the particle swarm.

[0116] In this context, a cooperative particle refers to another third-update particle randomly matched for each third-update particle in the third-update particle swarm. This cooperative particle can provide cooperative information for the current particle to determine the vibration amplitude. For example, the n+5th particle can be randomly matched as a cooperative particle for the nth particle.

[0117] Among them, the cooperation information factor refers to a numerical factor determined based on the distance between each third update particle and its corresponding cooperative particle and the current number of training iterations. It can control the magnitude of the vibration search. For example, the cooperation information factor has a larger value when the distance between the particle and the cooperative particle is greater.

[0118] Among them, vibration search amplitude refers to the magnitude of the particle's position offset during the vibration search process, which can determine the distance the particle jumps out of the current region. For example, when the vibration amplitude is large, the particle can move to a more distant hyperparameter space region.

[0119] Vibrational search update refers to the update operation that applies vibrational perturbation to the position of the third update particle based on the cooperative information factor and the current best particle. This can cause the particle position to undergo oscillatory changes, for example, by superimposing an oscillatory displacement related to the cooperative information factor on the current position.

[0120] The fourth update particle swarm refers to the particle swarm obtained after all the third update particles have completed the vibrational search update, which can be used as the initial particle swarm for the next iteration. For example, a new particle swarm composed of multiple fourth update particles.

[0121] This embodiment details the specific operational steps of the information fusion vibration search mechanism: First, a cooperative particle is randomly matched for each third-update particle in the third-update particle swarm. Then, a cooperative information factor is determined based on the distance between each particle and the cooperative particle and the current training iteration. Finally, the position of each particle is updated using vibration search based on the cooperative information factor and the current optimal particle. After all particles are updated, a fourth-update particle swarm is obtained, completing one full iteration update of the particle swarm. In this way, after completing the neighborhood gradient search, particles introduce cooperative information between the swarm by randomly matching cooperative particles, and the vibration amplitude is controlled by the cooperative information factor. Specifically, the cooperative information factor considers the distance between the current particle and the cooperative particle, as well as the current training iteration. In the early stages of training, when the distance between particles is large, the cooperative information factor value is large, resulting in significant vibrations to help particles jump out of their current region and explore other regions further away in the hyperparameter space. In the later stages of training, as the number of training iterations increases, the cooperative information factor value gradually decreases, causing the vibration amplitude to become smoother. When the particle swarm is close to the optimal region overall, it no longer generates excessively large disturbances, thus achieving stable convergence. Meanwhile, the oscillatory displacement introduced during the vibrational search update process causes non-monotonic changes in particle positions, effectively preventing particles from stagnating in local optima due to excessively small step sizes in the later stages of the search. This mechanism, acting as the final escape mechanism within the framework of the three sequentially executed mechanisms, enables the particle swarm to break through the limitations of local optima after completing the global movement of social learning and the local refinement of neighborhood gradient search. Thus, the three mechanisms together constitute a complete iterative logic from global exploration to local refinement and then to escaping local optima, ensuring that the particle swarm can approach the globally optimal hyperparameter combination after multiple iterations.

[0122] Based on the above embodiments and optional embodiments, an optional implementation method is provided, which is described in detail below.

[0123] An optional embodiment of the present invention provides a cable external force damage monitoring method based on edge computing, which solves the problems of high false alarm rate of single-mode monitoring, large cloud processing delay and / or insufficient model feature extraction capability in the prior art. Figure 2 A flowchart of a cable external force damage monitoring method based on edge computing is provided as an optional embodiment of the present invention. Figure 3 A flowchart illustrating the training of the first deep learning model using the gradient vibration particle swarm optimization algorithm, provided as an optional embodiment of the present invention, is described below:

[0124] Acquire multimodal data of the cable under normal operating conditions and multimodal data of the cable under external force damage conditions. Use the multimodal data as training samples and construct training labels from the corresponding states to obtain training data composed of training samples and training labels.

[0125] Preprocess the multimodal data corresponding to the training samples to obtain the first target modal data and the second target modal data corresponding to the training samples;

[0126] Based on the first target modal data, the second target modal data, and training labels, the gradient vibration particle swarm algorithm is used to train the first deep learning model and the second deep learning model to obtain the first feature extraction model and the second feature extraction model.

[0127] A third deep learning model is trained using the first target modality data, the second target modality data, training labels, the first feature extraction model, and the second feature extraction model to obtain a feature recognition model;

[0128] The first feature extraction model, the second feature extraction model, and the feature recognition model are deployed on edge computing nodes, and real-time multimodal data are collected through the edge computing nodes;

[0129] Real-time multimodal data is preprocessed by edge computing nodes to obtain first real-time modal data and second real-time modal data. A first feature extraction model is scheduled to extract features from the first real-time modal data and a second feature extraction model is scheduled to extract features from the second real-time modal data. The extracted features are then fused to obtain the fused features to be identified.

[0130] The edge computing node scheduling feature recognition model identifies the features to be identified and fused, obtains the cable external force damage identification results, and issues early warnings based on the cable external force damage identification results.

[0131] In one possible design approach, acquiring multimodal data of the cable under normal operating conditions and multimodal data of the cable under conditions of external force damage includes:

[0132] Under normal cable operation conditions, multimodal data of the cable's normal operation conditions are obtained by synchronously collecting cable vibration signals and cable image data through edge monitoring terminals deployed around the cable channel;

[0133] When the cable is damaged by external force, the cable vibration signal and cable image data are collected synchronously by the edge monitoring terminals deployed around the cable channel to obtain multimodal data of the cable under the condition of external force damage.

[0134] In one possible design approach, the multimodal data corresponding to the training samples are preprocessed to obtain the first target modality data and the second target modality data corresponding to the training samples, including:

[0135] The cable vibration signal in the training sample is denoised and converted into a time-frequency diagram to obtain the first target mode data corresponding to the training sample.

[0136] The cable image data in the training samples are enhanced and / or normalized to obtain the second target modality data corresponding to the training samples.

[0137] In one possible design approach, based on the first target modal data, the second target modal data, and training labels, a first deep learning model and a second deep learning model are trained using the gradient vibration particle swarm optimization algorithm to obtain a first feature extraction model and a second feature extraction model, including:

[0138] Using the first target modal data as the desired input data and the training labels as the desired output, the first deep learning model is trained using the gradient vibration particle swarm algorithm to obtain the first feature extraction model.

[0139] Using the second target modal data as the desired input data and the training labels as the desired output, the second deep learning model is trained using the gradient vibration particle swarm algorithm to obtain the second feature extraction model.

[0140] In one possible design approach, the first target modal data is used as the desired input data, and the training labels are used as the desired output. A first deep learning model is trained using the gradient vibration particle swarm optimization algorithm to obtain a first feature extraction model, including:

[0141] The particle swarm is initialized based on the hyperparameters of the first deep learning model;

[0142] For each particle in the particle swarm, the fitness of the particle is determined by using the first target mode data as the expected input data and the training label as the expected output.

[0143] Based on particle fitness, determine the historical best value of the particle and the best particle in the current training process;

[0144] Based on the historical best value and the best particle, an improved social learning mechanism combining adaptive weights and sine and cosine algorithms is used to perform social learning on the particle swarm, resulting in a socially learned particle swarm.

[0145] A gradient search mechanism is used to perform neighborhood gradient search on the particle swarm after social learning, resulting in a particle swarm after neighborhood gradient search.

[0146] An information fusion vibration mechanism is used to perform vibration search on the particle swarm after neighborhood gradient search, resulting in a particle swarm after vibration search.

[0147] Based on the particle swarm after vibration search, proceed to the next training, and repeat the training multiple times to output the globally optimal particle.

[0148] The hyperparameters in the globally optimal particle are used as the target hyperparameters of the first deep learning model to obtain the first feature extraction model.

[0149] In one possible design approach, based on historical best values ​​and the best particle, an improved social learning mechanism combining adaptive weights and sine / cosine algorithms is used to perform social learning on the particle swarm, resulting in a socially learned particle swarm, including:

[0150] The adaptive weights are obtained as follows:

[0151] ;

[0152] in, Indicates adaptive weights, Indicates the current number of training iterations. This represents the maximum value of the adaptive weights, and is set to 0.9; This represents the minimum value of the adaptive weights, and is set to 0.1; Represents the natural constant. This indicates the overall decay rate, and can be set to 3, 4, or 5. This indicates the preset maximum number of training iterations. This represents the fluctuation intensity coefficient, and is set to 0.1 or 0.2. Represents pi (π). This indicates the fluctuation period and is set to a value between [20, 50].

[0153] Based on adaptive weights, historical best values, and the best particle, a sine and cosine algorithm is used to perform social learning on the particle swarm, resulting in the following particle swarm after social learning:

[0154] ;

[0155] in, Let d represent the d-th dimension of the hyperparameter of the i-th particle in the particle swarm, where d = 1, 2, ..., Dim, and Dim represents the total dimension of the hyperparameter. Let d represent the d-th dimension hyperparameter of the i-th particle in the particle swarm after social learning. Represents a random number between (0, 1). Represents a random number between (0, 1). Represents a random number between (0, 1). Represents a random number between (0, 1). Denotes the d-th dimension hyperparameter of the optimal particle. Let d be the d-th dimension hyperparameter representing the historical best value of the i-th particle.

[0156] In one possible design approach, a gradient search mechanism is used to perform a neighborhood gradient search on the particle swarm after social learning, resulting in a particle swarm that includes:

[0157] For particles in the particle swarm after social learning, the pseudo gradient update information is obtained as follows:

[0158] ;

[0159] ;

[0160] in, This represents the pseudo-gradient update information during the t-th training process. This represents the pseudo-update information during the (t+1)th training iteration. Represents the interval [0,1] 3D random vector, Represents the optimal particle. This represents particles after random social learning. Indicates the step size parameter;

[0161] Based on the current number of training iterations, the gradient search coefficients for the first and second neighborhoods are obtained as follows:

[0162] ;

[0163] ;

[0164] ;

[0165] ;

[0166] in, Represents the gradient search coefficients in the first neighborhood. Indicates the gradient search coefficient of the second neighborhood. Represents a random number between (0, 1). Represents a random number between (0, 1). This represents the base value of the neighborhood gradient search coefficient. Indicates the coefficient of time variation. Represents pi (π). This represents the maximum value of the time variation coefficient, and is set to 1. This represents the minimum value of the time variation coefficient, and is set to 0.2;

[0167] Based on the pseudo-gradient update information, the first neighborhood gradient search coefficients, and the second neighborhood gradient search coefficients, a neighborhood gradient search is performed, resulting in the following particle swarm after the neighborhood gradient search:

[0168] ;

[0169] in, Let represent the d-th dimension hyperparameter of the particle after the m-th social learning iteration. Let represent the particle after the gradient search of the m-th neighborhood, and all particles after the gradient search of their neighborhoods constitute the particle swarm after the gradient search of their neighborhoods. This represents a random number that follows a standard normal distribution. Denotes the d-th dimension hyperparameter of the worst-case particle. The d-th parameter represents the pseudo-gradient update information. Denotes the d-th dimension hyperparameter of the optimal particle. This represents a constant term and is set to 0.01. This represents a random number between (0, 1).

[0170] In one possible design approach, an information fusion vibration mechanism is used to perform a vibration search on the particle swarm after the neighborhood gradient search, resulting in a particle swarm after the vibration search, including:

[0171] Based on the particle swarm after neighborhood gradient search, a random other particle is matched for each particle after neighborhood gradient search to obtain the cooperative particle corresponding to the particle after neighborhood gradient search.

[0172] Based on the cooperative particles corresponding to the particles after neighborhood gradient search, the cooperative information factor is obtained as follows:

[0173] ;

[0174] in, Indicates collaborative information factors. This represents the particle after the gradient search of the nth neighborhood. express The corresponding cooperative particles, Indicates the number of vibration searches;

[0175] Based on the cooperative information factor and the optimal particle, a vibrational search is performed on the particle swarm after the neighborhood gradient search, resulting in the following particle swarm after the vibrational search:

[0176] ;

[0177] in, Let d be the hyperparameter of the particle after the gradient search of the nth neighborhood. Let d represent the d-th dimension hyperparameter of the particle after the nth vibrational search, and all particles after the vibrational search constitute the particle swarm after the vibrational search. This represents the vibration search control coefficient. The sequence number indicates the number of vibration search attempts.

[0178] In one possible design approach, the second target modal data is used as the desired input data, and the training labels are used as the desired output. A gradient vibration particle swarm optimization algorithm is employed to train a second deep learning model, resulting in a second feature extraction model, including:

[0179] The particle swarm is initialized based on the hyperparameters of the second deep learning model;

[0180] For each particle in the particle swarm, the fitness of the particle is determined by using the second target modality data as the expected input data and the training label as the expected output.

[0181] Based on particle fitness, determine the historical best value of the particle and the best particle in the current training process;

[0182] Based on the historical best value and the best particle, an improved social learning mechanism combining adaptive weights and sine and cosine algorithms is used to perform social learning on the particle swarm, resulting in a socially learned particle swarm.

[0183] A gradient search mechanism is used to perform neighborhood gradient search on the particle swarm after social learning, resulting in a particle swarm after neighborhood gradient search.

[0184] An information fusion vibration mechanism is used to perform vibration search on the particle swarm after neighborhood gradient search, resulting in a particle swarm after vibration search.

[0185] Based on the particle swarm after vibration search, proceed to the next training, and repeat the training multiple times to output the globally optimal particle.

[0186] The hyperparameters in the globally optimal particle are used as the target hyperparameters of the second deep learning model to obtain the second feature extraction model.

[0187] In one possible design approach, a third deep learning model is trained using first target modality data, second target modality data, training labels, a first feature extraction model, and a second feature extraction model to obtain a feature recognition model, including:

[0188] The data features of the first target modality data are extracted using a first feature extraction model and the data features of the second target modality data are extracted using a second feature extraction model, and then fused to obtain fused data features;

[0189] The particle swarm is initialized based on the hyperparameters of the third deep learning model;

[0190] For each particle in the particle swarm, the fitness of the particle is determined by using the fused data features as the expected input data and the training labels as the expected output.

[0191] Based on particle fitness, determine the historical best value of the particle and the best particle in the current training process;

[0192] Based on the historical best value and the best particle, an improved social learning mechanism combining adaptive weights and sine and cosine algorithms is used to perform social learning on the particle swarm, resulting in a socially learned particle swarm.

[0193] A gradient search mechanism is used to perform neighborhood gradient search on the particle swarm after social learning, resulting in a particle swarm after neighborhood gradient search.

[0194] An information fusion vibration mechanism is used to perform vibration search on the particle swarm after neighborhood gradient search, resulting in a particle swarm after vibration search.

[0195] Based on the particle swarm after vibration search, proceed to the next training, and repeat the training multiple times to output the globally optimal particle.

[0196] The hyperparameters in the globally optimal particle are used as the target hyperparameters of the third deep learning model to obtain the feature recognition model.

[0197] This invention provides a cable external force damage monitoring method based on edge computing. It constructs a training set by collecting two types of multimodal data: vibration signals and image data under normal and external force damage conditions of the cable. After preprocessing, time-frequency maps and enhanced images are obtained as target modal data. A gradient vibration particle swarm optimization algorithm is used to train a convolutional neural network to obtain two types of feature extraction models, and then a feature recognition model is trained. The above models are deployed on edge computing nodes to collect and preprocess multimodal data in real time. After feature extraction and fusion, the data is input into the recognition model, outputting damage recognition results and triggering early warnings. This invention achieves decoupling of modal feature extraction and recognition, adaptively optimizes the decision boundary, improves the complementarity of multimodal features, and enhances monitoring accuracy. It is suitable for real-time monitoring scenarios of cable channels.

[0198] Specifically, it will be introduced below:

[0199] S1. Obtain multimodal data of the cable under normal operating conditions and multimodal data of the cable under external force damage conditions. Use the multimodal data as training samples and construct training labels for the corresponding states to obtain training data composed of training samples and training labels.

[0200] For example, edge monitoring terminals are deployed around a 10kV underground cable tunnel in a certain city. Under normal cable operation, the terminals simultaneously collect vibration signals and on-site image data when the cable is not subjected to external force damage; under simulated external force damage (such as excavator digging or electric drilling), they simultaneously collect vibration signals and image data when the cable is subjected to external force damage. These multimodal data are used as training samples, and the corresponding normal or damaged states are used as training labels. It is worth noting that this mainly simulates the damaged state, which can be simulated in a laboratory or real-world scenario before online monitoring, and training datasets can be collected.

[0201] S2. Preprocess the multimodal data corresponding to the training samples to obtain the first target modal data and the second target modal data corresponding to the training samples;

[0202] The vibration signal in the training sample can be denoised using wavelet thresholding, and the denoised one-dimensional vibration signal can be converted into a two-dimensional time-frequency map through continuous wavelet transform, which serves as the first target modal data; the image data in the training sample can be enhanced by histogram equalization and the pixel values ​​can be normalized to the [0,1] interval, which serves as the second target modal data.

[0203] S3. Based on the first target modal data, the second target modal data, and the training labels, the first deep learning model and the second deep learning model are trained using the gradient vibration particle swarm algorithm to obtain the first feature extraction model and the second feature extraction model.

[0204] Construct a first deep learning model (e.g., CNN, convolutional neural network) and a second deep learning model (e.g., CNN, convolutional neural network). Using the first target modality data as input and training labels as output, optimize the hyperparameters (e.g., learning rate, batch size, number of convolutional kernels) of the first deep learning model using the gradient vibrational particle swarm optimization algorithm to obtain the first feature extraction model. Similarly, using the second target modality data as input, optimize the second deep learning model using the same algorithm to obtain the second feature extraction model.

[0205] S4. Using the first target modality data, the second target modality data, training labels, the first feature extraction model, and the second feature extraction model, train the third deep learning model to obtain the feature recognition model;

[0206] The network parameters of the first and second feature extraction models can be fixed, and the data features of the first and second target modal data can be extracted and fused. Using the fused features as input and the training labels as output, a third deep learning model (such as a fully connected network classifier, a backpropagation neural network, or a long short-term memory network) can be constructed. The hyperparameters of the model are then optimized again using the gradient vibration particle swarm optimization algorithm to obtain the feature recognition model.

[0207] S5. Deploy the first feature extraction model, the second feature extraction model, and the feature recognition model in the edge computing node, and collect real-time multimodal data through the edge computing node;

[0208] S6. Preprocess the real-time multimodal data through edge computing nodes to obtain the first real-time modal data and the second real-time modal data, and schedule the first feature extraction model to extract features from the first real-time modal data and the second feature extraction model to extract features from the second real-time modal data, and fuse the extracted features to obtain the fused features to be identified.

[0209] S7. The feature to be identified is identified by the edge computing node scheduling feature identification model to obtain the cable external force damage identification result, and an early warning is given based on the cable external force damage identification result.

[0210] The first feature extraction model, the second feature extraction model, and the feature recognition model are packaged and deployed to edge computing nodes at the cable tunnel site. The edge computing nodes collect vibration and image data in real time, preprocess the data, and then feed it into the corresponding feature extraction model. The extracted features are fused and fed into the feature recognition model, which outputs the external force damage identification result. If the result is "damage," the edge node immediately sends an early warning signal to the monitoring center via the 5G network.

[0211] This invention overcomes the shortcomings of single-modal monitoring, which is susceptible to false alarms due to environmental interference, by collecting and fusing multimodal data, thus improving the robustness and accuracy of identifying external force damage under complex working conditions. Deploying the feature extraction and recognition models on edge computing nodes enables local data processing and on-site decision-making, significantly reducing network bandwidth consumption for data transmission back to the cloud and substantially shortening the response delay from anomaly occurrence to early warning. The gradient vibration particle swarm optimization algorithm is used to optimize and train the deep learning model, effectively avoiding the problem of traditional gradient descent methods easily getting trapped in local optima, and improving the accuracy of model feature extraction and state recognition.

[0212] In one possible design approach, acquiring multimodal data of the cable under normal operating conditions and multimodal data of the cable under conditions of external force damage includes:

[0213] Under normal cable operation conditions, multimodal data of the cable's normal operation conditions are obtained by synchronously collecting cable vibration signals and cable image data through edge monitoring terminals deployed around the cable channel;

[0214] When the cable is damaged by external force, the cable vibration signal and cable image data are collected synchronously by the edge monitoring terminals deployed around the cable channel to obtain multimodal data of the cable under the condition of external force damage.

[0215] Vibration signals are highly sensitive to stress waves from mechanical construction, and image data can intuitively reflect the presence and operating status of construction machinery on site. The synchronous acquisition of both in the time domain provides a spatiotemporally aligned data foundation for subsequent feature fusion, ensuring the completeness of the state representation.

[0216] In one possible design approach, the multimodal data corresponding to the training samples are preprocessed to obtain the first target modality data and the second target modality data corresponding to the training samples, including:

[0217] The cable vibration signal in the training sample is denoised and converted into a time-frequency diagram to obtain the first target mode data corresponding to the training sample.

[0218] The cable image data in the training samples are enhanced and / or normalized to obtain the second target modality data corresponding to the training samples.

[0219] Converting one-dimensional vibration signals into two-dimensional time-frequency maps fully preserves the joint distribution characteristics of vibration in both the time and frequency domains, facilitating subsequent deep feature mining using convolutional neural networks. Enhancing and normalizing image data eliminates interference from illumination variations and noise on image features, unifies the data distribution scale, and accelerates model convergence.

[0220] It is worth noting that, in addition to the preprocessing methods mentioned above, other preprocessing methods can also be used to preprocess the data to make it easier for deep learning models to recognize the data.

[0221] In one possible design approach, based on the first target modal data, the second target modal data, and training labels, a first deep learning model and a second deep learning model are trained using the gradient vibration particle swarm optimization algorithm to obtain a first feature extraction model and a second feature extraction model, including:

[0222] Using the first target modal data as the desired input data and the training labels as the desired output, the first deep learning model is trained using the gradient vibration particle swarm algorithm to obtain the first feature extraction model.

[0223] Using the second target modal data as the desired input data and the training labels as the desired output, the second deep learning model is trained using the gradient vibration particle swarm algorithm to obtain the second feature extraction model.

[0224] To address the heterogeneous characteristics of different modal data, we construct and train dedicated feature extraction models for each modality independently. This enables each model to specifically mine deep discriminative features within a single modality, avoiding the feature dilution problem caused by directly mixing heterogeneous data.

[0225] like Figure 3 As shown, using the first target modal data as the desired input data and the training labels as the desired output, the first deep learning model is trained using the gradient vibration particle swarm optimization algorithm to obtain the first feature extraction model, which includes:

[0226] S201. Initialize the particle swarm based on the hyperparameters of the first deep learning model;

[0227] S202. For each particle in the particle swarm, the fitness of the particle is determined by using the first target mode data as the expected input data and the training label as the expected output.

[0228] For example, by using the first target modality data as the expected input data and the training labels as the expected output, the root mean square loss function value is obtained. The fitness of the particle can be obtained by adding the root mean square loss function value to a preset minimum constant term (such as 0.01) and taking the reciprocal.

[0229] S203. Based on the particle fitness, determine the historical best value of the particle and the best particle in the current training process;

[0230] S204. Based on the historical best value and the best particle, an improved social learning mechanism combining adaptive weights and sine and cosine algorithms is used to perform social learning on the particle swarm to obtain the socially learned particle swarm.

[0231] S205. A gradient search mechanism is used to perform a neighborhood gradient search on the particle swarm after social learning, resulting in a particle swarm after the neighborhood gradient search.

[0232] S206. The information fusion vibration mechanism is used to perform vibration search on the particle swarm after the neighborhood gradient search, and the particle swarm after vibration search is obtained.

[0233] S207. Based on the particle swarm after vibration search, proceed to the next training and repeat the training multiple times to output the globally optimal particle; that is, output the optimal particle in the particle swarm after vibration search in the last training process.

[0234] S208. Use the hyperparameters in the globally optimal particle as the target hyperparameters of the first deep learning model to obtain the first feature extraction model.

[0235] By introducing a social learning mechanism that combines adaptive weights with sine and cosine algorithms, the algorithm's global exploration and local development capabilities are balanced; the neighborhood gradient search mechanism enhances the particle's fine search capability in the local space; and the information fusion vibration mechanism introduces random perturbations through the cooperative information between particles, effectively helping the algorithm escape local extrema. The three mechanisms work together to significantly improve the global convergence and optimal solution quality of the deep learning model's hyperparameter optimization.

[0236] In one possible design approach, based on historical best values ​​and the best particle, an improved social learning mechanism combining adaptive weights and sine / cosine algorithms is used to perform social learning on the particle swarm, resulting in a socially learned particle swarm, including:

[0237] The adaptive weights are obtained as follows:

[0238] ;

[0239] in, Indicates adaptive weights, Indicates the current number of training iterations. This represents the maximum value of the adaptive weights, and is set to 0.9; This represents the minimum value of the adaptive weights, and is set to 0.1; Represents the natural constant. This indicates the overall decay rate, and can be set to 3, 4, or 5. This indicates the preset maximum number of training iterations. This represents the fluctuation intensity coefficient, and is set to 0.1 or 0.2. Represents pi (π). This indicates the fluctuation period and is set to a value between [20, 50].

[0240] Based on adaptive weights, historical best values, and the best particle, a sine and cosine algorithm is used to perform social learning on the particle swarm, resulting in the following particle swarm after social learning:

[0241] ;

[0242] in, Let d represent the d-th dimension of the hyperparameter of the i-th particle in the particle swarm, where d = 1, 2, ..., Dim, and Dim represents the total dimension of the hyperparameter. Let d represent the d-th dimension hyperparameter of the i-th particle in the particle swarm after social learning. Represents a random number between (0, 1). Represents a random number between (0, 1). Represents a random number between (0, 1). Represents a random number between (0, 1). Denotes the d-th dimension hyperparameter of the optimal particle. Let d be the d-th dimension hyperparameter representing the historical best value of the i-th particle.

[0243] The adaptive weight formula combines the characteristics of exponential decay and periodic cosine fluctuation. It assigns larger weights in the early stage of training to expand the search range, and then the weights decay in the later stage and use the fluctuation characteristics to perform fine cyclical search near the extreme value to prevent premature convergence. The introduction of sine and cosine operators makes the particle update trajectory exhibit nonlinear sine and cosine fluctuations, which increases the diversity of the population and improves the ergodicity of the solution space in the social learning stage.

[0244] In one possible design approach, a gradient search mechanism is used to perform a neighborhood gradient search on the particle swarm after social learning, resulting in a particle swarm that includes:

[0245] For particles in the particle swarm after social learning, the pseudo gradient update information is obtained as follows:

[0246] ;

[0247] ;

[0248] in, This represents the pseudo-gradient update information during the t-th training process. This represents the pseudo-update information during the (t+1)th training iteration. Represents the interval [0,1] 3D random vector, Represents the optimal particle. This represents particles after random social learning. Indicates the step size parameter;

[0249] Based on the current number of training iterations, the gradient search coefficients for the first and second neighborhoods are obtained as follows:

[0250] ;

[0251] ;

[0252] ;

[0253] ;

[0254] in, Represents the gradient search coefficients in the first neighborhood. Indicates the gradient search coefficient of the second neighborhood. Represents a random number between (0, 1). Represents a random number between (0, 1). This represents the base value of the neighborhood gradient search coefficient. Indicates the coefficient of time variation. Represents pi (π). This represents the maximum value of the time variation coefficient, and is set to 1. This represents the minimum value of the time variation coefficient, and is set to 0.2;

[0255] Based on the pseudo-gradient update information, the first neighborhood gradient search coefficients, and the second neighborhood gradient search coefficients, a neighborhood gradient search is performed, resulting in the following particle swarm after the neighborhood gradient search:

[0256] ;

[0257] in, Let represent the d-th dimension hyperparameter of the particle after the m-th social learning iteration. Let represent the particle after the gradient search of the m-th neighborhood, and all particles after the gradient search of their neighborhoods constitute the particle swarm after the gradient search of their neighborhoods. This represents a random number that follows a standard normal distribution. Denotes the d-th dimension hyperparameter of the worst-case particle. The d-th parameter represents the pseudo-gradient update information. Denotes the d-th dimension hyperparameter of the optimal particle. This represents a constant term and is set to 0.01. This represents a random number between (0, 1).

[0258] By using pseudo-gradient update information to simulate the gradient descent direction and combining the first and second neighborhood gradient search coefficients that are dynamically adjusted nonlinearly with the number of iterations, the efficient approximation of particles to the neighborhood of the optimal solution is achieved. At the same time, a repulsion term for the worst particle is introduced, which makes the current particle actively move away from the worst particle while moving closer to the optimal particle. Combined with the guidance of normal distribution random perturbation, the efficiency and accuracy of local depth search are significantly enhanced.

[0259] In one possible design approach, an information fusion vibration mechanism is used to perform a vibration search on the particle swarm after the neighborhood gradient search, resulting in a particle swarm after the vibration search, including:

[0260] Based on the particle swarm after neighborhood gradient search, a random other particle is matched for each particle after neighborhood gradient search to obtain the cooperative particle corresponding to the particle after neighborhood gradient search.

[0261] Based on the cooperative particles corresponding to the particles after neighborhood gradient search, the cooperative information factor is obtained as follows:

[0262] ;

[0263] in, Indicates collaborative information factors. This represents the particle after the gradient search of the nth neighborhood. express The corresponding cooperative particles, Indicates the number of vibration searches;

[0264] Based on the cooperative information factor and the optimal particle, a vibrational search is performed on the particle swarm after the neighborhood gradient search, resulting in the following particle swarm after the vibrational search:

[0265] ;

[0266] in, Let d be the hyperparameter of the particle after the gradient search of the nth neighborhood. Let d represent the d-th dimension hyperparameter of the particle after the nth vibrational search, and all particles after the vibrational search constitute the particle swarm after the vibrational search. This represents the vibration search control coefficient. The sequence number indicates the number of vibration search attempts.

[0267] By randomly matching cooperative particles to calculate the cooperative information factor, the relative positional relationship between particles within the population is used as a reference benchmark for the vibration amplitude. As the number of iterations increases, the decay effect of the cooperative information factor and the alternating positive and negative sign mechanism together form a damped vibration effect, which generates large vibrations in the early stage of training to escape local optima, and the vibration tends to be gentle in the later stage to achieve stable convergence, thus ensuring the robustness of the optimization process.

[0268] In one possible design approach, the second target modal data is used as the desired input data, and the training labels are used as the desired output. A gradient vibration particle swarm optimization algorithm is employed to train a second deep learning model, resulting in a second feature extraction model, including:

[0269] The particle swarm is initialized based on the hyperparameters of the second deep learning model;

[0270] For each particle in the particle swarm, the fitness of the particle is determined by using the second target modality data as the expected input data and the training label as the expected output.

[0271] Based on particle fitness, determine the historical best value of the particle and the best particle in the current training process;

[0272] Based on the historical best value and the best particle, an improved social learning mechanism combining adaptive weights and sine and cosine algorithms is used to perform social learning on the particle swarm, resulting in a socially learned particle swarm.

[0273] A gradient search mechanism is used to perform neighborhood gradient search on the particle swarm after social learning, resulting in a particle swarm after neighborhood gradient search.

[0274] An information fusion vibration mechanism is used to perform vibration search on the particle swarm after neighborhood gradient search, resulting in a particle swarm after vibration search.

[0275] Based on the particle swarm after vibration search, proceed to the next training, and repeat the training multiple times to output the globally optimal particle.

[0276] The hyperparameters in the globally optimal particle are used as the target hyperparameters of the second deep learning model to obtain the second feature extraction model.

[0277] To optimize the hyperparameters of the second modality data, the gradient vibration particle swarm optimization algorithm, consistent with the one described above, is used. This ensures that the second feature extraction model also possesses the ability to deeply mine high-dimensional feature spaces and generalize performance, maintaining the symmetry and consistency of the multimodal branch model optimization.

[0278] In one possible design approach, a third deep learning model is trained using first target modality data, second target modality data, training labels, a first feature extraction model, and a second feature extraction model to obtain a feature recognition model, including:

[0279] The data features of the first target modality data are extracted using a first feature extraction model and the data features of the second target modality data are extracted using a second feature extraction model, and then fused to obtain fused data features;

[0280] The particle swarm is initialized based on the hyperparameters of the third deep learning model;

[0281] For each particle in the particle swarm, the fitness of the particle is determined by using the fused data features as the expected input data and the training labels as the expected output.

[0282] Based on particle fitness, determine the historical best value of the particle and the best particle in the current training process;

[0283] Based on the historical best value and the best particle, an improved social learning mechanism combining adaptive weights and sine and cosine algorithms is used to perform social learning on the particle swarm, resulting in a socially learned particle swarm.

[0284] A gradient search mechanism is used to perform neighborhood gradient search on the particle swarm after social learning, resulting in a particle swarm after neighborhood gradient search.

[0285] An information fusion vibration mechanism is used to perform vibration search on the particle swarm after neighborhood gradient search, resulting in a particle swarm after vibration search.

[0286] Based on the particle swarm after vibration search, proceed to the next training, and repeat the training multiple times to output the globally optimal particle.

[0287] The hyperparameters in the globally optimal particle are used as the target hyperparameters of the third deep learning model to obtain the feature recognition model.

[0288] Modal feature extraction and modal feature fusion recognition are decoupled. After feature-level fusion, the gradient vibration particle swarm algorithm is used to specifically optimize the recognition model, so that the recognition model can adaptively adjust the decision boundary for the high-dimensional joint feature distribution after fusion, maximize the mining of complementary information between multimodal features, and improve the accuracy of final state recognition.

[0289] For example, the first target modality data and the second feature extraction model can be set as a convolutional neural network. The fixed-dimensional data features output by the fully connected layer in the convolutional neural network are directly concatenated into a vector to obtain fused data features. The third deep learning model identifies the fused data features and finally obtains the cable external force damage identification result. When the cable external force damage identification result indicates that there is damage, a text message warning, a telephone voice warning, and / or an audible and visual warning can be sent to the management personnel.

[0290] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, because according to the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0291] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods according to the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0292] Example 2

[0293] According to embodiments of the present invention, an apparatus is also provided for implementing the above-described method for determining the result of cable external force damage based on edge computing. Figure 4 This is a structural block diagram of a cable external force damage determination device based on edge computing according to an embodiment of the present invention, as shown below. Figure 4 As shown, the device includes: an acquisition module 402, a determination module 404, an extraction module 406, a fusion module 408, and an identification module 410. The device will be described in detail below.

[0294] The acquisition module 402 is used to acquire a multimodal recognition model set deployed on edge computing nodes. This multimodal recognition model set includes multiple feature extraction models and feature recognition models. The feature extraction models extract features from different types of modal data. These models are obtained by training the corresponding initial deep learning models using the gradient vibration particle swarm optimization algorithm on the corresponding sample modal data. The sample modal data includes first-mode data under normal cable operation and second-mode data under cable damage conditions. The feature recognition model is trained based on sample fusion data. The gradient vibration particle swarm optimization algorithm includes a social learning mechanism, a neighborhood gradient search mechanism, and an information fusion vibration search mechanism. The social learning mechanism controls the balance between global search and local development of the particle swarm. The neighborhood gradient search mechanism searches within the neighborhood of the current optimal particle to improve optimization accuracy. The information fusion vibration search mechanism allows the particle swarm to escape local optima to explore more hyperparameter spaces. In each iteration of the gradient vibration particle swarm optimization algorithm, the social learning mechanism, the neighborhood gradient search mechanism, and the information fusion vibration search mechanism are executed sequentially to complete a full update of the particle swarm until the preset maximum number of iterations is reached. The hyperparameters corresponding to the current best particle are used as the target hyperparameters of the corresponding deep learning model, and each particle in the particle swarm represents a set of candidate hyperparameters. The determination module 404, connected to the acquisition module 402, is used to determine the real-time modal data corresponding to various modal types. The extraction module 406, connected to the determination module 404, is used to schedule the corresponding feature extraction model to extract features from the corresponding real-time modal data to obtain multiple real-time modal features. The fusion module 408, connected to the extraction module 406, is used to fuse multiple real-time features to obtain the fused features to be identified. The identification module 410, connected to the fusion module 408, is used to schedule the feature identification model to identify the fused features to be identified to obtain the cable external force damage identification result.

[0295] It should be noted that the above-mentioned acquisition module 402, determination module 404, extraction module 406, fusion module 408 and identification module 410 correspond to steps S102 to S110 in the method for determining the result of cable external force damage based on edge computing. The multiple modules and the corresponding steps are the same in terms of implementation examples and application scenarios, but are not limited to the content disclosed in the above embodiment 1.

[0296] Example 3

[0297] According to another aspect of the present invention, an electronic device is also provided, comprising: a processor; and a memory for storing processor-executable instructions, wherein the processor is configured to execute instructions to implement the edge computing-based cable external force damage determination method of any of the above embodiments.

[0298] Example 4

[0299] According to another aspect of the present invention, a computer-readable storage medium is also provided, which, when the instructions in the computer-readable storage medium are executed by a processor of an electronic device, enables the electronic device to perform any of the above-described edge computing-based cable external force damage determination methods.

[0300] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0301] In the above embodiments of the present invention, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0302] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0303] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0304] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0305] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0306] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. An edge computing-based cable external force damage result determination method, characterized by, Applied to edge computing nodes, including: A multimodal recognition model set deployed on edge computing nodes is obtained. This set includes multiple feature extraction models and feature recognition models. The feature extraction models extract features from different types of modal data. These models are obtained by training an initial deep learning model using a gradient vibration particle swarm optimization algorithm on corresponding sample modal data to optimize hyperparameters. The sample modal data includes first-mode data under normal cable operation and second-mode data under cable damage conditions. The feature recognition model is trained based on sample fusion data. The gradient vibration particle swarm optimization algorithm includes a social learning mechanism, a neighborhood gradient search mechanism, and information fusion vibration. The search mechanism includes a social learning mechanism to control the balance between global search and local development of the particle swarm, a neighborhood gradient search mechanism to search within the neighborhood of the current optimal particle to improve optimization accuracy, and an information fusion vibration search mechanism to enable the particle swarm to escape local optima and explore more hyperparameter spaces. In each iteration of the gradient vibration particle swarm algorithm, the social learning mechanism, the neighborhood gradient search mechanism, and the information fusion vibration search mechanism are executed sequentially to complete a full update of the particle swarm until the preset maximum number of iterations is reached. The hyperparameters corresponding to the current optimal particle are used as the target hyperparameters of the corresponding deep learning model. Each particle in the particle swarm represents a set of candidate hyperparameters. Determine the real-time modal data corresponding to each of the multiple modal types; The corresponding feature extraction model is scheduled to extract features from the corresponding real-time modal data, resulting in multiple real-time modal features; Multiple real-time features are fused to obtain the fused feature to be identified; The feature recognition model is scheduled to identify the fused features to be identified, and the cable external force damage identification result is obtained.

2. The method according to claim 1, characterized in that The social learning mechanism is used to update the position of the particle swarm through adaptive weights and sine and cosine operators to maintain population diversity. Maintaining population diversity is used to prevent the particle swarm from clustering to a local optimum less than a predetermined number of times during the optimization process, thereby maintaining global exploration capability. The neighborhood gradient search mechanism is used to guide particles to search within the neighborhood to approach the local optimum through pseudo-gradient information. The pseudo-gradient information is an alternative gradient vector constructed using the positional difference between the optimal particle and random particles. The neighborhood is the hyperparameter space region within a predetermined distance range of the current optimal particle. The local optimum is the hyperparameter combination that minimizes the loss function value of the deep learning model within the neighborhood. The loss function value is a quantification of the degree of difference between the predicted output and the actual output of the deep learning model on the training data. The information fusion vibration search mechanism is used to generate vibration perturbations through the cooperative information between particles to help particles jump out of local optima.

3. The method of claim 1, wherein, Before acquiring the multimodal recognition model set deployed on edge computing nodes, the following steps are also included: For any feature extraction model, the corresponding sample modal data is used as the expected input data, and the corresponding training label is used as the expected output. The gradient vibration particle swarm algorithm is used to train the corresponding initial deep learning model to obtain the corresponding feature extraction model.

4. The method according to claim 3, characterized in that Using the corresponding sample modal data as the expected input data and the corresponding training labels as the expected output, the gradient vibration particle swarm optimization algorithm is used to train the corresponding initial deep learning model to obtain the corresponding feature extraction model, including: Based on the initial hyperparameters of the initial deep learning model, the initial particle swarm is determined; For each particle in the initial particle swarm, the fitness of each particle is determined by using the sample modal data as the expected input data and the training label as the expected output. Based on the fitness of each particle, determine the historical best position of each particle and the current best particle of the particle swarm during the current training process; Based on the historical best position of each particle and the current best particle of the particle swarm, the gradient vibration particle swarm algorithm is used to train the corresponding initial deep learning model and output the global best particle. The hyperparameters in the globally optimal particle are used as the target hyperparameters of the initial deep learning model to obtain the trained feature extraction model.

5. The method of claim 1, wherein, The operation of the social learning mechanism includes: Obtain the first updated particle swarm; Obtain the adaptive weights corresponding to the first updated particle swarm at the current training iterations, wherein the adaptive weights adopt an exponential decay form and are superimposed with cosine fluctuations on the decay basis; For each first updated particle in the first updated particle swarm, based on the adaptive weight, the current position of the first updated particle in the current training iteration, the position of the current best particle in the current training process, and the historical best position of the first updated particle, the sine and cosine algorithms are used to determine the updated position of the first updated particle in the next training iteration. The calculated update position is used as the position after the social learning of the first update particle. After all the first update particles are updated, a second update particle group including multiple second update particles is obtained.

6. The method of claim 1, wherein, The operation of the neighborhood gradient search mechanism includes: Obtain the pseudo-gradient update information corresponding to the second updated particle swarm, and use the positional difference between the current best particle under the current training number and the second updated particle randomly selected from the second updated particle swarm, combined with the step size parameter, to generate a random vector with the same dimension as the hyperparameter, as the pseudo-gradient update information for the next iteration. The step size parameter is determined based on the distance between the current best particle and the randomly selected second updated particle and the previous pseudo-gradient update information. Based on the current number of training iterations, the first neighborhood gradient search coefficient and the second neighborhood gradient search coefficient are obtained. Both the first neighborhood gradient search coefficient and the second neighborhood gradient search coefficient change non-linearly with time. The first neighborhood gradient search coefficient is used to control the degree of repulsion of the second updated particle away from the worst particle, and the second neighborhood gradient search coefficient is used to control the degree of step size adjustment of the second updated particle toward the best particle. For each second updated particle in the second updated particle swarm, the position of the second updated particle is updated by neighborhood search based on the pseudo gradient update information, the first neighborhood gradient search coefficient and the second neighborhood gradient search coefficient, so as to obtain the position of the second updated particle after neighborhood gradient search. After all the second updated particles are updated, a third updated particle swarm including multiple third updated particles is obtained.

7. The method according to any one of claims 1 to 6, characterized in that, The operation of the information fusion vibration search mechanism includes: For each third update particle in the third update particle swarm, randomly match one other third update particle as a cooperative particle; For each third update particle, the cooperation information factor of the third update particle is determined based on the distance between the corresponding third update particle and the corresponding cooperative particle and the current training number, wherein the cooperation information factor is used to control the amplitude of vibration search. For each third update particle, the position of the third update particle is updated by vibration search based on the cooperative information factor and the current best particle, so as to obtain the position of the third update particle after vibration search. After all the third update particles are updated, the fourth update particle group after vibration search is obtained, thus completing one complete iteration update of the particle group.

8. An edge computing-based cable external force damage result determination apparatus, characterized by, Applied to edge computing nodes, including: The acquisition module is used to acquire a multimodal recognition model set deployed on edge computing nodes. This multimodal recognition model set includes multiple feature extraction models and feature recognition models. The multiple feature extraction models are used to extract features from different types of modal data. These models are obtained by training corresponding initial deep learning models using a gradient vibration particle swarm optimization algorithm on corresponding sample modal data to optimize hyperparameters. The sample modal data includes first modal data under normal cable operation and second modal data under cable damage conditions. The feature recognition model is trained based on sample fusion data. The gradient vibration particle swarm optimization algorithm includes a social learning mechanism, a neighborhood gradient search mechanism, and information fusion. The algorithm employs a combined vibrational search mechanism. The social learning mechanism controls the balance between global and local exploration in the particle swarm. The neighborhood gradient search mechanism searches within the neighborhood of the current optimal particle to improve optimization accuracy. The information fusion vibrational search mechanism enables the particle swarm to escape local optima and explore more hyperparameter spaces. In each iteration of the gradient vibrational particle swarm algorithm, the social learning mechanism, the neighborhood gradient search mechanism, and the information fusion vibrational search mechanism are executed sequentially to complete a full update of the particle swarm until a preset maximum number of iterations is reached. The hyperparameters corresponding to the current optimal particle are used as the target hyperparameters of the corresponding deep learning model. Each particle in the particle swarm represents a set of candidate hyperparameters. The determination module is used to determine the real-time modal data corresponding to various modal types. The extraction module is used to schedule the corresponding feature extraction model to extract features from the corresponding real-time modal data and obtain multiple real-time modal features; The fusion module is used to fuse multiple real-time features to obtain the fused features to be identified; The identification module is used to schedule the feature identification model to identify the fused features to be identified, and obtain the identification result of cable external force damage.

9. An electronic device, characterized in that, include: processor; Memory used to store the processor's executable instructions; The processor is configured to execute the instructions to implement the method for determining cable external force damage results based on edge computing as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, When the instructions in the computer-readable storage medium are executed by the processor of the electronic device, the electronic device is able to perform the edge computing-based cable external force damage determination method as described in any one of claims 1 to 7.