Real-time safety protection method based on sensor array

Abstract

The invention belongs to the technical field of safety protection, and discloses a real-time safety protection method based on a sensor array, and the method is based on a multi-mode sensor array system combining digital twin modeling, scene adaptation design and precise deployment, and is matched with a scene adaptive mixed domain attention reinforcement learning fusion model. High-dimensional features of sensing data are extracted from four dimensions of space, time, frequency and semantics, data weights of all sensors are dynamically distributed according to scene characteristics, and environmental interference is inhibited while threat associated features are enhanced; a scene-divided threat sample data set is constructed by combining a virtual threat sample generated by a digital twinning model with a real sample, and a lightweight model is optimized through transfer learning, model pruning and incremental training, so that the threat identification accuracy in a complex scene is improved, the adaptability of a system to novel threats is enhanced, and the threat identification efficiency is improved. The identification deviation caused by scene difference is reduced, and the accurate monitoring requirements of different protection scenes are met.

Description

Technical Field

[0001] This invention belongs to the field of security protection technology, specifically a real-time security protection method based on sensor arrays. Background Technology

[0002] Sensor-based security technologies have been widely applied in perimeter security, industrial monitoring, and public safety, enabling threat monitoring through the collection and analysis of multi-source sensor signals and providing crucial protection for personnel and property. Currently, while these technologies are widely used, the following technical challenges remain in adapting to complex scenarios and proactive protection responses:

[0003] Most existing technologies adopt a fixed-weight fusion strategy, which does not fully consider the differences in threat characteristics under different protection scenarios such as indoor enclosed spaces, outdoor open perimeters, and high-risk industrial areas. This makes it difficult for the fused data to accurately match the needs of the scenario, and ultimately reduces the accuracy of threat identification.

[0004] Existing technologies often only initiate protective actions after a threat occurs and triggers an alarm threshold, lacking the ability to predict the development trend of threats. Furthermore, the protection process is mostly based on a preset fixed pattern, which cannot be dynamically adjusted according to the evolution of threats, resulting in untimely protection, low handling efficiency, and easy secondary losses.

[0005] In addition, the sensor deployment of existing protection systems mostly adopts a fixed mode and fails to be adaptively optimized according to the protection priorities of different scenarios, resulting in many monitoring blind spots. Furthermore, in terms of threat location and tracing, most existing technologies can only achieve rough area alerts, and cannot accurately determine the location of the threat or trace the path of threat spread. Summary of the Invention

[0006] The purpose of this invention is to provide a real-time security protection method based on a sensor array to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a real-time security protection method based on a sensor array, comprising a deployment stage, a data acquisition and preprocessing stage, a data fusion stage, a threat identification stage, an active protection and prediction stage, a location and tracing stage, and an optimization and iteration stage;

[0008] Preferably, the deployment phase establishes a three-in-one multimodal sensor array deployment system integrating digital twin modeling, scene adaptation design, and precise deployment implementation. The multimodal sensors include vibration sensors, acoustic sensors, infrared thermal imaging sensors, microwave radar sensors, BeiDou and UWB dual-mode positioning sensors, and environmental perception sensors. First, a scene digital twin model is established, integrating scene topology parameters, threat type distribution, environmental interference characteristics, and historical protection data to achieve real-time correspondence between the physical scene and the virtual model. Based on the digital twin model, a scene feature database is built, including key parameter dimensions such as electromagnetic interference level and signal obstruction area.

[0009] Secondly, an improved hybrid genetic and particle swarm optimization algorithm is used to partition the protected area into risk zones and optimize the sensor deployment locations. The blind-spot-free coverage effect of the deployment scheme is verified through virtual roaming, while the sensor type and deployment density are matched with the threat distribution of the scenario.

[0010] The specific methods and blind spot determination criteria for virtual roaming verification are as follows:

[0011] Roaming path planning: Based on the scene topology parameters of the digital twin model, uniformly distributed roaming path points are generated with a path point spacing of ≤0.5 meters, covering the entire protected area. Each path point simulates the sensor signal receiving scenario.

[0012] Coverage strength threshold: The sensor signal received strength at each path point is ≥-85dBm. The electromagnetic interference level collected by the environmental perception sensor is dynamically corrected. For each level increase in the interference level, the threshold is lowered by 5dBm.

[0013] Blind zone determination rule: If the signal reception strength of 3 or more consecutive path points is less than the threshold, or the area of ​​a single blind zone is greater than 0.5m², then the deployment plan is determined to have blind zones, and the sensor deployment points need to be re-optimized;

[0014] Verification iteration: Repeat the virtual roaming verification after each optimization until the blind area ratio is ≤0.1% and there are no continuous blind path points.

[0015] The specific fusion logic and parameters of the improved genetic and particle swarm optimization (IGA-PSO) algorithm are as follows:

[0016] Algorithm objective: Minimize the area of ​​the monitoring blind zone. Sensor deployment costs The weighted sum, the fitness function is:

[0017]

[0018] : The set of coordinates of sensor deployment points, where n is the number of sensors;

[0019] Total area of ​​the protected zone, in m²;

[0020] Monitoring blind spots are given higher priority than cost.

[0021] Deployment cost weighting;

[0022] : Preset maximum deployment cost threshold, in yuan.

[0023] Fusion steps:

[0024] Initialization: Genetic algorithm population size = 50, particle swarm size = 50, total number of iterations = 100;

[0025] Particle Swarm Update: Velocity Update Formula:

[0026]

[0027] Position update formula:

[0028] =0.8 is the inertia weight. =1.5 is the learning factor; , is a random number; gbest represents the optimal solution for each individual particle; gbest represents the global optimal solution.

[0029] Genetic algorithm optimization: Every 20 iterations, the optimal solution of the particle swarm is substituted into the genetic algorithm. The selection operator adopts the roulette wheel method, the crossover operator is single-point crossover (crossover probability = 0.7), and the mutation operator is Gaussian mutation (mutation probability = 0.05).

[0030] Termination condition: The change in fitness function value over 10 consecutive iterations. Output the set of optimal deployment points.

[0031] To address the environmental interference characteristics of different scenarios, a multi-sensor cross-calibration mechanism is adopted. By leveraging the complementarity of data from each sensor and comparing the preset scenario-specific standard reference signal with the actual acquired signal, the sensor acquisition parameters are dynamically corrected. At the same time, environmental perception data is combined to compensate for environmental errors in real time.

[0032] The mathematical model and steps of the multi-sensor cross-calibration mechanism are as follows:

[0033] Calibration objective: Correct the raw sensor data. Environmental errors Output calibrated data ;

[0034] Error compensation formula:

[0035]

[0036] in:

[0037] Sensor numbers, i, j = 1, 2, ..., m, where m is the total number of sensors;

[0038] The scene-specific standard reference signal for sensor j is retrieved from the scene feature database;

[0039] The cross-calibration coefficient between sensor i and sensor j is obtained by fitting historical data, and its value ranges from [0.05, 0.3].

[0040] The comprehensive environmental interference value collected by the environmental perception sensor is normalized to [0, 1] and includes a weighted sum of temperature, electromagnetic interference, etc.

[0041] : Environmental error correction coefficient for sensor i, preset according to sensor type, such as acoustic sensor =0.2, infrared sensor =0.15;

[0042] Calibration steps:

[0043] Step 1: Retrieve standard reference signals from each sensor from the scene feature database.

[0044] Step 2: Collect raw data from each sensor Environmental interference value

[0045] Step 3: Substitute the data into the formula to calculate the calibrated data. Repeat calibration every 5 minutes and update dynamically. .

[0046] Preferably, the acquisition and preprocessing stage is based on the PTP time protocol to realize synchronous data acquisition of all sensors in the array, and adopts a three-level computing architecture of edge layer, fog node layer and cloud node layer.

[0047] The raw data collected is pre-processed locally at the edge layer, including removing environmental noise and sensor noise based on wavelet threshold denoising algorithm; data standardization and dimensionality reduction are completed at the fog node layer, and the national cryptographic SM4 encryption algorithm is introduced to encrypt the data; the cloud node layer is used for distributed storage and backup of data, and blockchain technology is used to store the pre-processed key data.

[0048] After preprocessing, the data is encapsulated into encrypted data packets of a unified format and transmitted to the local edge gateway via a 5G private network or industrial Ethernet.

[0049] Preferably, in the data fusion stage, a scene-adaptive hybrid domain attention reinforcement learning fusion model is established. The fusion model includes three-level modules: scene recognition, feature matching, and weight optimization. First, the current protection scene type is determined by combining the scene recognition algorithm with the real-time scene status of the digital twin model in the deployment stage.

[0050] The specific hierarchical structure of the hybrid domain attention module is as follows:

[0051] The input layer receives standardized multi-sensor data with dimensions of [number of sensors × time step × single sensor feature dimension]. The data passes through spatial attention sublayer, temporal attention sublayer, frequency attention sublayer, and semantic attention sublayer in sequence. Finally, the four-dimensional features are fused through the splicing layer, and the output is a high-dimensional feature vector with dimensions of [time step × 288].

[0052] The spatial attention sublayer uses a 3×3 convolutional kernel to extract spatial correlation features, and the output dimension is [number of sensors × time step × 64];

[0053] The temporal attention sublayer uses a bidirectional GRU network with 128 hidden layer neurons and an output dimension of [number of sensors × time step × 128].

[0054] The frequency attention sublayer extracts spectral features based on STFT transform and uses 1×1 convolution to reduce the dimension to 64.

[0055] The semantic attention sublayer combines threat semantic tags from the scene feature database and maps them to 32-dimensional semantic features through a fully connected layer.

[0056] The specific configuration of the reinforcement learning weight allocation module is as follows:

[0057] ① Intelligent agent: The action is based on the weight coefficients of each sensor data. The weight coefficients range from [0, 1] and the sum is 1.

[0058] ②State space: includes the current scene type, the signal-to-noise ratio of each sensor data, and the threat feature matching degree;

[0059] ③ Reward function: R = α × data fusion accuracy - β × weight allocation complexity, where α = 0.7, β = 0.3, fusion accuracy is calculated by the cosine similarity with the scene standard feature vector, and complexity is the variance of the weight coefficients;

[0060] ④ Training parameters: The optimizer is Adam, the learning rate is 0.001, the number of iterations is 500 rounds, the batch size is 32, and the experience replay buffer size is 10000.

[0061] Secondly, the hybrid domain attention module extracts high-dimensional features from the data of each sensor from four dimensions: spatial domain, temporal domain, frequency domain, and semantic domain. Finally, the reinforcement learning weight allocation module dynamically adjusts the weight ratio of each sensor data according to the scene-specific threat features.

[0062] The attention mechanism strengthens features that are strongly related to the current scene's threats and suppresses environmental interference features, ultimately fusing them into a global feature vector that adapts to the current scene. Meanwhile, the reinforcement learning module autonomously optimizes the weight allocation strategy through continuous interaction with the environment.

[0063] The input-output mapping relationship between the hybrid domain attention module and the scene:

[0064] ① Spatial domain input: Deployment coordinates (x, y, z) and signal strength of each sensor; Output: Spatial correlation weights between sensors;

[0065] ②Time domain input: Sensor data sequence of 100 consecutive time steps, output time series trend characteristics;

[0066] ③ Frequency domain input: STFT spectrum of the signal, outputting the weighting for distinguishing between high-frequency threat features and low-frequency interference features;

[0067] ④ Semantic domain input: scene threat labels, such as industrial leakage and illegal intrusion, output semantic matching score; after weighted fusion of four-dimensional features, the output global feature vector has a fit degree of ≥0.85 with the scene.

[0068] Preferably, the threat identification stage is based on the scenario-adaptive global feature vector from the data fusion stage, sets up a scenario-specific lightweight deep learning model, and combines the virtual sample generation capability of the digital twin model from the deployment stage to generate virtual threat samples based on the scenario feature database and the digital twin model, and then merges them with real samples to construct a scenario-specific threat sample dataset.

[0069] The specific hierarchical structure of the lightweight deep learning model is as follows:

[0070] ① Input layer: Receives global feature vectors, dimension = [time step × 288];

[0071] ② Feature extraction layer: Based on the improvement of MobileNetV3-Small, the last 3 depthwise separable convolutions of the original network are replaced with attention-enhanced convolutions (adding CBAM attention module), and the number of feature map channels is 32→64→128→256 respectively;

[0072] ③ Classification layer: It contains two fully connected layers (with 128 and 64 neurons respectively) and one Softmax output layer. The output dimension is equal to the number of threat categories in the scene, such as illegal intrusion, device leakage, open flame, etc., which are configured according to the specific scene.

[0073] The specific steps of model training are as follows:

[0074] ① Pre-training: The MobileNetV3-Small base model was pre-trained using the ImageNet dataset, with the first 80% of the network layers frozen;

[0075] ② Fine-tuning: Based on the threat sample dataset for different scenarios, the ratio of virtual samples to real samples is 3:1, the total number of samples is ≥5000, all layers are unfrozen for fine-tuning, the learning rate is 0.0005, the number of iterations is 300 rounds, the batch size is 64, and the loss function is cross-entropy loss;

[0076] ③ Model pruning: Structured pruning is adopted, with a pruning ratio of 30%, retaining convolutional kernels with a threat feature contribution of ≥0.6;

[0077] ④ Quantization: INT8 quantization is used, and the quantization calibration set is 10% of the scene sample dataset to ensure that the model inference accuracy decreases by ≤3% after quantization.

[0078] Incremental training parameters: Incremental training is triggered when the number of new samples is ≥500, learning rate = 0.0001, number of iterations = 100 rounds, knowledge distillation is used to preserve the performance of historical models, distillation temperature = 2.0.

[0079] The lightweight model is trained and optimized for different scenarios using transfer learning, enabling the model to identify specific threat types in the current scenario. Model pruning and quantization techniques are introduced to reduce the computational complexity of the model.

[0080] Meanwhile, a scenario-adaptive update mechanism is added to periodically collect new threat samples in the current scenario and iterative virtual samples generated by digital twins to incrementally train the model and improve the model's ability to identify new threats in the current scenario.

[0081] The triggering conditions and execution rules of the scene adaptive update mechanism are as follows:

[0082] Triggering conditions:

[0083] ① The number of newly added real threat samples is ≥500;

[0084] ② The model's recognition accuracy was <92% in three consecutive tests;

[0085] ③ The scene feature database is updated by ≥20%, for example, when new threat types are added or environmental interference features change; incremental training is triggered when any of these conditions are met.

[0086] Iterative virtual sample generation: During each incremental training, the digital twin model generates twice the number of iterative virtual samples based on the features of the newly added real samples, ensuring that the sample distribution is consistent with the newly added real samples;

[0087] Training stops when the model's recognition accuracy is ≥95% after incremental training and the validation set loss function value is ≤0.08. In this case, training stops and the model parameters are updated.

[0088] Preferably, the proactive protection prediction stage combines the threat types identified in the threat identification stage with the temporal characteristics in the data fusion stage to establish a proactive protection system that includes threat prediction, virtual simulation, level assessment, and dynamic adaptation of protection strategies; an improved LSTM neural network is used in conjunction with a digital twin model in the deployment stage to predict the future development trend of threats, and threat evolution simulation is performed in the digital twin model to simulate the implementation effect of different protection strategies;

[0089] The specific structure of the improved LSTM neural network is as follows:

[0090] ① Input layer: Receives threat type encoding vector and time-series feature vector, including 5 dimensions such as signal strength and spread rate. Dimension = [time step × 5], and after concatenation, the input dimension = [time step × (number of threat categories + 5)].

[0091] ② Hidden layers: Set up 3 LSTM units, with 256 neurons per layer, and add dropout layers and layer normalization;

[0092] ③ Output layer: The fully connected layer outputs the evolution parameters of the threat over the next 10 time steps, such as the spread range and intensity level. The output dimension is [10×3].

[0093] Model training parameters: The training dataset contains ≥3000 historical threat evolution data, learning rate = 0.001, number of iterations = 400 rounds, batch size = 48, loss function is MSE loss, and optimizer is RMSprop.

[0094] Reinforcement learning dynamically adjusts the specific configuration of the protective action sequence:

[0095] ①State space: includes threat level, evolution rate, and scenario constraints;

[0096] ② Action space: includes 8 types of protective actions, such as enhanced monitoring and area isolation, as well as the intensity of action execution;

[0097] ③ Reward function: R = γ × Threat control efficiency - δ × Protection cost, where γ = 0.8, δ = 0.2, control efficiency = (threat evolution mitigation magnitude / preset threshold), protection cost = normalized value of resource consumption for action execution;

[0098] ④ Training parameters: learning rate = 0.0008, number of iterations = 600 rounds, exploration rate ε decreases linearly from 0.9 to 0.1.

[0099] By combining threat types, prediction results, simulation effects and scenario characteristics, a scenario-adaptive threat level assessment system is established, which divides threats into three levels: potential threats, ongoing threats and urgent threats. A scenario-specific dynamic protection strategy library is established for different levels of threats, and a reinforcement learning algorithm is introduced to dynamically adjust the sequence of protection actions according to the real-time evolution of the threat.

[0100] When a potential threat is detected, a key monitoring enhancement mode is activated, combining BeiDou and UWB positioning deployed during the deployment phase to achieve precise monitoring of key areas; when a threat is occurring, area isolation, audible and visual alarms, and precise positioning and tracking are activated; when an emergency threat is detected, emergency shutdown, joint rescue, evacuation guidance, and remote visual command are activated, realizing the transformation from passive response to proactive prediction and response.

[0101] The logic of combining improved LSTM with digital twin models:

[0102] ① Input data association: The digital twin model outputs scene topology parameters (such as wall barrier coefficient and channel width) and environmental interference parameters (such as wind speed and temperature), which are used as additional input features of the improved LSTM. These are then concatenated with threat type and time series features and input into the model.

[0103] ② Output data feedback: The threat evolution prediction results output by LSTM are fed into the digital twin model to drive the threat evolution simulation in the virtual scene. The simulation results correct the prediction parameters of LSTM in reverse to ensure that the prediction error is ≤10%.

[0104] Preferably, the positioning and tracing stage establishes a BeiDou, UWB and inertial navigation fusion positioning model based on the spatial distribution information of the sensor array in the deployment stage, the topological features of the digital twin scene, and the multimodal time-series fusion data in the data fusion stage; by comparing the time difference and signal strength attenuation law of threat signals collected by different sensor nodes, the positioning algorithm is corrected by combining the scene topology parameters in the digital twin model.

[0105] The fusion localization model uses the Extended Kalman Filter (EKF) algorithm to achieve data fusion, and the specific model is as follows:

[0106] State equations (position and velocity of the target):

[0107]

[0108] in:

[0109] The state vector at time k, where x, y, and z are three-dimensional coordinates. , , For three-dimensional velocity;

[0110] State transition matrix, assumption of uniform motion;

[0111] Input matrix, T=0.1s is the sampling period;

[0112] : Three-dimensional acceleration acquired by inertial navigation;

[0113] Process noise (Gaussian distribution, variance)

[0114] ;

[0115] Observation equations (positioning output from BeiDou and UWB):

[0116]

[0117] in:

[0118] : Observation vector at time k ( BeiDou positioning coordinates UWB positioning coordinates);

[0119] : Observation matrix, extracting only location information;

[0120] Observation noise, BeiDou noise variance =4×10 -4 UWB noise variance =1×10 -5 ;

[0121] Digital twin topology correction:

[0122] If the localization result falls within the signal obstruction area of ​​the digital twin model (such as walls or obstacles), a correction coefficient is introduced.

[0123]

[0124] in The normal vector for the occluded region is extracted from the topology parameters of the digital twin scene. =0.3m is the corrected distance.

[0125] In terms of threat attribution, the time-series data stored on the blockchain during the pre-processing stage is combined with the current location data to trace the threat spread path and generate an attribution report that includes timestamps, location coordinates, and the threat evolution process. At the same time, the attribution data is fed back to the digital twin model in the deployment stage to optimize the formulation of protection strategies.

[0126] Preferably, the optimization iteration stage establishes a closed-loop optimization system for the entire chain based on the technical links formed by the aforementioned steps; during the execution of the protection strategy in the active protection prediction stage, scene adaptation effect data and active protection effect data are collected synchronously through the sensor array and digital twin model in the deployment stage. The scene adaptation effect data includes data fusion accuracy, threat identification accuracy, and positioning accuracy, while the active protection effect data includes threat handling time, whether secondary losses are avoided, and prediction accuracy.

[0127] Based on the evaluation results, the core modules of the system are optimized in a targeted manner. If the scene adaptation effect is poor, the sensor deployment parameters in the deployment phase, the hybrid domain attention weight rules in the data fusion phase, and the scene parameters of the digital twin model are adjusted. If the active protection effect is not good, the threat prediction model parameters in the active protection prediction phase and the scene-specific protection strategy library are updated.

[0128] Simultaneously, the threat data and protection effectiveness data after handling will be added to the scenario feature database during the deployment phase for iterative optimization of the digital twin model and calibration updates of the sensor array.

[0129] The beneficial effects of this invention are as follows:

[0130] 1. This invention is based on a multimodal sensor array system that combines digital twin modeling, scene-adaptive design, and precise deployment. It is paired with a scene-adaptive hybrid domain attention reinforcement learning fusion model to extract high-dimensional features from sensor data from four dimensions: space, time, frequency, and semantics. The weights of each sensor data are dynamically allocated according to scene characteristics, which strengthens threat correlation features while suppressing environmental interference. By combining virtual threat samples generated by the digital twin model with real samples, a scene-specific threat sample dataset is constructed. The lightweight model is optimized through transfer learning, model pruning, and incremental training, which improves the accuracy of threat identification in complex scenes. At the same time, it enhances the system's adaptability to new threats, reduces identification bias caused by scene differences, and meets the precise monitoring needs of different protection scenarios.

[0131] 2. This invention establishes a proactive protection system encompassing threat prediction, virtual simulation, level assessment, and dynamic adaptation strategies. It employs an improved LSTM neural network combined with a digital twin model to accurately predict threat development trends and simulate the implementation effects of different protection schemes. Based on threat type, prediction results, and scenario characteristics, a three-tiered threat level assessment system is established. A dedicated dynamic protection strategy library is developed for potential, ongoing, and emergency threats. Reinforcement learning algorithms are used to adjust the sequence of protection actions according to the real-time evolution of the threat. For potential threats, monitoring of key areas is strengthened; for ongoing threats, area isolation and location tracking are initiated; and for emergency threats, rescue and evacuation guidance are coordinated. This upgrades the protection mode from passive response to proactive prediction and response, shortening threat handling time and effectively reducing the risk of secondary losses.

[0132] 3. This invention employs an improved hybrid genetic and particle swarm optimization algorithm to perform risk zoning and sensor deployment point optimization in the protected area. Virtual roaming verification ensures blind-spot-free coverage, while a multi-sensor cross-calibration mechanism utilizes the complementarity of sensor data to compensate for environmental errors. A fusion positioning model integrating BeiDou, UWB, and inertial navigation is constructed, and a positioning algorithm is corrected based on the topological features of the digital twin scenario to meet the positioning needs of different scenarios. Based on blockchain-stored time-series and positioning data, a complete source tracing report containing timestamps, positioning coordinates, and the threat evolution process is generated. This solves the problems of monitoring blind spots and inaccurate positioning, and provides a reliable basis for threat tracing. The source tracing data then feeds back into the digital twin model. Attached Figure Description

[0133] Figure 1 This is an overall flowchart of the method of the present invention;

[0134] Figure 2 This is a flowchart of the data processing and threat identification process of the present invention;

[0135] Figure 3 This is a flowchart illustrating the active protection and closed-loop optimization process of this invention. Detailed Implementation

[0136] 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.

[0137] like Figures 1 to 3As shown, this embodiment of the invention provides a real-time security protection method based on a sensor array, including a deployment stage, a data acquisition and preprocessing stage, a data fusion stage, a threat identification stage, an active protection and prediction stage, a location and tracing stage, and an optimization and iteration stage. The specific implementation of each stage is as follows:

[0138] Based on the core characteristics of the protection scenario (such as indoor, outdoor, open, enclosed, densely populated, and concentrated equipment) and the protection level requirements, a multimodal sensor array deployment system integrating digital twin modeling, scenario adaptation design, and precise deployment implementation is established during the deployment phase. The multimodal sensors include vibration sensors, acoustic sensors, infrared thermal imaging sensors, microwave radar sensors, BeiDou and UWB dual-mode positioning sensors, and environmental perception sensors. The environmental perception sensors are used to collect environmental parameters such as temperature, humidity, light intensity, and electromagnetic interference intensity.

[0139] First, a scene digital twin model is established, integrating scene topology parameters, threat type distribution, environmental interference characteristics, and historical protection data to achieve real-time correspondence between the physical scene and the virtual model. Based on the digital twin model, a scene feature database is built, including key parameter dimensions such as electromagnetic interference level and signal obstruction area.

[0140] The real-time synchronization mechanism of the digital twin model is as follows:

[0141] ① Synchronization frequency: dynamically adjusted based on scene type. Indoor enclosed scenes are synchronized once every 0.5 seconds, outdoor open scenes are synchronized once every 1 second, and industrial high-risk scenes are synchronized once every 0.2 seconds.

[0142] ② Synchronized data types: including real-time sensor data, environmental interference parameters, threat evolution status data, and protective action execution feedback data;

[0143] ③ Update logic: The system adopts a combination of incremental update and full calibration. Incremental update only synchronizes changed data, such as threat signal strength and location coordinates. Full calibration is performed every 10 minutes to ensure that the deviation between the virtual model and the physical scene is ≤0.1 meters and ≤0.3dB.

[0144] ④ Data interaction protocol: The MQTT protocol is used to realize bidirectional data transmission between physical sensors and digital twin models, with a transmission delay of ≤50ms.

[0145] Secondly, an improved hybrid genetic and particle swarm optimization algorithm is adopted, combined with the virtual simulation capability of the digital twin model, to perform risk zoning of the protected area and optimize the sensor deployment points. The blind-spot-free coverage effect of the deployment scheme is verified through virtual roaming. At the same time, the sensor type and deployment density are matched with the threat distribution of the scenario. For example, the density of gas sensors and infrared thermal imaging sensors is strengthened in high-risk industrial areas, and the deployment density of Beidou / UWB dual-mode positioning sensors and microwave radar sensors is increased in outdoor perimeter areas.

[0146] To address the environmental interference characteristics of different scenarios, a multi-sensor cross-calibration mechanism is adopted. This mechanism leverages the complementarity of data from various sensors, such as cross-verification of BeiDou and UWB positioning data and collaborative calibration of acoustic and vibration signals. By comparing preset scenario-specific standard reference signals with actual acquired signals, sensor acquisition parameters, such as gain, threshold, and sampling frequency, are dynamically corrected. Simultaneously, environmental perception data is combined to compensate for errors caused by temperature, electromagnetic interference, etc., in real time, thereby improving the data acquisition accuracy of the sensor array in complex scenarios.

[0147] Based on the deployment and cross-calibration of the multimodal sensor array during the deployment phase, the acquisition and preprocessing phase uses the PTP precision time protocol to achieve millisecond-level synchronous data acquisition of all sensors in the array, avoiding data misalignment caused by acquisition time difference. At the same time, a three-level computing architecture of edge layer, fog node layer and cloud node layer is adopted.

[0148] The raw data is pre-processed locally at the edge layer (sensor nodes), including removing environmental noise and sensor noise based on wavelet threshold denoising algorithm. Deep preprocessing is performed at the fog node layer to standardize the data, mapping sensor data of different dimensions to the same numerical range. Dimensionality reduction is also performed, using an improved PCA algorithm to retain key data features and remove redundant data. The national standard SM4 encryption algorithm is introduced to encrypt the data. The cloud node layer is used for distributed data storage and backup, and blockchain technology is used to store and verify the pre-processed key data, ensuring the immutability and traceability of the data.

[0149] The specific parameters of the wavelet thresholding denoising algorithm are as follows:

[0150] 1. Wavelet basis function: Select db4 wavelet, decomposition level = 3 levels;

[0151] 2. Threshold type: An improved soft threshold function is used.

[0152] ,

[0153] Where the threshold , For data length, median This is an estimate of the noise standard deviation;

[0154] 3. Processing steps:

[0155] ① Perform three-level wavelet decomposition on the original data to obtain low-frequency approximation coefficients and high-frequency detail coefficients;

[0156] ② The improved soft thresholding function described above is used to process the high-frequency detail coefficients;

[0157] ③ Reconstruct the wavelet coefficients and output the denoised data.

[0158] The specific improvements and parameters of the improved PCA algorithm are as follows:

[0159] Improved logic: Introduce kernel principal component analysis (KPCA) mechanism to solve the problem of dimensionality reduction failure of linear PCA for nonlinear sensor data;

[0160] Kernel function selection: Use a polynomial kernel function. ,in This was determined through cross-validation;

[0161] Dimensionality reduction threshold: cumulative principal component contribution rate ≥ 95%, number of retained principal components ≤ 30% of the original feature dimension;

[0162] Execution steps:

[0163] ① Centralize the processing of standardized sensor data;

[0164] ② Calculate the kernel matrix and perform eigenvalue decomposition;

[0165] ③ Select principal components with a cumulative contribution rate of ≥95% and output the dimensionality-reduced feature vectors.

[0166] The data association and hash calculation logic for blockchain-based evidence storage is as follows:

[0167] Key data structure: The evidence storage data includes a quadruple of sensor ID, timestamp, preprocessed data, and scene identifier. ;

[0168] Hash calculation formula (using SHA-256 algorithm):

[0169]

[0170] : The block hash value of the k-th data item;

[0171] : The block hash value of the previous data (genesis block) );

[0172] XOR operation;

[0173] : The k-th quadruple data;

[0174] Fixed salt value to prevent tampering;

[0175] Time-series data association rules: based on timestamps An index is established with the sensor ID_s, and each piece of data stored in the blockchain contains the hash of the previous block, ensuring that the time sequence is traceable. If the hash value is discontinuous after tampering, it can be identified.

[0176] After preprocessing, the data is encapsulated into encrypted data packets of a unified format and transmitted to the local edge gateway via a 5G private network or industrial Ethernet.

[0177] For the standardized encrypted data packets output in the data acquisition and preprocessing stage, a scenario-adaptive hybrid domain attention reinforcement learning fusion model is established in the data fusion stage. The fusion model includes three-level modules: scenario recognition, feature matching, and weight optimization. First, the current protection scenario type is accurately determined by combining the scenario recognition algorithm with the real-time scenario status of the digital twin model in the deployment stage.

[0178] The scene recognition algorithm uses an improved SVM (Support Vector Machine) algorithm, with the following specific configuration:

[0179] Input features: Scene topology parameters and environmental disturbance features output by the digital twin model, totaling 8-dimensional feature vectors. All are normalized to [0, 1];

[0180] 2. Kernel Function: Radial Basis Function (RBF) is used.

[0181]

[0182] Feature vectors of two scenes;

[0183] =0.8: Kernel function parameters, determined through cross-validation;

[0184] Euclidean distance between eigenvectors;

[0185] Training parameters: penalty coefficient C=10, training samples are historical scene data in the scene feature database, classification output is 3 core scenes, including indoor enclosed, outdoor open, and industrial high-risk, training stops when the recognition accuracy is ≥95%.

[0186] Secondly, the hybrid domain attention module extracts high-dimensional features from each sensor's data across four dimensions: spatial domain (sensor deployment location association), temporal domain (signal timing changes), frequency domain (signal spectrum features), and semantic domain (threat type association). Finally, the reinforcement learning weight allocation module dynamically adjusts the weight ratio of each sensor's data based on scene-specific threat features. For example, in outdoor perimeter scenarios, the weight of microwave radar and BeiDou / UWB positioning data is increased, while infrared thermal imaging data is used as an auxiliary; in indoor enclosed scenarios, the weight of acoustic and infrared thermal imaging data is strengthened.

[0187] By strengthening features that are strongly correlated with the current scene's threats and suppressing environmental interference features through an attention mechanism, the data is ultimately fused into a global feature vector that adapts to the current scene. At the same time, the reinforcement learning module autonomously optimizes the weight allocation strategy through continuous interaction with the environment, thereby improving the scene-specificity and accuracy of data fusion.

[0188] The threat identification phase is based on the scenario-adaptive global feature vector from the data fusion phase. It sets up a scenario-specific lightweight deep learning model (based on an improvement of MobileNetV3) and combines the virtual sample generation capability of the digital twin model from the deployment phase. Based on the scenario feature database and digital twin model of this phase, a large number of virtual threat samples are generated, such as illegal intrusion and device leakage scenarios in different environments. These are then fused with real samples to construct a scenario-specific threat sample dataset.

[0189] The rules for generating virtual threat samples and the feature matching requirements are as follows:

[0190] Generation logic: Based on real threat samples in the scene feature database, a digital twin model is used to simulate different environmental interferences (such as electromagnetic interference intensity ±30%, temperature fluctuation ±5℃), threat intensity gradients (such as signal strength 0.2~1.0 times that of real samples), and scene topology changes (such as obstacle position offset 0~1 meter) to generate diverse virtual samples.

[0191] Feature matching criteria: The cosine similarity of the core features of the virtual sample and the real sample is ≥0.9; the core features include spectral peaks, temporal variation trends, spatial distribution patterns, etc.

[0192] Sample quantity ratio: 3-5 virtual samples are generated for each real sample. The total sample size of the threat sample dataset for each scenario must meet the following requirements: ≥8000 for indoor enclosed scenarios, ≥10000 for outdoor open scenarios, and ≥12000 for industrial high-risk scenarios.

[0193] Sample labeling rules: Virtual samples are automatically labeled with "Scene Type - Threat Type - Environmental Interference Level - Threat Intensity", which is consistent with the labeling format of real samples to ensure the consistency of training data.

[0194] Transfer learning is used to train and optimize the lightweight model in different scenarios. For example, indoor scenarios focus on illegal intrusion and left-behind items, while high-risk industrial areas focus on equipment leaks and open flame hazards, enabling the model to identify specific threat types in the current scenario. At the same time, model pruning and quantization techniques are introduced to reduce the computational complexity of the model and ensure millisecond-level inference at the fog node layer to guarantee real-time performance.

[0195] Meanwhile, a scenario-adaptive update mechanism is added to periodically collect new threat samples (real samples) in the current scenario and iterative virtual samples generated by digital twins to incrementally train the model and continuously improve the model's ability to identify new threats in the current scenario.

[0196] The proactive protection prediction phase combines the threat types identified in the threat identification phase with the temporal characteristics of the data fused in the data fusion phase, such as the rate of change of threat signal strength and the spread speed, to establish a proactive protection system that includes threat prediction, virtual simulation, level assessment, and dynamic adaptation of protection strategies. An improved LSTM neural network is used in conjunction with a digital twin model in the deployment phase to predict the future development trend of threats, such as the spread range of leaked gas and the spread speed of fire. Threat evolution simulation is performed in the digital twin model to simulate the implementation effect of different protection strategies.

[0197] By combining threat types, prediction results, simulation effects, and scenario characteristics, a scenario-adaptive threat level assessment system is established, classifying threats into three levels: potential threats, emerging threats, and urgent threats. A scenario-specific dynamic protection strategy library is established for different threat levels, and reinforcement learning algorithms are introduced to dynamically adjust the sequence of protection actions based on the real-time evolution of the threat (such as whether potential threats have escalated or whether emerging threats have spread).

[0198] The specific execution details of the scenario-specific dynamic protection strategy library are as follows:

[0199] Such as enclosed indoor environments like computer rooms and archives:

[0200] ① Potential threat: The frame rate of the infrared thermal imaging sensor is increased to 30fps, and the sampling interval for BeiDou / UWB positioning is shortened to 0.5 seconds;

[0201] ② If an illegal intrusion or other threat is incurred: the area is isolated and the access control is locked. The protection radius is equal to the location coordinates ±3 meters. The frequency of the audible and visual alarm is 2Hz, the sound level is 85dB, and the light flashing frequency is 2 times / second.

[0202] ③ In case of emergency threats such as fire: the air conditioning system is activated to cut off ventilation, the dry powder fire extinguisher is activated to cover a range of ±5 meters from the location coordinates, the evacuation guidance indicator lights are turned on according to the preset escape route, and the remote visual command system is automatically connected to the security center.

[0203] Such as open outdoor scenes like the perimeter of a park:

[0204] ① Potential threat: The microwave radar sensor's detection range is increased by 20%, and it can continuously track the target's movement trajectory;

[0205] ② If a threat is posed, such as climbing over the fence: the area isolation triggers the high-voltage start of the electronic fence, with a voltage of 10kV and a safe current of ≤10mA. The audible and visual alarm covers a range of 50 meters around the perimeter.

[0206] ③ In case of emergency threats such as suspicious persons carrying dangerous items: the park patrol robot will be coordinated to locate the coordinates, with a response time of ≤3 minutes, the access control of surrounding entrances and exits will be closed, and an alarm will be simultaneously reported to the local public security system.

[0207] High-risk industrial settings such as chemical workshops:

[0208] ① Potential threat: The gas sensor sampling interval is shortened to 1 second, and concentration data is uploaded in real time;

[0209] ② If a gas leak or other threat is in progress: the isolation valve is closed due to area isolation, valves within 50 meters upstream and downstream of the leak point are activated, and audible and visual alarms cover the entire workshop area;

[0210] ③ In case of emergency threats such as explosion risk: emergency shutdown triggers the main power supply of the production line to be cut off, the linkage sprinkler system is activated, covering the positioning coordinates ±10 meters, the evacuation guidance broadcast plays escape instructions in a loop, and the rescue personnel positioning system marks the dangerous area.

[0211] The quantitative thresholds for Level 3 threat level are as follows (based on temporal and scenario features in the fused data):

[0212] Potential threat determination: Threat signal strength ≤ 0.3 × scene threat threshold, the scene threat threshold is retrieved from the scene feature database, and the signal strength change rate ≤ 0.05dB / s, with no diffusion trend. No diffusion trend means that the positioning coordinate deviation between adjacent time steps is ≤ 0.5 meters;

[0213] Threat determination in progress: 0.3 × scene threat threshold < threat signal strength ≤ 0.7 × scene threat threshold, signal strength change rate 0.05~0.2dB / s, diffusion speed ≤ 1 meter / minute, not touching core protection areas, such as core equipment areas or densely populated areas;

[0214] Emergency threat determination: Threat signal strength > 0.7 × scenario threat threshold, or signal strength change rate > 0.2 dB / s, or spread speed > 1 meter / minute, or has reached the core protection area;

[0215] The scenario threat threshold is the average signal strength of the protection actions triggered by historical threats in each scenario, and is updated in real time by the scenario feature database.

[0216] When a potential threat is detected, a key monitoring enhancement mode is activated, combining BeiDou and UWB positioning deployed during the deployment phase to achieve precise monitoring of key areas; when a threat is occurring, area isolation, audible and visual alarms, and precise positioning and tracking are activated; when an emergency threat is detected, emergency shutdown, joint rescue, evacuation guidance, and remote visual command are activated, realizing the transformation from passive response to proactive prediction and response.

[0217] To support the accurate execution of dynamic protection strategies during the proactive protection prediction phase, the positioning and tracing phase establishes a BeiDou, UWB, and inertial navigation fusion positioning model based on the spatial distribution information of the sensor array during the deployment phase, the topological features of the digital twin scene, and the multimodal temporal fusion data during the data fusion phase. By comparing the time difference and signal strength attenuation patterns of threat signals collected by different sensor nodes, and combining the scene topology parameters (such as wall obstruction and terrain undulations) in the digital twin model, the positioning algorithm is corrected.

[0218] In indoor scenarios, UWB positioning is the primary method, supplemented by inertial navigation. In outdoor scenarios, BeiDou high-precision positioning is the primary method, supplemented by UWB, achieving seamless switching between indoor and outdoor positioning. The indoor positioning accuracy is better than 0.1 meters, and the outdoor positioning accuracy is better than 2 centimeters.

[0219] In terms of threat attribution, the time-series data stored on the blockchain during the pre-processing stage is combined with the current location data to trace the threat spread path and generate an attribution report that includes timestamps, location coordinates, and the threat evolution process. At the same time, the attribution data is fed back to the digital twin model in the deployment stage to optimize the formulation of protection strategies.

[0220] The block structure and data association index rules for blockchain evidence storage are as follows:

[0221] Block structure:

[0222] ① Block header: contains the hash value of the previous block, timestamp, and data checksum;

[0223] ② Block body: contains sensor ID, preprocessed data, positioning coordinates, and scene identifier;

[0224] Association Indexing Rules: A primary index is built using sensor ID and timestamp, and a secondary index is built using location coordinates and threat type. When tracing the source, the time-series data of the target sensor is queried through the primary index, and then the data of different sensors and location data of the same threat event are associated through the secondary index.

[0225] Data query efficiency: The response time for a single traceability data query is ≤100ms, and it supports filtering queries by time range and location area.

[0226] The optimization iteration phase establishes a closed-loop optimization system across the entire chain based on the technical links formed in the aforementioned steps. During the execution of the protection strategy in the proactive protection prediction phase, scenario adaptation effect data and proactive protection effect data are collected synchronously through the sensor array and digital twin model in the deployment phase. The scenario adaptation effect data includes the data fusion accuracy in the data fusion phase, the threat identification accuracy in the threat identification phase, and the positioning accuracy in the positioning and tracing phase. The proactive protection effect data includes the threat handling time, whether secondary losses are avoided, and the prediction accuracy in the proactive protection prediction phase.

[0227] The specific calculation methods for scene adaptation effect and active protection effect are as follows:

[0228]

[0229] Number of fused data samples;

[0230] : The magnitude of the fused feature vector of the i-th sample;

[0231] : Scene standard feature vector magnitude, retrieved from the scene feature database.

[0232] Threat identification accuracy (

[0233]

[0234] A true positive result indicates an actual threat that is identified as a threat.

[0235] True negative, meaning not actually a threat and identified as non-threat;

[0236] False positives: Detected as threats when they are not actually threats.

[0237] False negative: The actual threat was not identified.

[0238] Positioning accuracy (

[0239]

[0240] Number of location samples;

[0241] : The result of the j-th localization attempt;

[0242] Threatens actual location, manual calibration or high-precision benchmark positioning.

[0243] Prediction accuracy (

[0244]

[0245] Number of threat prediction samples;

[0246] : The predicted evolution time of the k-th threat, such as the time it takes for it to spread to a specified range;

[0247] : The actual evolution time of the kth threat.

[0248] Based on the evaluation results, the core modules of the system are optimized in a targeted manner. If the scene adaptation effect is poor, the sensor deployment parameters in the deployment phase, the hybrid domain attention weight rules in the data fusion phase, and the scene parameters of the digital twin model are adjusted. If the active protection effect is not good, the threat prediction model parameters in the active protection prediction phase and the scene-specific protection strategy library are updated.

[0249] Simultaneously, the threat data and protection effect data after handling will be added to the scenario feature database during the deployment phase, providing data support for the iterative optimization of the digital twin model and the calibration update of the sensor array, thereby continuously improving the system's scenario adaptability, threat identification accuracy and active protection effect.

[0250] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0251] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A real-time security protection method based on a sensor array, characterized in that, It includes the deployment phase, the data collection and preprocessing phase, the data fusion phase, the threat identification phase, the proactive protection and prediction phase, the location and tracing phase, and the optimization and iteration phase; Deployment phase: Deploy a multimodal sensor array and establish a digital twin model to verify the blind-spot-free coverage effect of the deployment scheme and compensate for environmental errors through a cross-calibration mechanism; Data acquisition and preprocessing stage: A three-level collaborative computing architecture is adopted to realize synchronous data acquisition from sensors, perform noise reduction and encryption, and then upload the data to the blockchain for evidence storage. The data is encapsulated into a unified format and transmitted to the edge gateway. Data fusion stage: Based on the real-time scene status of the digital twin model, high-dimensional features of sensor data are extracted from multiple dimensions through the scene adaptive fusion model, the weights of each sensor data are dynamically allocated, and a global feature vector adapted to the scene is generated. Threat identification phase: Based on scenario-specific threat sample datasets, a lightweight deep learning model is trained and optimized to achieve real-time identification of scenario threats; Proactive protection prediction phase: Establish a proactive protection system, predict threat development trends by combining threat types and temporal characteristics, establish a hierarchical dynamic protection strategy library, and adjust protection actions according to the evolution of threats; Location and tracing phase: Establish a fusion positioning model of BeiDou, UWB and inertial navigation, combine digital twin scene topology features to correct the positioning algorithm, and complete the full-link source tracing of threats based on time-series evidence data; Optimization and iteration phase: Collect scenario adaptation effect data and active protection effect data, optimize system parameters based on evaluation results, and supplement data for digital twin model optimization.

2. The real-time security protection method based on sensor array according to claim 1, characterized in that, The deployment phase establishes a three-in-one multimodal sensor array deployment system integrating digital twin modeling, scenario adaptation design, and precise deployment implementation. The multimodal sensors include vibration sensors, acoustic sensors, infrared thermal imaging sensors, microwave radar sensors, BeiDou and UWB dual-mode positioning sensors, and environmental perception sensors. The digital twin model integrates scene topology parameters, threat type distribution, environmental interference characteristics, and historical protection data to build a scene feature database, which includes key parameter dimensions such as electromagnetic interference level and signal obstruction area. By optimizing the algorithm, the risk zoning of the protected area and the sensor deployment points were optimized, and the blind-spot-free coverage effect of the deployment scheme was verified. The multi-sensor cross-calibration mechanism utilizes the complementarity of data from each sensor to dynamically correct the acquired parameters and compensate for environmental errors by combining environmental perception data.

3. The real-time security protection method based on sensor array according to claim 2, characterized in that, The acquisition and preprocessing stage achieves synchronous acquisition by the sensors through a time synchronization protocol. The three-level collaborative computing architecture includes an edge layer, a fog node layer, and a cloud node layer. The edge layer performs preliminary preprocessing of the raw data, removing environmental noise and sensor noise. The fog node layer completes data standardization, data dimensionality reduction, and data encryption. The cloud node layer enables distributed data storage backup and critical data authentication; After preprocessing, the data is transmitted to the edge gateway.

4. The real-time security protection method based on sensor array according to claim 3, characterized in that, The fusion model in the data fusion stage includes three modules: scene recognition, feature matching, and weight optimization. By combining the scene recognition algorithm with the real-time scene status of the digital twin model, the current protection scene type is determined. Secondly, the hybrid domain attention module extracts high-dimensional features from the sensor data from four dimensions: spatial domain, temporal domain, frequency domain, and semantic domain. Finally, the reinforcement learning weight allocation module dynamically adjusts the weight ratio of the data based on scene-specific threat features. The system enhances threat-related features and suppresses environmental interference features through an attention mechanism, and then integrates them to generate a global feature vector and autonomously optimizes the weight allocation strategy.

5. The real-time security protection method based on sensor array according to claim 4, characterized in that, The scenario-specific threat sample dataset is constructed by fusing virtual threat samples generated by a digital twin model with real samples; Transfer learning is used to train and optimize the lightweight model in different scenarios, thereby reducing the computational complexity of the model. An adaptive update mechanism for scenarios is added, which incrementally trains the model by adding new threat samples and iterative virtual samples generated by digital twins, thereby improving the model's ability to identify new threats.

6. The real-time security protection method based on sensor array according to claim 5, characterized in that, The proactive protection system comprises four modules: threat prediction, virtual simulation, level assessment, and dynamic adaptation of protection strategies. It uses a time-series prediction model combined with a digital twin model to predict future threat trends and performs threat evolution simulation in the digital twin model to simulate the implementation effects of different protection strategies. Based on threat type, prediction results, simulation effects, and scenario characteristics, a scenario-adaptive threat level assessment system is established, which classifies threats into three levels: potential threats, ongoing threats, and urgent threats. A reinforcement learning algorithm is introduced to dynamically adjust the sequence of protective actions according to the real-time evolution of the threat, thereby achieving proactive prediction and response.

7. The real-time security protection method based on sensor array according to claim 6, characterized in that, The fusion positioning model combines the spatial distribution information of the sensor array, the topological features of the digital twin scene, and multimodal temporal fusion data; by comparing the time difference and signal strength attenuation pattern of threat signals collected by different sensor nodes, the positioning algorithm is corrected by combining the scene topological parameters in the digital twin model. Threat attribution is based on blockchain-stored time-series and location data to trace the path of threat spread, generate attribution reports that include timestamps, location coordinates, and the threat evolution process, and feed the attribution data back to the digital twin model to optimize protection strategies.

8. The real-time security protection method based on sensor array according to claim 7, characterized in that, The optimization iteration phase establishes a full-link closed-loop optimization system. The scenario adaptation effect data includes data fusion accuracy, threat identification accuracy, and positioning accuracy. The active protection effect data includes threat handling time, whether secondary losses are avoided, and prediction accuracy. Based on the evaluation results, we will optimize sensor deployment parameters, fusion model weight rules, threat prediction model parameters, and protection strategy library in a targeted manner. The threat data and protection effectiveness data after the treatment are added to the scene feature database for iterative optimization of the digital twin model and calibration update of the sensor array.

Images