Intelligent security remote inspection and control method based on AI

Abstract

The invention relates to the technical field of security management, in particular to an AI-based intelligent security remote inspection and control method, which comprises the steps of front-end intelligent equipment deployment, intelligent inspection execution, data processing and intelligent analysis and remote control and linkage response. Compared with the defects of incomplete information and high false alarm rate due to the fact that equipment appearance anomaly detection mainly depends on single-modal data in the prior art, the scheme adopts multi-modal data preprocessing and feature extraction, constructs a cross-modal alignment network architecture, dynamically fuses laser radar, RGB-D images and IMU data by using a self-attention mechanism, and improves the detection accuracy of the equipment appearance anomaly. According to the method, more accurate equipment appearance abnormity identification is realized by combining environment self-adaptive adjustment, and the model is compressed and deployed to an edge end through a knowledge distillation technology, so that real-time reasoning and analysis are realized, the intelligent level of security inspection and the accuracy of abnormity detection are remarkably improved, the false alarm rate is reduced, and the practicability and reliability of the system are enhanced.

Description

Technical Field

[0001] The present invention relates to the field of security management technology, and in particular to an AI-based intelligent security remote inspection and control method. Background Art

[0002] In modern security systems, as safety demands continue to rise, traditional inspection methods are no longer able to meet the demands for efficiency and accuracy. Manual inspections are not only time-consuming and labor-intensive, but also susceptible to human factors, leading to frequent problems such as missed inspections and false alarms. Therefore, the development of AI-based intelligent security remote inspection and control methods is particularly important.

[0003] Currently, device appearance anomaly detection primarily relies on single-modality data, such as images or videos. This approach is often limited by information integrity and struggles to fully and accurately reflect the device's true state. To overcome this limitation, it is necessary to fuse data from multiple modalities, such as LiDAR, RGB-D images, and IMU data, to obtain more comprehensive device information. However, data from different modalities differ in format, characteristics, and spatiotemporal distribution. Therefore, effective preprocessing, feature extraction, and alignment fusion present a pressing technical challenge.

[0004] Furthermore, real-time security inspections are crucial. Traditional analytical methods often fail to meet the demands of rapid response. Therefore, it is necessary to develop a method that can perform real-time reasoning and analysis at the edge. This requires algorithmic models to be both highly accurate and highly efficient, enabling fast and accurate anomaly detection within limited computing resources. Summary of the Invention

[0005] In order to overcome the problems raised in the above background technology, the present invention proposes an AI-based intelligent security remote inspection and control method.

[0006] The technical solution of the present invention is: an AI-based intelligent security remote inspection and control method, comprising the following steps: S11: Front-end intelligent equipment deployment, including data collection equipment, inspection robots, and supporting facilities for inspection robots; S12: Intelligent inspection execution, controlling the inspection robot to autonomously navigate along a preset training route, and performing real-time data collection and preliminary analysis; S13: Data processing and intelligent analysis: edge computing nodes are used to process and intelligently analyze the data collected by inspection robots and data acquisition devices. S14: Remote control and linkage response. After receiving an abnormal alarm, the monitoring center sends remote control instructions to the AI ​​inspection equipment through the communication network.

[0007] Preferably, when deploying front-end intelligent devices, the deployed devices specifically include: A11: Environmental data collection equipment, including lidar and high-definition cameras, is used to collect geographic environmental data within the inspection range; A12: Equipment monitoring equipment, including vibration sensors and smart meters, is used to collect equipment operating data within the inspection range; A13: Inspection robot, used to conduct inspections within the inspection range according to the set inspection route and manual instructions; A14: The supporting facilities of the inspection robot include a charging base and a signal base station, which are used to provide wireless communication network and charging services for the inspection robot during the inspection process.

[0008] Preferably, when controlling the inspection robot to autonomously navigate according to a preset training route and to collect and perform preliminary data analysis in real time, the following steps are specifically included: S21: Load and initialize the inspection route. Import the preset inspection route into the inspection robot's navigation system, build the inspection task sequence based on the high-precision map, and perform equipment self-test. S22: Environmental perception and real-time positioning, which collects environmental data through the environmental perception equipment and positioning equipment on the inspection robot; S23: Dynamic path planning and obstacle avoidance: real-time processing and analysis of data collected by environmental perception devices and positioning devices, and real-time obstacle avoidance and dynamic path planning; S24: Multimodal data synchronous collection, which is performed by the inspection equipment carried by the inspection robot. The inspection equipment includes a laser radar, a high-definition camera, a vibration sensor, and an infrared camera. S25: Edge-side data preprocessing and preliminary analysis: The edge computing nodes process and analyze the collected data in real time to identify abnormal appearance and temperature of the equipment. S26: Real-time data transmission and recording: uploading the collected data and the calculation results of the edge computing node in real time, and recording and storing them; S27: Manual command response and task switching, using a hierarchical alarm mechanism to report abnormal data in real time, and reconstructing the task queue through ROS when receiving tasks issued by the remote inspection center.

[0009] Preferably, when a hierarchical alarm mechanism is used to report abnormal data in real time, the rules of the hierarchical alarm mechanism are: A11: Low-risk anomalies are recorded in local storage and uploaded in batches after the task is completed. Low-risk anomalies include instrument reading deviations. A12: For high-risk events, they are immediately pushed to the monitoring center via the 5G slicing network, triggering audible and visual alarms. High-risk anomalies include intrusion behavior and sudden temperature rise of the device.

[0010] Preferably, when the edge computing node processes and analyzes the collected data in real time to identify abnormalities in the appearance of the device, the following steps are specifically included: S31: Multimodal data preprocessing and feature extraction. First, PTP is used to synchronize the lidar, high-definition camera, and IMU to eliminate timestamp deviations. An extrinsic calibration matrix is ​​then used to unify all coordinate systems to the inspection robot's coordinate system. Feature extraction is then performed. For lidar data, PointNet++ is used to extract local geometric features and generate 128-dimensional feature vectors. For high-definition camera data, ResNet-18 is used to extract semantic features from RGB images. These features are then combined with the 3D coordinates of the depth map to encode 256-dimensional features. For IMU data, sliding window filtering is performed on the accelerometer and gyroscope signals to extract time-frequency domain features. S32: Cross-modal alignment network architecture design, building independent encoder branches for LiDAR, RGB-D image, and IMU data, extracting feature vectors from each modality, and using self-attention to calculate inter-modal correlation weights. It dynamically fuses multi-source data, generates dynamic weight coefficients based on real-time light intensity and motion speed, and performs environmental adaptive adjustments through a gating mechanism. S33: Training and optimization strategy, training the constructed network model, wherein the training adopted is multi-scenario adversarial training; S34: Edge deployment and real-time inference: Use knowledge distillation to compress the constructed network model into a tiny version and deploy it inside the inspection robot. It processes and analyzes the collected data in real time to identify abnormal appearance and temperature of the equipment.

[0011] As a preference, when using self-attention to calculate the inter-modality correlation weights and dynamically fuse multi-source data, the principle formula is: ; Among them, Q, K and V are the projection matrices obtained by linear transformation of lidar, RGB-D image and IMU data respectively, and dk is a hyperparameter. .

[0012] Preferably, when real-time processing and analysis of data collected by the environmental perception device and the positioning device are performed, and real-time obstacle avoidance and dynamic path planning are performed, the following steps are specifically included: S41: Modeling and decomposing multi-objective problems. Path planning objectives are quantified into three objective functions: path length, risk level, and task timeliness. The Tchebycheff decomposition method is used to dynamically assign weights to ensure that the trade-offs between objectives can be calculated. The principle formula of the Tchebycheff decomposition method is: ; in, is a uniformly distributed weight vector, where , is the current ideal reference point, is the Tchebycheff function, x is the decision variable, representing the path plan to be optimized, Represents the weight vector of the jth target, where j=1,2,3. In , when j = 1, 2, and 3, they correspond to the quantitative values ​​of path length, risk level, and task timeliness, respectively; S42: Initialize the weight vector and neighborhood structure. Generate an initial solution set covering the Pareto front based on the uniformly distributed weight vector, and build a neighborhood relationship through Euclidean distance so that adjacent subproblems can share optimization information. S43: Evolutionary operations and dynamic updates, combining crossover and mutation operations to generate new path solutions, using the Tchebycheff function to evaluate their performance, dynamically updating the population and ideal reference point to ensure that the algorithm continues to converge to a better Pareto frontier area; S44: Pareto optimal path set generation, maintaining external archives to retain non-dominated solutions, and screening evenly distributed path sets through congestion sorting or clustering methods to ensure solution diversity and global optimality; S45: Dynamic environment adaptive optimization, which adjusts the weight vector distribution and neighborhood structure in real time according to environmental changes, prioritizes key objectives, and enhances the algorithm's robustness to dynamic scenarios; S46: Output and Application, maps the path set in EP to the three-dimensional target space, displays the trade-off relationship between different targets, provides an interactive decision interface for manual selection of the optimal path, and verifies the algorithm performance through hypervolume and reverse generation distance indicators.

[0013] Preferably, when performing real-time processing and analysis on the data collected by the environmental perception device and the positioning device, and performing real-time obstacle avoidance and dynamic path planning, it also includes establishing an energy consumption prediction model, using the energy consumption prediction model to predict the energy consumption data of the patrol robot, and modifying the patrol path according to the prediction results of the energy consumption prediction model and the position of the charging station.

[0014] Preferably, when establishing an energy consumption prediction model and using the energy consumption prediction model to predict the energy consumption data of the inspection robot, the following steps are specifically included: S51: Data preparation and feature engineering, integrating historical energy consumption data with terrain features, processing continuous and categorical variables through standardization and one-hot encoding, and constructing a sliding window to generate a time series-static feature fusion dataset; S52: LSTM network architecture design. This involves designing a network structure consisting of an input layer, multiple layers of LSTM units, and a fully connected layer. The input layer integrates time-series energy consumption data with static terrain features, uses the LSTM gating mechanism to capture long-term dependencies, and ultimately outputs the predicted energy consumption per unit distance through the fully connected layer. S53: Model training and tuning, using mean squared error as the loss function, combined with the Adam optimizer to adaptively adjust the learning rate, dividing the model into training sets, validation sets, and test sets for training, and dynamically stopping the overfitting trend through early stopping; S54: Prediction and performance verification: Generate energy consumption prediction values ​​under different terrain parameters based on the test set, calculate indicators to quantify model accuracy, and visually verify the model's fitting effect on complex terrain energy consumption patterns by comparing the predicted curve with the actual value; S55: Dynamic environmental adaptability, introducing an incremental learning mechanism to fine-tune the model weights online when new terrain data is added or energy consumption patterns change.

[0015] Preferably, when processing and intelligently analyzing the data collected by the inspection robot and the data acquisition device through the edge computing node, the following steps are specifically included: S61: Data reception and preprocessing: receiving data from the inspection robot through a standardized communication protocol and preprocessing the data; S62: Marking of key data: reading and identifying the results of the inspection robot's preliminary analysis, and marking abnormal data; S63: Intelligent analysis, using AI models to analyze and identify abnormal data in real time; S64: Real-time response and decision-making, judging the abnormality level based on the analysis results of the AI ​​model, and triggering a graded response based on the abnormality level, where the abnormality level includes low-risk events, medium-risk events, and high-risk events.

[0016] Beneficial effects of the present invention: 1. Compared with existing technologies that mainly rely on single-modal data for device appearance anomaly detection, which has the disadvantages of incomplete information and high false alarm rate, this solution adopts multimodal data preprocessing and feature extraction. By building a cross-modal alignment network architecture, it uses the self-attention mechanism to dynamically fuse lidar, RGB-D image and IMU data, combined with environmental adaptive adjustment, to achieve more accurate device appearance anomaly recognition. The model is compressed and deployed to the edge through knowledge distillation technology, realizing real-time reasoning and analysis. This significantly improves the intelligence level of security inspections and the accuracy of anomaly detection, reduces the false alarm rate, and enhances the practicality and reliability of the system. 2. Compared with existing technologies that mainly rely on static path planning methods and lack flexible response to environmental changes and target diversity, resulting in suboptimal and poorly adaptable path planning results, this solution uses multi-objective problem modeling and Tchebycheff decomposition method, combined with evolutionary algorithms for dynamic path planning. By adjusting weight vectors and neighborhood structures in real time, it achieves comprehensive optimization of path length, risk level, and task timeliness. This solution can generate a Pareto optimal path set and provide an interactive decision-making interface. This not only enhances the intelligence and flexibility of path planning, but also significantly improves the algorithm's adaptability to dynamic environments and the global optimality of the solution, providing a more efficient, safe, and reliable path planning solution for security inspections. 3. Compared with the existing technology that mainly relies on empirical formulas or simple statistical models for energy consumption prediction, it lacks the accurate capture of different terrain characteristics and dynamic energy consumption patterns, resulting in insufficient precision and adaptability in energy consumption management in path planning. This solution adopts the LSTM network architecture to establish an energy consumption prediction model. By integrating time-series energy consumption data with static terrain characteristics, it captures long-term dependencies and achieves accurate prediction of energy consumption per unit distance. Combined with the incremental learning mechanism, it dynamically adapts to new terrain data or changes in energy consumption patterns. This solution not only improves the accuracy and robustness of energy consumption prediction, but also can intelligently adjust the inspection path according to the prediction results and the location of the charging station, effectively optimize the energy consumption management of the inspection robot, extend the duration of a single inspection task, and improve the overall inspection efficiency and energy utilization rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 Shown is a flow chart of the AI-based intelligent security remote inspection and control method of the present invention; Figure 2 Shown is a schematic diagram of part of the workflow of intelligent inspection execution in the AI-based intelligent security remote inspection and control method of the present invention. DETAILED DESCRIPTION

[0018] The present invention will be further described below with reference to the accompanying drawings and examples.

[0019] See also Figure 1-2The present invention provides an embodiment: an AI-based intelligent security remote inspection and control method, comprising the following steps: S11: Front-end intelligent equipment deployment, including data collection equipment, inspection robots, and supporting facilities for inspection robots; S12: Intelligent inspection execution, controlling the inspection robot to autonomously navigate along a preset training route, and performing real-time data collection and preliminary analysis; S13: Data processing and intelligent analysis: edge computing nodes are used to process and intelligently analyze the data collected by inspection robots and data acquisition devices. S14: Remote control and linkage response. After receiving an abnormal alarm, the monitoring center sends remote control instructions to the AI ​​inspection equipment through the communication network.

[0020] As described above, the present invention significantly improves the efficiency and accuracy of security inspections by deploying front-end intelligent devices, realizing autonomous navigation inspections and real-time data analysis, utilizing edge computing for efficient processing and intelligent identification, and supporting remote control and rapid linkage response, thereby achieving immediate discovery and effective control of potential security threats.

[0021] Preferably, when deploying front-end intelligent devices, the deployed devices specifically include: A11: Environmental data collection equipment, including lidar and high-definition cameras, is used to collect geographic environmental data within the inspection range; A12: Equipment monitoring equipment, including vibration sensors and smart meters, is used to collect equipment operating data within the inspection range; A13: Inspection robot, used to conduct inspections within the inspection range according to the set inspection route and manual instructions; A14: The supporting facilities of the inspection robot include a charging base and a signal base station, which are used to provide wireless communication network and charging services for the inspection robot during the inspection process.

[0022] As described above, the present invention integrates environmental data acquisition equipment (lidar, high-definition cameras) to accurately capture geographic environmental information, and equipment monitoring equipment (vibration sensors, smart meters) to monitor the equipment operating status in real time. It is combined with inspection robots to autonomously and efficiently perform inspection tasks, and supplemented by charging stations and signal base stations to ensure worry-free communication and battery life. This solution comprehensively improves the intelligence level and comprehensive monitoring capabilities of security inspections, ensuring the comprehensiveness and continuity of inspection work.

[0023] Preferably, when controlling the inspection robot to autonomously navigate according to a preset training route and to collect and perform preliminary data analysis in real time, the following steps are specifically included: S21: Load and initialize the inspection route. Import the preset inspection route into the inspection robot's navigation system, build the inspection task sequence based on the high-precision map, and perform equipment self-test. S22: Environmental perception and real-time positioning, which collects environmental data through the environmental perception equipment and positioning equipment on the inspection robot; S23: Dynamic path planning and obstacle avoidance: real-time processing and analysis of data collected by environmental perception devices and positioning devices, and real-time obstacle avoidance and dynamic path planning; S24: Multimodal data synchronous collection, which is performed by the inspection equipment carried by the inspection robot. The inspection equipment includes a laser radar, a high-definition camera, a vibration sensor, and an infrared camera. S25: Edge-side data preprocessing and preliminary analysis: The edge computing nodes process and analyze the collected data in real time to identify abnormal appearance and temperature of the equipment. S26: Real-time data transmission and recording: uploading the collected data and the calculation results of the edge computing node in real time, and recording and storing them; S27: Manual command response and task switching, using a hierarchical alarm mechanism to report abnormal data in real time, and reconstructing the task queue through ROS when receiving tasks issued by the remote inspection center.

[0024] As described above, the present invention loads a preset route and initializes it through a patrol robot, combines high-precision maps with environmental perception equipment to achieve real-time positioning and obstacle avoidance, synchronously collects multimodal data, uses edge computing nodes for real-time preprocessing and preliminary analysis to identify anomalies, transmits and records data in real time and issues graded alarms, and flexibly switches tasks in response to remote commands, greatly improving patrol efficiency and accuracy and ensuring the comprehensiveness and response speed of security patrols.

[0025] Preferably, when a hierarchical alarm mechanism is used to report abnormal data in real time, the rules of the hierarchical alarm mechanism are: A11: Low-risk anomalies are recorded in local storage and uploaded in batches after the task is completed. Low-risk anomalies include instrument reading deviations. A12: For high-risk events, they are immediately pushed to the monitoring center via the 5G slicing network, triggering audible and visual alarms. High-risk anomalies include intrusion behavior and sudden temperature rise of the device.

[0026] As described above, the present invention adopts a hierarchical alarm mechanism to locally record and batch upload low-risk anomalies such as instrument reading deviations, effectively managing data volume. For high-risk events such as intrusions and sudden temperature rises in equipment, these are immediately pushed to the monitoring center at high speed via the 5G slicing network and trigger audible and visual alarms, ensuring a rapid response and handling of emergencies, significantly improving the timeliness and effectiveness of security monitoring.

[0027] Preferably, when the edge computing node processes and analyzes the collected data in real time to identify abnormalities in the appearance of the device, the following steps are specifically included: S31: Multimodal data preprocessing and feature extraction. First, PTP is used to synchronize the lidar, high-definition camera, and IMU to eliminate timestamp deviations. An extrinsic calibration matrix is ​​then used to unify all coordinate systems to the inspection robot's coordinate system. Feature extraction is then performed. For lidar data, PointNet++ is used to extract local geometric features and generate 128-dimensional feature vectors. For high-definition camera data, ResNet-18 is used to extract semantic features from RGB images. These features are then combined with the 3D coordinates of the depth map to encode 256-dimensional features. For IMU data, sliding window filtering is performed on the accelerometer and gyroscope signals to extract time-frequency domain features. S32: Cross-modal alignment network architecture design, building independent encoder branches for LiDAR, RGB-D image, and IMU data, extracting feature vectors from each modality, and using self-attention to calculate inter-modal correlation weights. It dynamically fuses multi-source data, generates dynamic weight coefficients based on real-time light intensity and motion speed, and performs environmental adaptive adjustments through a gating mechanism. S33: Training and optimization strategy, training the constructed network model, wherein the training adopted is multi-scenario adversarial training; S34: Edge deployment and real-time inference: Use knowledge distillation to compress the constructed network model into a tiny version and deploy it inside the inspection robot. It processes and analyzes the collected data in real time to identify abnormal appearance and temperature of the equipment.

[0028] As a preference, when using self-attention to calculate the inter-modality correlation weights and dynamically fuse multi-source data, the principle formula is: ; Among them, Q, K and V are the projection matrices obtained by linear transformation of lidar, RGB-D image and IMU data respectively, and dk is a hyperparameter. .

[0029] As described above, compared to existing technologies that primarily rely on single-modal data for device appearance anomaly detection, which suffers from incomplete information and high false alarm rates, this solution utilizes multimodal data preprocessing and feature extraction. By constructing a cross-modal alignment network architecture, this solution dynamically fuses lidar, RGB-D imagery, and IMU data using a self-attention mechanism, combined with environmental adaptive adjustment, to achieve more accurate device appearance anomaly identification. Furthermore, knowledge distillation technology compresses the model and deploys it to the edge, enabling real-time reasoning and analysis. This significantly improves the intelligence level of security patrols and the accuracy of anomaly detection, reduces false alarm rates, and enhances the system's practicality and reliability.

[0030] Preferably, when real-time processing and analysis of data collected by the environmental perception device and the positioning device are performed, and real-time obstacle avoidance and dynamic path planning are performed, the following steps are specifically included: S41: Modeling and decomposing multi-objective problems. Path planning objectives are quantified into three objective functions: path length, risk level, and task timeliness. The Tchebycheff decomposition method is used to dynamically assign weights to ensure that the trade-offs between objectives can be calculated. The principle formula of the Tchebycheff decomposition method is: ; in, is a uniformly distributed weight vector, where , is the current ideal reference point, is the Tchebycheff function, x is the decision variable, representing the path plan to be optimized, Represents the weight vector of the jth target, where j=1,2,3. In , when j = 1, 2, and 3, they correspond to the quantitative values ​​of path length, risk level, and task timeliness, respectively; S42: Initialize the weight vector and neighborhood structure. Generate an initial solution set covering the Pareto front based on the uniformly distributed weight vector, and build a neighborhood relationship through Euclidean distance so that adjacent subproblems can share optimization information. S43: Evolutionary operations and dynamic updates, combining crossover and mutation operations to generate new path solutions, using the Tchebycheff function to evaluate their performance, dynamically updating the population and ideal reference point to ensure that the algorithm continues to converge to a better Pareto frontier area; S44: Pareto optimal path set generation, maintaining external archives to retain non-dominated solutions, and screening evenly distributed path sets through congestion sorting or clustering methods to ensure solution diversity and global optimality; S45: Dynamic environment adaptive optimization, which adjusts the weight vector distribution and neighborhood structure in real time according to environmental changes, prioritizes key objectives, and enhances the algorithm's robustness to dynamic scenarios; S46: Output and Application, maps the path set in EP to the three-dimensional target space, displays the trade-off relationship between different targets, provides an interactive decision interface for manual selection of the optimal path, and verifies the algorithm performance through hypervolume and reverse generation distance indicators.

[0031] As described above, compared to the prior art, which primarily relies on static path planning methods and lacks the flexibility to respond to environmental changes and target diversity, the present invention often results in suboptimal and poorly adaptable path planning results. This solution employs multi-objective problem modeling and the Tchebycheff decomposition method, combined with an evolutionary algorithm for dynamic path planning. By adjusting weight vectors and neighborhood structures in real time, it achieves comprehensive optimization of path length, risk level, and task timeliness. This solution can generate a Pareto-optimal path set and provide an interactive decision-making interface. This not only enhances the intelligence and flexibility of path planning, but also significantly improves the algorithm's adaptability to dynamic environments and the global optimality of the solution, providing a more efficient, safe, and reliable path planning solution for security inspections.

[0032] Preferably, when performing real-time processing and analysis on the data collected by the environmental perception device and the positioning device, and performing real-time obstacle avoidance and dynamic path planning, it also includes establishing an energy consumption prediction model, using the energy consumption prediction model to predict the energy consumption data of the patrol robot, and modifying the patrol path according to the prediction results of the energy consumption prediction model and the position of the charging station.

[0033] Preferably, when establishing an energy consumption prediction model and using the energy consumption prediction model to predict the energy consumption data of the inspection robot, the following steps are specifically included: S51: Data preparation and feature engineering, integrating historical energy consumption data with terrain features, processing continuous and categorical variables through standardization and one-hot encoding, and constructing a sliding window to generate a time series-static feature fusion dataset; S52: LSTM network architecture design. This involves designing a network structure consisting of an input layer, multiple layers of LSTM units, and a fully connected layer. The input layer integrates time-series energy consumption data with static terrain features, uses the LSTM gating mechanism to capture long-term dependencies, and ultimately outputs the predicted energy consumption per unit distance through the fully connected layer. S53: Model training and tuning, using mean squared error as the loss function, combined with the Adam optimizer to adaptively adjust the learning rate, dividing the model into training sets, validation sets, and test sets for training, and dynamically stopping the overfitting trend through early stopping; S54: Prediction and performance verification: Generate energy consumption prediction values ​​under different terrain parameters based on the test set, calculate indicators to quantify model accuracy, and visually verify the model's fitting effect on complex terrain energy consumption patterns by comparing the predicted curve with the actual value; S55: Dynamic environmental adaptability, introducing an incremental learning mechanism to fine-tune the model weights online when new terrain data is added or energy consumption patterns change.

[0034] As described above, compared with the existing technology, the present invention mainly relies on empirical formulas or simple statistical models for energy consumption prediction, which lacks the accurate capture of different terrain characteristics and dynamic energy consumption patterns, resulting in insufficient refinement and adaptability of energy consumption management in path planning. This solution adopts an LSTM network architecture to establish an energy consumption prediction model. By fusing time-series energy consumption data with static terrain characteristics, it captures long-term dependencies and achieves accurate prediction of energy consumption per unit distance. It is combined with an incremental learning mechanism to dynamically adapt to new terrain data or changes in energy consumption patterns. This solution not only improves the accuracy and robustness of energy consumption prediction, but also can intelligently adjust the inspection path according to the prediction results and the location of the charging station, effectively optimize the energy consumption management of the inspection robot, extend the duration of a single inspection task, and improve the overall inspection efficiency and energy utilization.

[0035] Preferably, when processing and intelligently analyzing the data collected by the inspection robot and the data acquisition device through the edge computing node, the following steps are specifically included: S61: Data reception and preprocessing: receiving data from the inspection robot through a standardized communication protocol and preprocessing the data; S62: Marking of key data: reading and identifying the results of the inspection robot's preliminary analysis, and marking abnormal data; S63: Intelligent analysis, using AI models to analyze and identify abnormal data in real time; S64: Real-time response and decision-making, judging the abnormality level based on the analysis results of the AI ​​model, and triggering a graded response based on the abnormality level, where the abnormality level includes low-risk events, medium-risk events, and high-risk events.

[0036] As described above, the present invention receives and pre-processes data from patrol robots and data acquisition equipment through edge computing nodes, focuses on marking abnormal data and uses AI models for real-time intelligent analysis. It can accurately judge the abnormality level and trigger graded responses. This technical solution effectively improves the timeliness of data processing and the accuracy of intelligent analysis, providing security patrols with faster and more accurate abnormality identification and response capabilities.

[0037] The embodiments of the present invention are described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Various changes can be made within the scope of knowledge of those skilled in the art without departing from the spirit of the present invention.

Claims

1. AI-based intelligent security remote inspection and control method, characterized by: The following steps are included: S11: Front-end intelligent equipment deployment, including data collection equipment, inspection robots, and supporting facilities for inspection robots; S12: Intelligent inspection execution, controlling the inspection robot to autonomously navigate according to the preset training route, and performing real-time data collection and preliminary analysis; S13: Data processing and intelligent analysis: edge computing nodes are used to process and intelligently analyze the data collected by inspection robots and data acquisition devices. S14: Remote control and linkage response. After receiving an abnormal alarm, the monitoring center sends remote control instructions to the AI ​​inspection equipment through the communication network.

2. The AI-based intelligent security remote inspection and control method according to claim 1 is characterized by: When deploying front-end intelligent devices, the deployed devices include: A11: Environmental data collection equipment, including lidar and high-definition cameras, is used to collect geographic environmental data within the inspection range; A12: Equipment monitoring equipment, including vibration sensors and smart meters, is used to collect equipment operating data within the inspection range; A13: Inspection robot, used to conduct inspections within the inspection range according to the set inspection route and manual instructions; A14: The supporting facilities of the inspection robot include a charging base and a signal base station, which are used to provide wireless communication network and charging services for the inspection robot during the inspection process.

3. The AI-based intelligent security remote inspection and control method according to claim 2 is characterized by: When controlling the inspection robot to autonomously navigate according to the preset training route and collect and perform preliminary data analysis in real time, it specifically includes: S21: Load and initialize the inspection route. Import the preset inspection route into the inspection robot's navigation system, build the inspection task sequence based on the high-precision map, and perform equipment self-test. S22: Environmental perception and real-time positioning, which collects environmental data through the environmental perception equipment and positioning equipment on the inspection robot; S23: Dynamic path planning and obstacle avoidance: real-time processing and analysis of data collected by environmental perception devices and positioning devices, and real-time obstacle avoidance and dynamic path planning; S24: Multimodal data synchronous collection, which is performed by the inspection equipment carried by the inspection robot. The inspection equipment includes a laser radar, a high-definition camera, a vibration sensor, and an infrared camera. S25: Edge-side data preprocessing and preliminary analysis: The edge computing nodes process and analyze the collected data in real time to identify abnormal appearance and temperature of the equipment. S26: Real-time data transmission and recording: uploading the collected data and the calculation results of the edge computing node in real time, and recording and storing them; S27: Manual command response and task switching, using a hierarchical alarm mechanism to report abnormal data in real time, and reconstructing the task queue through ROS when receiving tasks issued by the remote inspection center.

4. The AI-based intelligent security remote inspection and control method according to claim 3 is characterized by: When using a hierarchical alarm mechanism to report abnormal data in real time, the rules of the hierarchical alarm mechanism are as follows: A11: Low-risk anomalies are recorded in local storage and uploaded in batches after the task is completed. Low-risk anomalies include instrument reading deviations. A12: For high-risk events, they are immediately pushed to the monitoring center via the 5G slicing network, triggering audible and visual alarms. High-risk anomalies include intrusion behavior and sudden temperature rise of the device.

5. The AI-based intelligent security remote inspection and control method according to claim 4 is characterized in that: When edge computing nodes process and analyze collected data in real time to identify abnormal device appearance, the following steps are specifically performed: S31: Multimodal data preprocessing and feature extraction. First, PTP is used to synchronize the lidar, high-definition camera, and IMU to eliminate timestamp deviations. An extrinsic calibration matrix is ​​then used to unify all coordinate systems to the inspection robot's coordinate system. Feature extraction is then performed. For lidar data, PointNet++ is used to extract local geometric features and generate 128-dimensional feature vectors. For high-definition camera data, ResNet-18 is used to extract semantic features from RGB images. These features are then combined with the 3D coordinates of the depth map to encode 256-dimensional features. For IMU data, sliding window filtering is performed on the accelerometer and gyroscope signals to extract time-frequency domain features. S32: Cross-modal alignment network architecture design, building independent encoder branches for LiDAR, RGB-D image, and IMU data, extracting feature vectors from each modality, and using self-attention to calculate inter-modal correlation weights. It dynamically fuses multi-source data, generates dynamic weight coefficients based on real-time light intensity and motion speed, and performs environmental adaptive adjustments through a gating mechanism. S33: Training and optimization strategy, training the constructed network model, wherein the training adopted is multi-scenario adversarial training; S34: Edge deployment and real-time inference: Use knowledge distillation to compress the constructed network model into a tiny version and deploy it inside the inspection robot. It processes and analyzes the collected data in real time to identify abnormal appearance and temperature of the equipment.

6. The AI-based intelligent security remote inspection and control method according to claim 5 is characterized by: When using self-attention to calculate the correlation weights between modalities and dynamically fuse multi-source data, the principle formula is: ； Among them, Q, K and V are the projection matrices obtained by linear transformation of lidar, RGB-D image and IMU data respectively, and dk is a hyperparameter. .

7. The AI-based intelligent security remote inspection and control method according to claim 6 is characterized by: When processing and analyzing data collected by environmental perception devices and positioning devices in real time, and performing real-time obstacle avoidance and dynamic path planning, it specifically includes: S41: Modeling and decomposing multi-objective problems. Path planning objectives are quantified into three objective functions: path length, risk level, and task timeliness. The Tchebycheff decomposition method is used to dynamically assign weights to ensure that the trade-offs between objectives can be calculated. The principle formula of the Tchebycheff decomposition method is: ； in, is a uniformly distributed weight vector, where , is the current ideal reference point, is the Tchebycheff function, x is the decision variable, representing the path plan to be optimized, Represents the weight vector of the jth target, where j=1,2,3. In , when j = 1, 2, and 3, they correspond to the quantitative values ​​of path length, risk level, and task timeliness, respectively; S42: Initialize the weight vector and neighborhood structure. Generate an initial solution set covering the Pareto front based on the uniformly distributed weight vector, and build a neighborhood relationship through Euclidean distance so that adjacent subproblems can share optimization information. S43: Evolutionary operations and dynamic updates, combining crossover and mutation operations to generate new path solutions, using the Tchebycheff function to evaluate their performance, dynamically updating the population and ideal reference point to ensure that the algorithm continues to converge to a better Pareto frontier area; S44: Pareto optimal path set generation, maintaining external archives to retain non-dominated solutions, and screening evenly distributed path sets through congestion sorting or clustering methods to ensure solution diversity and global optimality; S45: Dynamic environment adaptive optimization, which adjusts the weight vector distribution and neighborhood structure in real time according to environmental changes, prioritizes key objectives, and enhances the algorithm's robustness to dynamic scenarios; S46: Output and Application, maps the path set in EP to the three-dimensional target space, displays the trade-off relationship between different targets, provides an interactive decision interface for manual selection of the optimal path, and verifies the algorithm performance through hypervolume and reverse generation distance indicators.

8. The AI-based intelligent security remote inspection and control method according to claim 7 is characterized by: When performing real-time processing and analysis on the data collected by the environmental perception equipment and positioning equipment, and performing real-time obstacle avoidance and dynamic path planning, it also includes establishing an energy consumption prediction model, using the energy consumption prediction model to predict the energy consumption data of the patrol robot, and modifying the patrol path according to the prediction results of the energy consumption prediction model and the location of the charging station.

9. The AI-based intelligent security remote inspection and control method according to claim 8, characterized in that: When establishing an energy consumption prediction model and using the energy consumption prediction model to predict the energy consumption data of the inspection robot, the following are specifically included: S51: Data preparation and feature engineering, integrating historical energy consumption data with terrain features, processing continuous and categorical variables through standardization and one-hot encoding, and constructing a sliding window to generate a time series-static feature fusion dataset; S52: LSTM network architecture design. This involves designing a network structure consisting of an input layer, multiple layers of LSTM units, and a fully connected layer. The input layer integrates time-series energy consumption data with static terrain features, uses the LSTM gating mechanism to capture long-term dependencies, and ultimately outputs the predicted energy consumption per unit distance through the fully connected layer. S53: Model training and tuning, using mean squared error as the loss function, combined with the Adam optimizer to adaptively adjust the learning rate, dividing the model into training sets, validation sets, and test sets for training, and dynamically stopping the overfitting trend through early stopping; S54: Prediction and performance verification: Generate energy consumption prediction values ​​under different terrain parameters based on the test set, calculate indicators to quantify model accuracy, and visually verify the model's fitting effect on complex terrain energy consumption patterns by comparing the predicted curve with the actual value; S55: Dynamic environmental adaptability, introducing an incremental learning mechanism to fine-tune the model weights online when new terrain data is added or energy consumption patterns change.

10. The AI-based intelligent security remote inspection and control method according to claim 9 is characterized in that: When edge computing nodes are used to process and intelligently analyze data collected by inspection robots and data acquisition devices, the following are specifically involved: S61: Data reception and preprocessing: receiving data from the inspection robot through a standardized communication protocol and preprocessing the data; S62: Marking of key data: reading and identifying the results of the inspection robot's preliminary analysis, and marking abnormal data; S63: Intelligent analysis, using AI models to analyze and identify abnormal data in real time; S64: Real-time response and decision-making, judging the abnormality level based on the analysis results of the AI ​​model, and triggering a graded response based on the abnormality level, where the abnormality level includes low-risk events, medium-risk events, and high-risk events.

Images