110KV cable external damage prevention monitoring method and 110KV cable external damage prevention monitoring system

Abstract

The invention discloses a 110 kilovolt cable external damage prevention monitoring method and system. The method comprises the following steps: S1, monitoring deployment and sensing network establishment; s2, synchronously collecting multi-source data; s3, visual risk target identification; s4, multi-modal signal feature extraction is carried out; s5, pre-judging the risk of the edge end; s6, multi-dimensional result fusion and accurate judgment are carried out; according to the method, the risk perception precision is improved through a three-dimensional adaptive feature processing mechanism, and the accuracy of risk target identification and feature extraction is improved; data transmission time delay is reduced through edge end localization processing, multi-dimensional fusion analysis is achieved by means of the powerful computing power of the cloud end, the contradiction that real-time performance and accuracy in traditional monitoring are difficult to consider at the same time is solved, and the efficiency and reliability of risk judgment are improved. Through a dynamic risk field coupling determination mechanism, dynamic quantification and accurate early warning of risks are realized, pertinence and effectiveness of cable external damage prevention monitoring are improved, and threats of external damage events to cable operation are reduced.

Description

Technical Field

[0001] This invention relates to the field of power system operation and maintenance technology, specifically to a method and system for monitoring external damage to 110 kV cables. Background Technology

[0002] As the core carrier of power transmission in urban and suburban areas, 110 kV cables are laid in various scenarios, including direct burial, tunnels, and cable trays. The main external damage risks during operation stem from construction machinery operations, falling trees, and accidental human impacts. To ensure safe cable operation, existing monitoring methods have initially incorporated vibration sensors, acoustic sensors, and AI vision acquisition units. By deploying multi-source equipment to build a basic sensing network, preliminary data collection on the cable's surrounding environment and abnormal signals is achieved, providing raw data support for risk identification. However, certain shortcomings still exist in practical power operation and maintenance applications.

[0003] Firstly, in densely populated construction areas on the outskirts of cities and towns, the high-frequency operation of construction machinery generates complex vibration and acoustic signals. Traditional monitoring technologies often rely on single-modal data for risk identification. Some solutions only capture mechanical targets through visual acquisition units, ignoring the temporal correlation between vibration signals and target movements. This leads to misjudging risks when the machinery is stationary and missing risks during operation due to visual obstruction. Furthermore, in densely wooded suburban areas, the low-frequency signals generated by swaying trees are easily confused with environmental noise. If existing technologies only use vibration sensors to collect data, they lack correlation analysis of the movement trajectory of visual targets, often misjudging normal tree swaying as a risk of falling. This single-modal data acquisition and isolated feature analysis model is difficult to adapt to the multi-source risk characteristics in complex scenarios, thus limiting the accuracy of risk identification.

[0004] Secondly, regarding the need for comprehensive cable monitoring, existing technologies typically deploy edge nodes near the monitoring terminal and rely on a cloud platform for data aggregation. However, the computing power configuration of edge nodes often exhibits uneven regional characteristics: edge nodes in remote suburban areas have lower computing power due to cost constraints, while high-computing-power nodes are configured in densely populated construction areas due to the large amount of data. Traditional solutions do not design differentiated task allocation mechanisms for differences in computing power. When low-computing-power nodes are forced to perform complex feature extraction, data backlog is likely to occur, resulting in high latency in risk assessment and failing to meet the early warning needs of sudden operations of construction machinery. When high-computing-power nodes only perform simple data forwarding, it results in resource waste. On the other hand, relying entirely on the cloud for feature processing will further exacerbate the problem of insufficient real-time performance due to transmission link delays, creating a contradiction between real-time performance and processing accuracy.

[0005] Third, the evolution of cable damage risk is significantly influenced by dynamic environmental factors. For example, strong winds can accelerate the spread of tree fall risk, and changes in the timeliness of construction permits can alter regional risk levels. However, existing risk assessment technologies are mostly based on fixed parameters to build early warning models. Some solutions use a preset risk radius to divide the early warning area without adjusting the risk impact range of tree fall based on real-time wind speed. Some solutions only set risk coefficients based on historical damage data and cannot correct the assessment parameters through external dynamic data such as construction permits and meteorological warnings. This static risk assessment model leads to a disconnect between the early warning range and the actual risk evolution. For example, in the event of sudden short-term strong winds, traditional fixed-range early warnings may miss newly added high-risk areas, and warnings continue to be pushed after construction is completed, reducing the targeting and efficiency of operation and maintenance work.

[0006] Therefore, it is necessary to design a monitoring method and system for preventing external damage to 110 kV cables. Summary of the Invention

[0007] The purpose of this invention is to provide a method and system for monitoring external damage to 110 kV cables, in order to solve the problems mentioned in the background art, such as the reliance on isolated data in traditional single-modal monitoring leading to misjudgment and missed judgment in risk identification in scenarios such as densely populated construction areas in urban-rural fringe areas and densely wooded suburban areas. It also addresses the contradiction between insufficient real-time performance and resource waste caused by the fixed task allocation in traditional solutions when the computing power distribution of edge nodes is uneven, and solves the problem that existing static early warning models cannot adapt to dynamic environmental changes.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] Firstly, a method for monitoring external damage to 110 kV cables is provided, including the following steps:

[0010] S1: Monitoring Deployment and Sensing Network Construction: Based on the laying scenario of 110 kV cables and high-risk areas for external damage, monitoring terminals integrating high-precision positioning modules are deployed. These terminals integrate vibration sensors, acoustic sensors, and AI vision acquisition units to meet high-voltage insulation and electromagnetic interference resistance requirements. Edge nodes are deployed simultaneously to establish communication connections with all monitoring terminals, constructing a multi-source sensing network covering the entire cable area and including monitoring terminals and edge nodes. By standardizing the communication connections between monitoring terminals and edge nodes, signal compatibility and interoperability between different types of monitoring equipment are ensured, laying a hardware collaborative foundation for the subsequent unified acquisition and synchronization of multi-source data.

[0011] S2: Multi-source data synchronous acquisition: The multi-source sensing network constructed in S1 is activated to acquire multi-source data in real time. The multi-source data includes vibration signals, acoustic signals, environmental video streams, and monitoring terminal location data around the cable. The timestamp synchronization and low-latency transmission of the multi-source data are achieved through edge nodes, and the synchronized multi-source data is output. The time synchronization and low-latency transmission adopt a distributed clock calibration mechanism to avoid the timing misalignment of multi-source data caused by the clock deviation of a single node, and to ensure the time correlation of vibration signals, acoustic signals, and environmental video streams, providing a reliable timing reference for cross-modal feature verification.

[0012] S3: Visual Risk Target Recognition: A lightweight visual recognition model is deployed on the edge nodes involved in S2. The environmental video stream from the synchronized multi-source data output by S2 is input, and it is analyzed in real time to extract relevant features and time series information of external damage risk targets, preliminarily determine the target intrusion risk, and output visual recognition data. The deployment of the lightweight visual recognition model adopts a computing power adaptation strategy. Based on the hardware performance of the edge nodes, redundant parameters of the model are trimmed to reduce the computing power consumption at the edge while ensuring recognition accuracy, so as to realize real-time analysis of high-concurrency video streams.

[0013] S4: Multimodal Signal Feature Extraction: A multi-scale signal decomposition algorithm is adopted. The vibration and acoustic signals from the synchronized multi-source data output by S2, along with the temporal information from the visual recognition data output by S3, are input. The vibration and acoustic signals are processed to extract key parameters. Cross-modal temporal correlation verification is performed by correlating the temporal information, isolated features are removed, and multimodal feature data is constructed and output. The multi-scale signal decomposition algorithm can adapt to external damage feature signals in different frequency bands. By hierarchically decomposing the vibration and acoustic signals, it can uncover the feature differences corresponding to different external damage behaviors. Cross-modal temporal correlation verification can filter out interference signals unrelated to the visual target, improving the effectiveness of the multimodal feature data.

[0014] S5: Edge Risk Pre-judgment: Two heterogeneous models are deployed on the edge nodes involved in S2. Feature processing tasks are allocated based on the computing power characteristics of the edge nodes. The input data includes the visual recognition data output by S3, the multimodal feature data output by S4, and the preliminary intrusion risk judgment result of S3. The lightweight decision model performs a fusion pre-judgment on the input data and outputs a three-level preliminary risk result. The lightweight decision model dynamically adjusts the feature weights to adapt to the risk judgment requirements of different external damage scenarios.

[0015] S6: Multi-dimensional Result Fusion and Precise Judgment: Input the three-level preliminary risk results output by S5, the basic deployment data output by S1, and the real-time environmental data from the synchronized multi-source data output by S2. Upload the preliminary risk results to the cloud to construct a dynamic diffusion-type risk field. Integrate high-precision positioning spatial correlation, risk field diffusion intensity, and multi-modal feature confidence for three-dimensional coupling analysis. Access external dynamic data to correct parameters, accurately determine the authenticity and risk level of external damage events, and output the final judgment result. The construction of the dynamic diffusion-type risk field breaks through the limitations of fixed-area risk judgment and can quantify the spatial impact range of external damage risks. The three-dimensional coupling analysis integrates spatial, intensity, and feature information to achieve complementary verification of multi-dimensional information and reduce the probability of misjudgment in single-dimensional judgment.

[0016] As a further technical solution of the present invention, in S1, the laying scenarios include underground direct burial, tunnels, and cable trays, and the high-risk areas for external damage include key areas with dense construction and abundant trees; the basic deployment data includes the spacing between monitoring terminals. And high-precision positioning information, deployment spacing pass:

[0017]

[0018] in, To establish the baseline value for the spacing, This represents the regional risk level coefficient. Both are scene complexity coefficients, determined based on historical monitoring data and scene characteristic statistics; This serves as the foundational data for subsequent signal propagation speed calculations and location correlation analysis.

[0019] As a further technical solution of the present invention, in S2, the synchronized multi-source data includes the adjusted sampling frequency. Frame rate Vibration signal after timestamp synchronization Acoustic signals, environmental video streams, and monitoring terminal location data are all collected through a multi-source sensing network and synchronously processed by edge nodes; the sampling frequency is dynamically adjusted based on the external damage risk patterns. Frame rate Define the risk coefficient for a given period:

[0020]

[0021] in, The frequency of historical external damage events during the target period. The average frequency of historical external damage events over the entire time period; vibration signal sampling frequency adjustment:

[0022]

[0023] Environmental video stream frame rate adjustment:

[0024]

[0025] in, The sampling frequency reference value for the vibration signal. This serves as the baseline value for the environmental video stream frame rate. To adjust the gain coefficient; timestamp synchronization passed:

[0026]

[0027] in, The original acquisition time for each sensor. The time deviation between the sensor and the edge node is obtained by calibration using the time synchronization protocol.

[0028] As a further technical solution of the present invention, in S3, the external damage risk targets include construction machinery and tall trees; the visual recognition data includes target contour features. Characteristics of motion trajectory The timing of visual target appearance and visual recognition feature vector All four are generated by analyzing environmental video streams using a lightweight visual recognition model; the quantification of target contour features is achieved through:

[0029]

[0030] in, The length of the target contour edge line segment. The area of ​​the target's bounding rectangle;

[0031] The characteristics of the motion trajectory are calculated by:

[0032]

[0033] in, For the goal of Frame coordinates, The coordinates of the center point of the target trajectory. This represents the number of trajectory frames.

[0034] For integration and The high-dimensional feature vectors are output through the fully connected layer of the model; The timestamp at which the target first appears in the video stream is calibrated by the edge node clock.

[0035] Combined with the risk coefficient of external damage during the time period Adjust visual recognition sensitivity:

[0036]

[0037] in, To identify the sensitivity benchmark, This is the sensitivity adjustment coefficient;

[0038] The weighting formula for enhancing the capture of the movement trajectories of tall trees is:

[0039]

[0040] in, As the baseline value for trajectory capture weights, This refers to the wind speed influence coefficient.

[0041] Preliminary intrusion risk assessment results are based on Generates by comparing with a preset threshold, when If the feature similarity exceeds the threshold, it is considered to have an intrusion risk; otherwise, it is considered to have no risk.

[0042] As a further technical solution of the present invention, in S4, the multimodal feature data includes a multimodal feature matrix. and the spectral feature vector of the disassembled vibration signal Acoustic signal spectral eigenvectors All three are generated through multi-scale signal decomposition and cross-modal verification processing, specifically including: multi-scale signal processing using wavelet packet decomposition algorithm:

[0043]

[0044] in, For the first Layer The decomposition coefficients of each node. The vibration signal output by S2 Define wavelet packet basis functions; define scene adaptation coefficients. The low-frequency mechanical vibration characteristic weight adjustment formula is determined based on the statistical analysis of scene signal interference intensity.

[0045]

[0046] in, The low-frequency feature weight benchmark value, The feature contribution coefficient is used; the formula for adjusting the high-frequency interference filtering threshold is:

[0047]

[0048] in, This is the baseline value for the high-frequency filtering threshold. The noise suppression coefficient is used; the cross-modal temporal correlation verification formula is:

[0049]

[0050] in, The timing of the visual target appearance output by S3. For vibration or acoustic signal triggering timing, , The spacing between the monitoring terminals output by S1 is set. For the speed of signal propagation, when The corresponding features are retained when they are active, otherwise they are discarded; the formula for constructing the multimodal feature matrix is ​​as follows: ,in This is for the Hadamard product operation.

[0051] As a further technical solution of the present invention, in S5, the feature processing task is specifically allocated by combining the computing power characteristics of edge nodes through node computing power ratio quantization. The two heterogeneous models are the visual feature processing model and the multimodal feature processing model, which respectively process the visual recognition data of S3 and the multimodal feature data of S4.

[0052] Computing power adaptation and task allocation: Defining the node computing power ratio ,in The computing power of a single edge node. The arithmetic mean of the computing power of all edge nodes; when At this time, this node undertakes complex feature extraction and preliminary distillation tasks, matching high computing power, and the task weight formula is as follows: ;when At that time, the node focuses on the core feature screening task, adapting to low computing power scenarios, and the task weight formula is: ,in As the baseline value for task weight, For complex tasks, weight coefficients The core task weighting coefficient;

[0053] Calculation of preliminary risk results at level three: The outputs of the two heterogeneous models are fused using a lightweight decision-making model, as shown in the formula below. ,in , , For feature weights, For bias terms, It is the Sigmoid activation function. This is the preliminary intrusion risk assessment result for S3. A value of 1 indicates a risk, and 0 indicates no risk. , The signal feature vector of S4, This is the visual recognition feature vector of S3;

[0054] Level L1, slight risk. This is a serious risk for L2. This is an L3 emergency risk.

[0055] As a further technical solution of the present invention, in step S6, the final determination result includes an external damage event authenticity label and a quantified risk level, as follows:

[0056] Dynamic risk field construction: Input the basic deployment data of S1, using its high-precision positioning coordinates As the coordinates of the risk source, output by S5 Mapping the initial strength of risk sources A risk field is constructed based on the Gaussian diffusion model, and the formula for the risk field strength is: ,in Let be the coordinates of any point in the field. The diffusion coefficient is denoted as .

[0057] diffusion coefficient calculate: , The initial diffusion coefficient is... The environmental impact coefficient is calculated from real-time environmental data of S2. Regional risk coefficient;

[0058] 3D Coupling Determination: Confidence of Spatial Correlation in High-Precision Positioning , The distance between the monitoring point of S1 and the cable. To determine the effective distance threshold and the confidence level of the risk field diffusion intensity. Multimodal feature confidence , The multimodal feature matrix of S4; coupling formula , ;

[0059] Corrected by external dynamic data The corrected formula is as follows: , This is a correction amount;

[0060] Accurate Judgment and Output: Authenticity: It is true and marked with T. False and marked with F; Grade correction: Final risk level assessment: Corresponding to L1, Corresponding to L2, Corresponding to L3.

[0061] As a further technical solution of the present invention, it also includes S7: graded early warning and cross-platform linkage: inputting the risk level in the final judgment result output by S6, the high-precision positioning coordinates in the basic deployment data output by S1, the real-time signal data fragments in the synchronized multi-source data output by S2, and the video screenshots in the visual recognition data output by S3, triggering graded responses. Minor risks are pushed through a pop-up window on the monitoring platform, serious risks are pushed through SMS and the operation and maintenance APP, and emergency risks additionally trigger on-site audible and visual alarms and power dispatching system alarms; outputting push information including high-precision positioning coordinates, risk level, real-time signal data fragments and video screenshots, serving as the basis for operation and maintenance personnel to handle on-site, and the handling results serve as input for subsequent data feedback; the graded response mechanism matches different early warning channels according to the risk level, ensuring that operation and maintenance personnel can prioritize handling high-level events according to the urgency of the risk; the structured encapsulation of push information facilitates information parsing and display on different terminals, improving handling efficiency.

[0062] As a further technical solution of the present invention, it also includes S8: data feedback and model iterative optimization: inputting the operation and maintenance site handling results output by S7, updating the external damage historical sample library and risk field parameter library; the edge end responds to S3... With S4 , The original feature matrix formed Compression is performed using the following formula:

[0063]

[0064] in, For S3 output With S4 output , The original feature matrix formed, For convolution kernel weights, For convolution bias, The ReLU activation function is used; features are compressed at each edge in the cloud. Take the average to get Then, the features are optimized using a cloud-based distillation formula, which is:

[0065]

[0066] in, For the weight of the distillation layer, For distillation layer bias, The feature vectors at each edge are compressed; the model optimization uses stochastic gradient descent, and the loss function is:

[0067]

[0068] in, This is the actual label for the external damage incident. For the model prediction results, For weighted regularization terms, The regularization coefficient is used to output the optimized model parameters, which are then fed back to S3, S4, and S5 to achieve dynamic model iteration and form a closed-loop technology.

[0069] Secondly, a 110 kV cable external damage prevention monitoring system is provided, including a multi-source sensing network module, an edge processing module, a cloud fusion judgment module, a hierarchical early warning linkage module, and a data feedback optimization module, with each module communicating with each other in sequence;

[0070] The multi-source sensing network module includes several monitoring terminals that integrate high-precision positioning modules. The monitoring terminals integrate vibration sensors, acoustic sensors, and AI vision acquisition units to meet the requirements of high voltage insulation and electromagnetic interference resistance. They are used to perform the monitoring deployment of S1 and the multi-source data acquisition of S2, and output the basic deployment data of S1 and the synchronized multi-source data of S2.

[0071] The edge processing module, deployed on edge nodes, integrates a lightweight visual recognition model, two types of heterogeneous models, and a lightweight decision model. It is used to perform visual analysis and initial risk assessment in S3, multimodal feature extraction in S4, and fusion pre-judgment in S5. It outputs visual recognition data, multimodal feature data, and three-level preliminary risk results to the cloud fusion judgment module, while receiving updated parameters from the data feedback optimization module.

[0072] The cloud-based fusion judgment module is deployed on a cloud server. It has a built-in Gaussian diffusion model and a three-dimensional coupled judgment algorithm. It is used to perform dynamic risk field construction and multi-dimensional fusion judgment of S6, and to access external dynamic data to correct parameters and output the final judgment result of S6.

[0073] The tiered early warning and linkage module includes a monitoring platform, an operation and maintenance APP, on-site audible and visual alarms, and an interface to the power dispatching system. It is used to execute the tiered response and information push of S7 and output the push information of S7.

[0074] The data feedback optimization module is used to record the on-site handling results of S7's operation and maintenance, update the sample library and parameter library, perform iterative optimization of S8's model through the edge-cloud bidirectional distillation algorithm, output the optimized model parameters and feed them back to the edge processing module, forming a technical closed loop.

[0075] Thirdly, a computer-readable storage medium is provided, on which a computer program is stored, wherein when the computer program is executed by a processor, it implements all the steps of the 110 kV cable external damage monitoring method described in any of the first aspects.

[0076] Fourthly, a 110 kV cable external damage monitoring electronic device is provided, comprising a processor, a memory, and a communication interface, wherein the processor, memory, and communication interface are interconnected, wherein the communication interface is used for data interaction with a monitoring terminal, an edge node, and a cloud server; the memory is used to store a computer program; and the processor is used to call the computer program in the memory to execute all the steps of any of the 110 kV cable external damage monitoring methods described in the first aspect.

[0077] Compared with existing technologies, the beneficial effects of this 110 kV cable external damage monitoring method and system are:

[0078] In high-risk scenarios of external damage in densely populated construction areas in urban-rural fringe areas and densely wooded suburban areas, this technology improves the accuracy of risk perception through a three-dimensional adaptive feature processing mechanism. The lightweight visual recognition model in S3 extracts the contour features of external damage risk targets. Characteristics of motion trajectory Time sequence information S4 synchronously receives vibration signals For acoustic signals, wavelet packet decomposition algorithm is used to complete multi-scale signal processing, combined with cross-modal temporal correlation verification formula. Remove isolated features and construct a multimodal feature matrix This multi-dimensional feature fusion method avoids the one-sidedness of single-modality data, makes the feature extraction process more in line with the actual occurrence pattern of external damage risks, effectively solves the problem of misjudgment and omission caused by single features in traditional monitoring, and improves the accuracy of risk target identification and feature extraction.

[0079] To address the uneven distribution of computing power across multiple edge nodes and the need for real-time monitoring, this technology achieves a balance between processing efficiency and judgment accuracy through an edge-cloud collaborative optimization mechanism. In S5, this is based on the node computing power ratio... Dynamically allocate feature processing tasks: nodes with sufficient computing power undertake complex feature extraction and preliminary distillation, while nodes with limited computing power focus on core feature selection, in conjunction with a lightweight decision model using formulas. Quickly output preliminary risk results at level three The S6 cloud platform constructs a dynamic diffusion-type risk field based on a Gaussian diffusion model, and uses a three-dimensional coupling judgment formula. And correct them simultaneously. get This collaborative model achieves accurate judgment by reducing data transmission latency through localized processing at the edge and leveraging the powerful computing power of the cloud to achieve multi-dimensional fusion analysis. It resolves the contradiction between real-time performance and accuracy in traditional monitoring, thereby improving the efficiency and reliability of risk assessment.

[0080] In response to the dynamic nature of external damage risks that change with the environment, this technology achieves dynamic quantification and precise early warning of risks through a dynamic risk field coupling judgment mechanism. The risk field intensity formula in S6... Real-time reflection of risk diffusion status; formula for calculating the risk diffusion coefficient. Incorporating environmental impact coefficient Regional risk coefficient Dynamic factors, combined with high-precision positioning spatial correlation confidence level Confidence level of risk field diffusion intensity and multimodal feature confidence A comprehensive analysis was conducted, and external dynamic data was corrected. The parameters enable dynamic adjustment of the model, ensuring that the risk assessment results remain consistent with the actual environment and risk evolution trend. This provides accurate basis for S7 level early warning, improves the pertinence and effectiveness of cable external damage monitoring, and reduces the threat of external damage events to cable operation. Attached Figure Description

[0081] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0082] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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 are within the scope of protection of the present invention.

[0083] Please see the appendix Figure 1 The present invention provides an embodiment 1: a method for monitoring external damage to 110 kV cables, comprising the following steps:

[0084] S1: Monitoring Deployment and Sensing Network Construction: Based on the laying scenario of 110 kV cables and high-risk areas for external damage, monitoring terminals integrating high-precision positioning modules are deployed. These terminals integrate vibration sensors, acoustic sensors, and AI vision acquisition units to meet high-voltage insulation and electromagnetic interference resistance requirements. Edge nodes are deployed simultaneously to establish communication connections with all monitoring terminals, constructing a multi-source sensing network covering the entire cable area and including monitoring terminals and edge nodes. By standardizing the communication connections between monitoring terminals and edge nodes, signal compatibility and interoperability between different types of monitoring equipment are ensured, laying a hardware collaborative foundation for the subsequent unified acquisition and synchronization of multi-source data.

[0085] The laying scenarios include underground direct burial, tunnels, and cable trays, while high-risk areas for external damage include key areas with dense construction and abundant vegetation; basic deployment data includes the spacing between monitoring terminals. And high-precision positioning information, deployment spacing pass:

[0086]

[0087] in, To establish the baseline value for the spacing, This represents the regional risk level coefficient. Both are scene complexity coefficients, determined based on historical monitoring data and scene characteristic statistics; This serves as the foundational data for subsequent signal propagation speed calculations and location correlation analysis;

[0088] S2: Multi-source data synchronous acquisition: The multi-source sensing network constructed in S1 is activated to acquire multi-source data in real time. The multi-source data includes vibration signals, acoustic signals, environmental video streams, and monitoring terminal location data around the cable. The timestamp synchronization and low-latency transmission of the multi-source data are achieved through edge nodes, and the synchronized multi-source data is output. The time synchronization and low-latency transmission adopt a distributed clock calibration mechanism to avoid the timing misalignment of multi-source data caused by the clock deviation of a single node, and to ensure the time correlation of vibration signals, acoustic signals, and environmental video streams, providing a reliable timing reference for cross-modal feature verification.

[0089] Synchronized multi-source data includes the adjusted sampling frequency Frame rate Vibration signal after timestamp synchronization Acoustic signals, environmental video streams, and monitoring terminal location data are all collected through a multi-source sensing network and synchronously processed by edge nodes; the sampling frequency is dynamically adjusted based on the external damage risk patterns. Frame rate Define the risk coefficient for a given period:

[0090]

[0091] in, The frequency of historical external damage events during the target period. The average frequency of historical external damage events over the entire time period; vibration signal sampling frequency adjustment:

[0092]

[0093] Environmental video stream frame rate adjustment:

[0094]

[0095] in, The sampling frequency reference value for the vibration signal. This serves as the baseline value for the environmental video stream frame rate. To adjust the gain coefficient; timestamp synchronization passed:

[0096]

[0097] in, The original acquisition time for each sensor. The time deviation between the sensor and the edge node is obtained by calibration using the time synchronization protocol.

[0098] S3: Visual Risk Target Recognition: A lightweight visual recognition model is deployed on the edge nodes involved in S2. The environmental video stream from the synchronized multi-source data output by S2 is input, and it is analyzed in real time to extract relevant features and time series information of external damage risk targets, preliminarily determine the target intrusion risk, and output visual recognition data. The deployment of the lightweight visual recognition model adopts a computing power adaptation strategy. Based on the hardware performance of the edge nodes, redundant parameters of the model are trimmed to reduce the computing power consumption at the edge while ensuring recognition accuracy, so as to realize real-time analysis of high-concurrency video streams.

[0099] External damage risk targets include construction machinery and tall trees; visual recognition data includes target outline features. Characteristics of motion trajectory The timing of visual target appearance and visual recognition feature vector All four are generated by analyzing environmental video streams using a lightweight visual recognition model; the quantification of target contour features is achieved through:

[0100]

[0101] in, The length of the target contour edge line segment. The area of ​​the target's bounding rectangle;

[0102] The characteristics of the motion trajectory are calculated by:

[0103]

[0104] in, For the goal of Frame coordinates, The coordinates of the center point of the target trajectory. This represents the number of trajectory frames.

[0105] For integration and The high-dimensional feature vectors are output through the fully connected layer of the model; The timestamp at which the target first appears in the video stream is calibrated by the edge node clock.

[0106] Combined with the risk coefficient of external damage during the time period Adjust visual recognition sensitivity:

[0107]

[0108] in, To identify the sensitivity benchmark, This is the sensitivity adjustment coefficient;

[0109] The weighting formula for enhancing the capture of the movement trajectories of tall trees is:

[0110]

[0111] in, As the baseline value for trajectory capture weights, This refers to the wind speed influence coefficient.

[0112] Preliminary intrusion risk assessment results are based on Generates by comparing with a preset threshold, when If the feature similarity exceeds the threshold, it is considered to have an intrusion risk; otherwise, it is considered to have no risk.

[0113] S4: Multimodal Signal Feature Extraction: A multi-scale signal decomposition algorithm is adopted. The vibration and acoustic signals from the synchronized multi-source data output by S2, along with the temporal information from the visual recognition data output by S3, are input. The vibration and acoustic signals are processed to extract key parameters. Cross-modal temporal correlation verification is performed by correlating the temporal information, isolated features are removed, and multimodal feature data is constructed and output. The multi-scale signal decomposition algorithm can adapt to external damage feature signals in different frequency bands. By hierarchically decomposing the vibration and acoustic signals, it can uncover the feature differences corresponding to different external damage behaviors. Cross-modal temporal correlation verification can filter out interference signals unrelated to the visual target, improving the effectiveness of the multimodal feature data.

[0114] Multimodal feature data includes multimodal feature matrices. and the spectral feature vector of the disassembled vibration signal Acoustic signal spectral eigenvectors All three are generated through multi-scale signal decomposition and cross-modal verification processing, specifically including: multi-scale signal processing using wavelet packet decomposition algorithm:

[0115]

[0116] in, For the first Layer The decomposition coefficients of each node. The vibration signal output by S2 Define wavelet packet basis functions; define scene adaptation coefficients. The low-frequency mechanical vibration characteristic weight adjustment formula is determined based on the statistical analysis of scene signal interference intensity.

[0117]

[0118] in, The low-frequency feature weight benchmark value, The feature contribution coefficient is used; the formula for adjusting the high-frequency interference filtering threshold is:

[0119]

[0120] in, This is the baseline value for the high-frequency filtering threshold. The noise suppression coefficient is used; the cross-modal temporal correlation verification formula is:

[0121]

[0122] in, The timing of the visual target appearance output by S3. For vibration or acoustic signal triggering timing, , The spacing between the monitoring terminals output by S1 is set. For the speed of signal propagation, when The corresponding features are retained when they are active, otherwise they are discarded; the formula for constructing the multimodal feature matrix is ​​as follows: ,in For Hadamard product operations;

[0123] S5: Edge Risk Pre-judgment: Two heterogeneous models are deployed on the edge nodes involved in S2. Feature processing tasks are allocated based on the computing power characteristics of the edge nodes. The input data includes the visual recognition data output by S3, the multimodal feature data output by S4, and the preliminary intrusion risk judgment result of S3. The lightweight decision model performs a fusion pre-judgment on the input data and outputs a three-level preliminary risk result. The lightweight decision model dynamically adjusts the feature weights to adapt to the risk judgment requirements of different external damage scenarios.

[0124] The feature processing tasks are allocated based on the computing power characteristics of edge nodes, which is specifically achieved through node computing power ratio quantization. The two types of heterogeneous models are the visual feature processing model and the multimodal feature processing model, which respectively process the visual recognition data of S3 and the multimodal feature data of S4.

[0125] Computing power adaptation and task allocation: Defining the node computing power ratio ,in The computing power of a single edge node. The arithmetic mean of the computing power of all edge nodes; when At this time, this node undertakes complex feature extraction and preliminary distillation tasks, matching high computing power, and the task weight formula is as follows: ;when At that time, the node focuses on the core feature screening task, adapting to low computing power scenarios, and the task weight formula is: ,in As the baseline value for task weight, For complex tasks, weight coefficients The core task weighting coefficient;

[0126] Calculation of preliminary risk results at level three: The outputs of the two heterogeneous models are fused using a lightweight decision-making model, as shown in the formula below. ,in , , For feature weights, For bias terms, It is the Sigmoid activation function. This is the preliminary intrusion risk assessment result for S3. A value of 1 indicates a risk, and 0 indicates no risk. , The signal feature vector of S4, This is the visual recognition feature vector of S3;

[0127] Level L1, slight risk. This is a serious risk for L2. Level 3 emergency risk;

[0128] S6: Multi-dimensional Result Fusion and Precise Judgment: Input the three-level preliminary risk results output by S5, the basic deployment data output by S1, and the real-time environmental data from the synchronized multi-source data output by S2. Upload the preliminary risk results to the cloud to construct a dynamic diffusion-type risk field. Integrate high-precision positioning spatial correlation, risk field diffusion intensity, and multi-modal feature confidence for three-dimensional coupling analysis. Access external dynamic data to correct parameters, accurately determine the authenticity and risk level of external damage events, and output the final judgment result. The construction of the dynamic diffusion-type risk field breaks through the limitations of fixed-area risk judgment and can quantify the spatial impact range of external damage risks. The three-dimensional coupling analysis integrates spatial, intensity, and feature information to achieve complementary verification of multi-dimensional information and reduce the probability of misjudgment in single-dimensional judgment.

[0129] The final assessment includes a label indicating the authenticity of the external damage incident and a quantified risk level, as detailed below:

[0130] Dynamic risk field construction: Input the basic deployment data of S1, using its high-precision positioning coordinates As the coordinates of the risk source, output by S5 Mapping the initial strength of risk sources A risk field is constructed based on the Gaussian diffusion model, and the formula for the risk field strength is: ,in Let be the coordinates of any point in the field. The diffusion coefficient is denoted as .

[0131] diffusion coefficient calculate: , The initial diffusion coefficient is... The environmental impact coefficient is calculated from real-time environmental data of S2. Regional risk coefficient;

[0132] 3D Coupling Determination: Confidence of Spatial Correlation in High-Precision Positioning , The distance between the monitoring point of S1 and the cable. To determine the effective distance threshold and the confidence level of the risk field diffusion intensity. Multimodal feature confidence , The multimodal feature matrix of S4; coupling formula , ;

[0133] Corrected by external dynamic data The corrected formula is as follows: , This is a correction amount;

[0134] Accurate Judgment and Output: Authenticity: It is true and marked with T. False and marked with F; Grade correction: Final risk level assessment: Corresponding to L1, Corresponding to L2, Corresponding to L3;

[0135] S7: Tiered Early Warning and Cross-Platform Linkage: Inputs include the risk level from the final judgment result output by S6, the high-precision positioning coordinates from the basic deployment data output by S1, the real-time signal data fragments from the synchronized multi-source data output by S2, and the video screenshots from the visual recognition data output by S3. This triggers a tiered response: minor risks are pushed via a pop-up notification on the monitoring platform, severe risks via SMS and the maintenance app, and emergency risks additionally trigger on-site audible and visual alarms and power dispatch system alarms. Outputs include push notifications containing high-precision positioning coordinates, risk level, real-time signal data fragments, and video screenshots, serving as the basis for on-site handling by maintenance personnel. The handling results serve as input for subsequent data feedback. The tiered response mechanism matches different early warning channels according to the risk level, ensuring that maintenance personnel can prioritize handling high-level events based on the urgency of the risk. The structured encapsulation of push notifications facilitates information parsing and display on different terminals, improving handling efficiency.

[0136] S8: Data Feedback and Model Iteration Optimization: Input the on-site maintenance handling results output by S7, and update the historical external damage sample library and risk field parameter library; the edge terminal responds to S3... With S4 , The original feature matrix formed Compression is performed using the following formula:

[0137]

[0138] in, For S3 output With S4 output , The original feature matrix formed, For convolution kernel weights, For convolution bias, The ReLU activation function is used; features are compressed at each edge in the cloud. Take the average to get Then, the features are optimized using a cloud-based distillation formula, which is:

[0139]

[0140] in, For the weight of the distillation layer, For distillation layer bias, The feature vectors at each edge are compressed; the model optimization uses stochastic gradient descent, and the loss function is:

[0141]

[0142] in, This is the actual label for the external damage incident. For the model prediction results, For weighted regularization terms, The regularization coefficient is used to output the optimized model parameters, which are then fed back to S3, S4, and S5 to achieve dynamic model iteration and form a closed-loop technology.

[0143] The present invention provides an embodiment 2: a 110 kV cable external damage prevention monitoring system, which includes a multi-source sensing network module, an edge processing module, a cloud fusion judgment module, a hierarchical early warning linkage module, and a data feedback optimization module, and each module is connected in sequence for communication;

[0144] The multi-source sensing network module includes several monitoring terminals that integrate high-precision positioning modules. The monitoring terminals integrate vibration sensors, acoustic sensors, and AI vision acquisition units to meet the requirements of high voltage insulation and electromagnetic interference resistance. They are used to perform the monitoring deployment of S1 and the multi-source data acquisition of S2, and output the basic deployment data of S1 and the synchronized multi-source data of S2.

[0145] The edge processing module, deployed on edge nodes, integrates a lightweight visual recognition model, two types of heterogeneous models, and a lightweight decision model. It is used to perform visual analysis and initial risk assessment in S3, multimodal feature extraction in S4, and fusion pre-judgment in S5. It outputs visual recognition data, multimodal feature data, and three-level preliminary risk results to the cloud fusion judgment module, while receiving updated parameters from the data feedback optimization module.

[0146] The cloud-based fusion judgment module is deployed on a cloud server. It has a built-in Gaussian diffusion model and a three-dimensional coupled judgment algorithm. It is used to perform dynamic risk field construction and multi-dimensional fusion judgment of S6, and to access external dynamic data to correct parameters and output the final judgment result of S6.

[0147] The tiered early warning and linkage module includes a monitoring platform, an operation and maintenance APP, on-site audible and visual alarms, and an interface to the power dispatching system. It is used to execute the tiered response and information push of S7 and output the push information of S7.

[0148] The data feedback optimization module is used to record the on-site handling results of S7's operation and maintenance, update the sample library and parameter library, perform iterative optimization of S8's model through the edge-cloud bidirectional distillation algorithm, output the optimized model parameters and feed them back to the edge processing module, forming a technical closed loop.

[0149] This invention provides an embodiment 3, which is applied to a 110 kV cable operation and maintenance project in a suburban area of ​​a city. The cable laying scenario involves a tunnel exit section with a length of 800 meters and an underground direct burial section with a length of 1.2 kilometers. High-risk areas for external damage include: the urban road widening construction area on the north side of the cable and the poplar planting area in the green belt on the south side. The construction area has intensive construction machinery operations, and the trees in the planting area are 8-12 meters tall and are prone to falling in strong winds. The monitoring range covers the entire cable area for 2 kilometers, and it is necessary to achieve accurate monitoring and early warning of two core risks: construction machinery intrusion and tree falling.

[0150] S1 Monitoring Deployment and Sensing Network Construction

[0151] The laying scenarios are clearly defined as laying at the tunnel exit and direct underground burial. The high-risk areas for external damage are designated as a 100m x 30m construction area and a 200m x 50m tree area.

[0152] The monitoring terminal uses an integrated device with a Beidou high-precision positioning module, which integrates MEMS vibration sensors, electret acoustic sensors, and a 2-megapixel AI vision acquisition unit. The device meets the insulation level and electromagnetic interference resistance level of a 110 kV high-voltage environment.

[0153] Layout spacing calculation: Take the baseline value of layout spacing. =50 meters, regional risk level coefficient =1.3, scene complexity coefficient =1.1, substitute into the formula In practice, 60 monitoring terminals were deployed at 35-meter intervals to form a multi-source sensing network covering the entire area.

[0154] Output basic deployment data: Spacing between monitoring terminals =35 meters, high-precision positioning coordinates of each terminal.

[0155] S2 Multi-Source Data Synchronous Acquisition

[0156] Risk coefficient calculation for the target period: The target period is 9:00-18:00 during the day, and the frequency of historical external damage events during the target period. =15 times, average frequency of historical external damage events throughout the entire time period =10 times, get ;

[0157] Data Acquisition Parameter Adjustment: Vibration Signal Sampling Frequency Reference Value =100Hz, adjust gain coefficient =0.8, substitute into the formula Hz; Environmental video stream frame rate benchmark =25fps, similarly fps;

[0158] Timestamp synchronization: Employs the NTP time synchronization protocol to calibrate the time deviation between the sensor and the edge node. ≤1ms, achieving timing consistency of vibration signals, acoustic signals, and video streams;

[0159] Output synchronized multi-source data: vibration signal sampled at 140Hz 35fps environmental video stream, 20-1000Hz acoustic signals, and monitoring terminal location data.

[0160] S3 Visual Risk Target Recognition

[0161] The external damage risk targets are excavators and 10-meter-tall poplar trees. The edge nodes are equipped with a lightweight visual recognition model of YOLOv8-nano.

[0162] Feature quantization: Excavator outline edge segments =20 items, total length =12 meters, area of ​​the circumscribed rectangle =6 square meters, outline features Poplar tree movement trajectory =30 frames, coordinates Fluctuation range ±1.2 meters, trajectory center point Motion trajectory characteristics rice;

[0163] Sensitivity Adjustment: Identify Sensitivity Baseline Value =0.6, sensitivity adjustment coefficient =0.5, substitute into the formula ;

[0164] Track capture weights: Standard track capture weight baseline values =0.4, the wind speed influence coefficient corresponding to real-time wind speed level 3. =0.3, substitute into the formula ;

[0165] Output visual recognition data: contour features =2. Characteristics of motion trajectory =0.8 meters, visual target appearance sequence =10:23:15.2s, visual recognition feature vector 128-dimensional, preliminary intrusion risk assessment results: If the feature similarity exceeds the threshold, it is determined that there is a risk of intrusion.

[0166] S4 Multimodal Signal Feature Extraction

[0167] Multi-scale signal processing: The db4 wavelet packet decomposition algorithm is used to decompose the signal into multiple levels. =3, the decomposition coefficient of the 5th node in the 3rd layer. Through formula Calculate and extract the peak amplitude of the vibration signal (0.8g) and its spectral feature vector. 128-dimensional acoustic signal with a peak spectral density of 60 dB, a duration of 8 seconds, and a spectral feature vector. 128 dimensions;

[0168] Scene adaptation adjustment: Construction area scene adaptation coefficient =1.3, Low-frequency feature weight benchmark value =0.6, characteristic contribution coefficient =0.7, substitute into the formula High-frequency filtering threshold benchmark value =200Hz, noise suppression coefficient =0.8, substitute into the formula Hz;

[0169] Cross-modal timing verification: vibration signal trigger timing =10:23:15.5s, monitoring terminal deployment spacing =35 meters, acoustic signal propagation speed =340m / s, s, substitute into the formula Because the time difference between the movement of construction machinery and the propagation of signals is extremely small, adjustments are needed. For 2 seconds, Determine timing matching;

[0170] Constructing the multimodal feature matrix: Substituting into the formula A 384×1 dimensional multimodal feature matrix was obtained. ;

[0171] Output multimodal feature data: feature matrix Vibration signal spectral feature vector Acoustic signal spectral eigenvectors .

[0172] S5 Edge Risk Pre-assessment

[0173] The heterogeneous models consist of a YOLOv8-nano visual recognition model and a CNN signal processing model, with two edge nodes configured: one for the construction area node and the other for the computing power. =2.6 TOPS, computing power of suburban nodes =1.8 TOPS, average computing power TOPS;

[0174] Computing power ratio and task allocation: computing power ratio Task weight It is responsible for complex feature extraction and preliminary distillation; computing power ratio Task weight Focus on core feature selection;

[0175] Risk prediction: Feature weights =0.3、 =0.4、 =0.3, bias term =0.1, substitute into the formula It was determined to be an emergency risk;

[0176] Output preliminary risk results at level three: Emergency Risk. =0.73.

[0177] S6 Multi-dimensional Result Fusion and Precise Judgment

[0178] Dynamic risk field construction: risk source coordinates =116.400°E, 39.900°N, initial intensity of the risk source =0.9, diffusion coefficient Construction machinery reference diffusion coefficient =8 meters, environmental impact factor =0.2, weighted by wind speed of level 3 and a small amount of precipitation, regional risk coefficient. Substituting into meters, risk field strength ;

[0179] Three-dimensional coupling determination: actual distance between monitoring point and cable =5 meters, effective positioning distance threshold =20 meters, ; ; Weighting coefficient =0.2、 =0.3、 =0.5, substitute into the formula ;

[0180] External data correction: Construction permit validity period: 3 days. =-0.05, Environmental Impact Factor After correction rice, ;

[0181] Level Correction: This is a serious risk. ;

[0182] The final judgment result is: the external damage incident was indeed caused by an excavator illegally approaching the cable, and the risk level is serious.

[0183] S7 Tiered Early Warning and Cross-Platform Collaboration

[0184] Triggering an emergency risk classification response: The monitoring platform sends a pop-up notification, sends SMS and APP alerts to 3 maintenance personnel, and activates 2 on-site audible and visual alarms;

[0185] Push notification: High-precision positioning coordinates (116.402°E, 39.903°N), severe risk level, 5-second vibration signal clip sampled at 140Hz, and screenshot of excavator operation video.

[0186] S8 Data Feedback and Model Iteration Optimization

[0187] Data Update: Record of Handling Results: Excavator operation stopped, no cable damage, incident confirmed. =1, Update the external damage historical sample database: Add 1 emergency risk sample, risk field parameter database: construction area Adjusted to 0.32;

[0188] Feature compression and distillation: Original feature matrix = 128+ 128+ 128 dimensions = 384 dimensions, convolution kernel weights =384×512 dimensions, convolution bias =512 dimensions, substituting into the formula This yields a 512-dimensional compressed feature vector. , Global feature vector in the cloud Distillation layer weight =512×256 dimensions, distillation bias =256 dimensions, substitute into the formula This yields a 256-dimensional lightweight feature vector.

[0189] Model optimization: Regularization coefficient =0.005, loss function ,in The stochastic gradient descent method is used for iterative optimization.

[0190] Feedback application: The optimized model parameters are pushed to all edge nodes for subsequent monitoring tasks.

[0191] In summary, this invention improves the accuracy of risk perception in high-risk scenarios such as densely populated construction areas in urban-rural fringe areas and densely wooded suburban areas. The lightweight visual recognition model in S3 extracts the contour features of external damage risk targets. Characteristics of motion trajectory Time sequence information S4 synchronously receives vibration signals For acoustic signals, wavelet packet decomposition algorithm is used to complete multi-scale signal processing, combined with cross-modal temporal correlation verification formula. Remove isolated features and construct a multimodal feature matrix This multi-dimensional feature fusion method avoids the one-sidedness of single-modality data, makes the feature extraction process more in line with the actual occurrence pattern of external damage risks, effectively solves the problem of misjudgment and omission caused by single features in traditional monitoring, and improves the accuracy of risk target identification and feature extraction.

[0192] To address the uneven distribution of computing power across multiple edge nodes and the need for real-time monitoring, this technology achieves a balance between processing efficiency and judgment accuracy through an edge-cloud collaborative optimization mechanism. In S5, this is based on the node computing power ratio... Dynamically allocate feature processing tasks: nodes with sufficient computing power undertake complex feature extraction and preliminary distillation, while nodes with limited computing power focus on core feature selection, in conjunction with a lightweight decision model using formulas. Quickly output preliminary risk results at level three The S6 cloud platform constructs a dynamic diffusion-type risk field based on a Gaussian diffusion model, and uses a three-dimensional coupling judgment formula. And correct them simultaneously. get This collaborative model achieves accurate judgment by reducing data transmission latency through localized processing at the edge and leveraging the powerful computing power of the cloud to achieve multi-dimensional fusion analysis. It resolves the contradiction between real-time performance and accuracy in traditional monitoring, thereby improving the efficiency and reliability of risk assessment.

[0193] In response to the dynamic nature of external damage risks that change with the environment, this technology achieves dynamic quantification and precise early warning of risks through a dynamic risk field coupling judgment mechanism. The risk field intensity formula in S6... Real-time reflection of risk diffusion status; formula for calculating the risk diffusion coefficient. Incorporating environmental impact coefficient Regional risk coefficient Dynamic factors, combined with high-precision positioning spatial correlation confidence level Confidence level of risk field diffusion intensity and multimodal feature confidence A comprehensive analysis was conducted, and external dynamic data was corrected. The parameters enable dynamic adjustment of the model, ensuring that the risk assessment results remain consistent with the actual environment and risk evolution trend. This provides accurate basis for S7 level early warning, improves the pertinence and effectiveness of cable external damage monitoring, and reduces the threat of external damage events to cable operation.

[0194] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. 110 kV cable external damage monitoring method, characterized in that: Includes the following steps: S1: Monitoring Deployment and Sensing Network Construction: Based on the laying scenario of 110 kV cables and high-risk areas for external damage, monitoring terminals integrating high-precision positioning modules are deployed. The monitoring terminals integrate vibration sensors, acoustic sensors, and AI vision acquisition units to meet the requirements of high-voltage insulation and anti-electromagnetic interference. Edge nodes are deployed simultaneously to establish communication connections with all monitoring terminals, and a multi-source sensing network covering the entire cable area, including monitoring terminals and edge nodes, is constructed. S2: Multi-source data synchronous acquisition: Start the multi-source sensing network built by S1 to acquire multi-source data in real time. The multi-source data includes: vibration signals, acoustic signals, environmental video streams and monitoring terminal location data around the cable. The timestamp synchronization and low-latency transmission of the multi-source data are realized through edge nodes, and the synchronized multi-source data is output. S3: Visual Risk Target Recognition: Deploy a lightweight visual recognition model at the edge nodes involved in S2, input the environmental video stream from the synchronized multi-source data output by S2, perform real-time analysis, extract relevant features and time series information of external damage risk targets, preliminarily determine the target intrusion risk, and output visual recognition data; S4: Multimodal signal feature extraction: Using a multi-scale signal decomposition algorithm, the vibration signal and acoustic signal from the synchronized multi-source data output by S2 and the temporal information from the visual recognition data output by S3 are input. The vibration signal and acoustic signal are processed to extract key parameters, and the temporal information is correlated to perform cross-modal temporal correlation verification. Isolated features are removed, multimodal feature data is constructed, and multimodal feature data is output. S5: Edge Risk Pre-judgment: Two types of heterogeneous models are deployed on the edge nodes involved in S2. Feature processing tasks are allocated based on the computing power characteristics of the edge nodes. The visual recognition data output by S3, the multimodal feature data output by S4, and the preliminary intrusion risk judgment results of S3 are input. The input data are fused and pre-judged through a lightweight decision model, and a three-level preliminary risk result is output. S6: Multi-dimensional result fusion and accurate judgment: Input the three-level preliminary risk results output by S5, the basic deployment data output by S1, and the real-time environmental data from the synchronized multi-source data output by S2. Upload the preliminary risk results to the cloud, construct a dynamic diffusion-type risk field, and perform three-dimensional coupling analysis by integrating high-precision positioning spatial correlation, risk field diffusion intensity, and multi-modal feature confidence. Access external dynamic data to correct parameters, accurately determine the authenticity and risk level of external damage events, and output the final judgment result.

2. The method for monitoring external damage to 110 kV cables according to claim 1, characterized in that: In S1, the laying scenarios include underground direct burial, tunnels, and cable trays, and the high-risk areas for external damage include key areas with dense construction and abundant trees; the basic deployment data includes the spacing between monitoring terminals. And high-precision positioning information, deployment spacing pass: in, To establish the baseline value for the spacing, This represents the regional risk level coefficient. Both are scene complexity coefficients, determined based on historical monitoring data and scene characteristic statistics; This serves as the foundational data for subsequent signal propagation speed calculations and location correlation analysis.

3. The method for monitoring external damage to 110 kV cables according to claim 1, characterized in that: In step S2, the synchronized multi-source data includes the adjusted sampling frequency. Frame rate Vibration signal after timestamp synchronization Acoustic signals, environmental video streams, and monitoring terminal location data are all collected through a multi-source sensing network and synchronously processed by edge nodes; the sampling frequency is dynamically adjusted based on the external damage risk patterns. Frame rate Define the risk coefficient for a given period: in, The frequency of historical external damage events during the target period. The average frequency of historical external damage events over the entire time period; vibration signal sampling frequency adjustment: Environmental video stream frame rate adjustment: in, The sampling frequency reference value for the vibration signal. This serves as the baseline value for the environmental video stream frame rate. To adjust the gain coefficient; timestamp synchronization passed: in, The original acquisition time for each sensor. The time deviation between the sensor and the edge node is obtained by calibration using the time synchronization protocol.

4. The method for monitoring external damage to 110 kV cables according to claim 1, characterized in that: In S3, external damage risk targets include construction machinery and tall trees; visual recognition data includes target outline features. Characteristics of motion trajectory The timing of visual target appearance and visual recognition feature vector All four are generated by analyzing environmental video streams using a lightweight visual recognition model; the quantification of target contour features is achieved through: in, The length of the target contour edge line segment. The area of ​​the target's bounding rectangle; The characteristics of the motion trajectory are calculated by: in, For the goal of Frame coordinates, The coordinates of the center point of the target trajectory. This represents the number of trajectory frames. For integration and The high-dimensional feature vectors are output through the fully connected layer of the model; The timestamp at which the target first appears in the video stream is calibrated by the edge node clock. Combined with the risk coefficient of external damage during the time period Adjust visual recognition sensitivity: in, To identify the sensitivity benchmark, This is the sensitivity adjustment coefficient; The weighting formula for enhancing the capture of the movement trajectories of tall trees is: in, As the baseline value for trajectory capture weights, This refers to the wind speed influence coefficient. Preliminary intrusion risk assessment results are based on Generates by comparing with a preset threshold, when If the feature similarity exceeds the threshold, it is considered to have an intrusion risk; otherwise, it is considered to have no risk.

5. The method for monitoring external damage to 110 kV cables according to claim 1, characterized in that: In step S4, the multimodal feature data includes a multimodal feature matrix. and the spectral feature vector of the disassembled vibration signal Acoustic signal spectral eigenvectors All three are generated through multi-scale signal decomposition and cross-modal verification processing, specifically including: multi-scale signal processing using wavelet packet decomposition algorithm: in, For the first Layer The decomposition coefficients of each node. The vibration signal output by S2 Define wavelet packet basis functions; define scene adaptation coefficients. The low-frequency mechanical vibration characteristic weight adjustment formula is determined based on the statistical analysis of scene signal interference intensity. in, The low-frequency feature weight benchmark value, The feature contribution coefficient is used; the formula for adjusting the high-frequency interference filtering threshold is: in, This is the baseline value for the high-frequency filtering threshold. The noise suppression coefficient is used; the cross-modal temporal correlation verification formula is: in, The timing of the visual target appearance output by S3. For vibration or acoustic signal triggering timing, , The spacing between the monitoring terminals output by S1 is set. For the speed of signal propagation, when Time Retention , Otherwise, discard; the formula for constructing the multimodal feature matrix is: ,in This is for the Hadamard product operation.

6. The method for monitoring external damage to 110 kV cables according to claim 1, characterized in that: In S5, the feature processing task is allocated based on the computing power characteristics of the edge nodes, which is specifically achieved through node computing power ratio quantization. The two heterogeneous models are the visual feature processing model and the multimodal feature processing model, which respectively process the visual recognition data of S3 and the multimodal feature data of S4. Computing power adaptation and task allocation: Defining the node computing power ratio ,in The computing power of a single edge node. The arithmetic mean of the computing power of all edge nodes; when At this time, this node undertakes complex feature extraction and preliminary distillation tasks, matching high computing power, and the task weight formula is as follows: ;when At that time, the node focuses on the core feature screening task, adapting to low computing power scenarios, and the task weight formula is: ,in As the baseline value for task weight, For complex tasks, weight coefficients The core task weighting coefficient; Calculation of preliminary risk results at level three: The outputs of the two heterogeneous models are fused using a lightweight decision-making model, as shown in the formula below. ,in , , For feature weights, For bias terms, It is the Sigmoid activation function. This is the preliminary intrusion risk assessment result for S3. A value of 1 indicates a risk, and 0 indicates no risk. , The signal feature vector of S4, This is the visual recognition feature vector of S3; Level L1, slight risk. This is a serious risk for L2. This is an L3 emergency risk.

7. The method for monitoring external damage to 110 kV cables according to claim 1, characterized in that: In S6, the final determination result includes the authenticity label of the external damage event and the quantitative risk level, as detailed below: Dynamic risk field construction: Input the basic deployment data of S1, using its high-precision positioning coordinates As the coordinates of the risk source, output by S5 Mapping the initial strength of risk sources A risk field is constructed based on the Gaussian diffusion model, and the formula for the risk field strength is: ,in Let be the coordinates of any point in the field. The diffusion coefficient is denoted as . diffusion coefficient calculate: , The initial diffusion coefficient is... The environmental impact coefficient is calculated from real-time environmental data of S2. Regional risk coefficient; 3D Coupling Determination: Confidence of Spatial Correlation in High-Precision Positioning , The distance between the monitoring point of S1 and the cable. To determine the effective distance threshold and the confidence level of the risk field diffusion intensity. Multimodal feature confidence , This is the multimodal feature matrix of S4; Coupling Formula , ; Corrected by external dynamic data The corrected formula is as follows: , This is a correction amount; Accurate Judgment and Output: Authenticity: It is true and marked with T. False and marked with F; Grade correction: Final risk level assessment: Corresponding to L1, Corresponding to L2, Corresponding to L3.

8. The method for monitoring external damage to 110 kV cables according to claim 1, characterized in that: It also includes S7: Tiered early warning and cross-platform linkage: Input the risk level in the final judgment result output by S6, the high-precision positioning coordinates in the basic deployment data output by S1, the real-time signal data fragments in the synchronized multi-source data output by S2, and the video screenshots in the visual recognition data output by S3 to trigger tiered responses. Minor risks are pushed through a pop-up window on the monitoring platform, serious risks are pushed through SMS and the operation and maintenance APP, and emergency risks additionally trigger on-site audible and visual alarms and power dispatching system alarms; the output includes push information containing high-precision positioning coordinates, risk level, real-time signal data fragments and video screenshots, which serve as the basis for operation and maintenance personnel to handle on-site, and the handling results serve as input for subsequent data feedback.

9. The method for monitoring external damage to 110 kV cables according to claim 1, characterized in that: It also includes S8: Data Feedback and Model Iteration Optimization: Inputting the on-site operation and maintenance handling results from S7, updating the historical external damage sample library and risk field parameter library; edge devices respond to S3... With S4 , The original feature matrix formed Compression is performed using the following formula: in, For S3 output With S4 output , The original feature matrix formed, For convolution kernel weights, For convolution bias, The ReLU activation function is used; features are compressed at each edge in the cloud. Take the average to get Then, the features are optimized using a cloud-based distillation formula, which is: in, For the weight of the distillation layer, For distillation layer bias, The feature vectors at each edge are compressed; the model optimization uses stochastic gradient descent, and the loss function is: in, This is the actual label for the external damage incident. For the model prediction results, For weighted regularization terms, The regularization coefficient is used to output the optimized model parameters, which are then fed back to S3, S4, and S5 to achieve dynamic model iteration and form a closed-loop technology. A 10.110 kV cable external damage monitoring system, characterized in that: It includes a multi-source sensing network module, an edge processing module, a cloud-to-cloud fusion judgment module, a hierarchical early warning and linkage module, and a data feedback optimization module, with each module communicating with each other in sequence; The multi-source sensing network module includes several monitoring terminals that integrate high-precision positioning modules. The monitoring terminals integrate vibration sensors, acoustic sensors, and AI vision acquisition units to meet the requirements of high voltage insulation and electromagnetic interference resistance. They are used to perform the monitoring deployment of S1 and the multi-source data acquisition of S2, and output the basic deployment data of S1 and the synchronized multi-source data of S2. The edge processing module, deployed on edge nodes, integrates a lightweight visual recognition model, two types of heterogeneous models, and a lightweight decision model. It is used to perform visual analysis and initial risk assessment in S3, multimodal feature extraction in S4, and fusion pre-judgment in S5. It outputs visual recognition data, multimodal feature data, and three-level preliminary risk results to the cloud fusion judgment module, while receiving updated parameters from the data feedback optimization module. The cloud-based fusion judgment module is deployed on a cloud server. It has a built-in Gaussian diffusion model and a three-dimensional coupled judgment algorithm. It is used to perform dynamic risk field construction and multi-dimensional fusion judgment of S6, and to access external dynamic data to correct parameters and output the final judgment result of S6. The tiered early warning and linkage module includes a monitoring platform, an operation and maintenance APP, on-site audible and visual alarms, and an interface to the power dispatching system. It is used to execute the tiered response and information push of S7 and output the push information of S7. The data feedback optimization module is used to record the on-site handling results of S7's operation and maintenance, update the sample library and parameter library, perform iterative optimization of S8's model through the edge-cloud bidirectional distillation algorithm, output the optimized model parameters and feed them back to the edge processing module, forming a technical closed loop.

Images