Substation airspace foreign object intrusion safety sensing method and system based on SSD algorithm
By using a multi-source sensor cluster and an improved SSD algorithm, combined with attention mechanisms and multi-scale feature enhancement, the problem of accurate detection and rapid response of substation security systems in harsh environments has been solved, achieving all-weather and efficient security perception of airspace intrusion.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- XINYANG POWER SUPPLY OF HENAN ELECTRIC POWER CORP
- Filing Date
- 2026-01-19
- Publication Date
- 2026-04-24
AI Technical Summary
Existing substation security systems are inadequate in terms of environmental adaptability, intelligence level, and response speed, especially in severe weather conditions where they struggle to achieve accurate and real-time detection and effective linkage of airspace intrusions.
Data is collected using a multi-source sensor cluster. Target recognition and risk assessment are performed through an improved SSD algorithm. By combining attention mechanism, multi-scale feature enhancement, and CIoU loss function, high-precision detection of small and occluded targets is achieved. Intelligent risk assessment is performed based on target behavior, triggering graded early warnings and coordinated responses.
It achieves high-precision airspace intrusion detection around the clock, improving the system's environmental adaptability and intelligence, shortening response time, and increasing the accuracy and efficiency of alarms.
Smart Images
Figure CN121921890A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of substation safety protection technology, specifically relating to a substation airspace intrusion security perception method and system based on the SSD (Single Shot MultiBox Detector) algorithm. Background Technology
[0002] Substations are critical nodes in the power system, and their safe and stable operation is of paramount importance. Intrusion of foreign objects (such as floating debris, balloons, drones, birds, and small animals) into the substation airspace can cause serious accidents such as short circuits and power outages. Traditional substation security relies primarily on physical fencing, manual inspections, and fixed video surveillance. Significant drawbacks include: manual inspections cannot achieve 24 / 7 real-time coverage; fixed video surveillance is susceptible to lighting and weather conditions, relies on manual monitoring, is inefficient, has a slow response time (averaging over 30 minutes), and a high rate of false alarms and missed alarms. In recent years, traditional computer vision methods have been attempted for substation security, but they still have shortcomings: poor environmental adaptability, drastically reduced detection performance at night, in rain, snow, fog, or other adverse weather conditions, and weak ability to identify small or occluded targets. While simple multi-sensor overlay schemes increase data dimensionality, they lack effective fusion mechanisms and lightweight models, leading to system redundancy and heavy data processing burdens, failing to achieve a closed-loop intelligent response from perception to early warning to coordinated action. Therefore, there is an urgent need for a security sensing system for airspace intrusion that can achieve accurate, real-time, and adaptive sensing in complex environments and effectively link with existing substation systems. Summary of the Invention
[0003] The purpose of this invention is to overcome the shortcomings of the prior art and provide a substation airspace intrusion security perception method and system based on the SSD algorithm, thus solving the technical problems mentioned in the background art.
[0004] The objective of this invention is achieved as follows: a substation airspace intrusion security perception method based on the SSD algorithm, comprising the following steps: S1, real-time acquisition of environmental data through a multi-source sensor cluster deployed at the substation perimeter; the multi-source sensor cluster includes at least a high-definition visible light camera, an infrared thermal imager, and a lidar; S2, preprocessing and spatiotemporal registration of the acquired environmental data to form a fused data stream; S3, inputting the fused data stream into an improved SSD target detection model for airspace intrusion target identification; the improved SSD model optimizes the detection capability for small targets and occluded targets by introducing an attention mechanism and a multi-scale feature enhancement module; S4, if an intrusion target is identified, its trajectory, speed, and morphological features are extracted, and the threat level is determined based on preset risk assessment rules; S5, a graded early warning is triggered according to the threat level, and the early warning information is sent to the substation security system.
[0005] By constructing a complete intelligent security closed loop encompassing perception, judgment, and response, the principle is to overcome the limitations of single sensors by utilizing multi-source sensors. An improved SSD algorithm enables accurate and robust intrusion target identification, and intelligent risk assessment is performed based on target behavior, ultimately driving a tiered automated response. In implementation, high-definition visible light cameras, infrared thermal imagers, and lidar are deployed around the substation perimeter to form a collaborative observation network, capturing target texture, thermal radiation, and 3D point cloud information. Heterogeneous data is synchronized in time, aligned spatially, and fused at the feature level to form a more comprehensive and complementary "fused data stream," providing high-quality input for subsequent identification. This fused data stream is then input into a pre-trained, specially optimized SSD target detection model. This model quickly and accurately selects "out-of-air" targets, such as drones, balloons, and birds, from complex backgrounds. The identified targets are tracked, and their trajectory, speed, size, and other characteristics are analyzed. Combined with their distance from critical equipment, behavioral intent (such as whether they are pointing at core equipment), and historical case matching, a quantified threat level is calculated. Based on the determined threat level, the system automatically executes different levels of response measures, ranging from "record only" to "audio-visual alarm + device linkage," and pushes the early warning information to the security system, thus solving the technical problems of poor environmental adaptability, low level of intelligence, and delayed response in existing technologies.
[0006] In step S2, the preprocessing and spatiotemporal registration of the collected environmental data specifically includes: performing Gaussian filtering and histogram equalization on the visible light video stream to enhance image quality; using a temperature threshold segmentation method to extract abnormal heat source areas from the infrared thermal imaging data; generating a depth map based on the lidar point cloud data and unifying the multi-source data to the same spatiotemporal coordinate system through coordinate transformation; and using an adaptive weighted algorithm to fuse visible light features, thermal infrared features, and depth features to output the fused data stream. The principle is to extract effective information using the most suitable preprocessing method for the data characteristics of different sensors, and then integrate multi-dimensional information into a coherent and enhanced data representation through coordinate unification and weighted fusion.
[0007] During implementation, Gaussian filtering and histogram equalization are used to improve image quality in visible light video; temperature threshold segmentation is used to directly extract potential thermal targets from infrared data; and depth maps are generated from LiDAR point clouds to provide distance information. Using calibration parameters, images and point cloud data from different sensors and coordinate systems are mapped to the same unified world coordinate system and time series, ensuring that information from the same spatiotemporal point is "seen." This is not a simple superposition, but rather a dynamic adjustment of the weights of visible light, infrared, and depth features based on the environment (e.g., day / night, clear / foggy). For example, the contribution of infrared and depth features is enhanced in foggy weather to generate an optimal "fused data stream." This solves the blind spot problem of single sensors (e.g., visible light cameras failing at night), achieving all-weather, all-time perception.
[0008] In step S3, the improved SSD object detection model specifically includes: introducing a dual mechanism of channel attention and spatial attention between the convolutional feature layers of the SSD model; the channel attention module generates a channel weight matrix through average pooling and max pooling operations, and the spatial attention module generates a spatial weight matrix through convolutional dimensionality reduction and activation functions; adding an ultra-small object detection layer to the feature pyramid network to detect intrusion targets with pixel areas below a set threshold; replacing ordinary convolution with an MSConv module combining depthwise separable convolution and shuffling operations to reduce the number of model parameters; and using the CIoU loss function instead of the traditional Smooth L1 loss function to comprehensively consider factors such as the distance between the center points of the predicted box and the ground truth box and the aspect ratio, thereby improving localization accuracy. In implementation, a channel attention module (focusing on "what are the important feature channels") and a spatial attention module (focusing on "where are the important regions in the feature map") are added after the feature layer, enabling the model to adaptively strengthen features related to the target. A dedicated detection head is added to the shallow layers (high resolution, many details) of the feature pyramid network (FPN) to detect targets with very small pixel areas (such as distant birds). The lightweight MSConv module replaces some standard convolutions with depthwise separable convolutions, significantly reducing computation and parameters. Combined with channel shuffling, it promotes information exchange between channels, maintaining model performance while achieving lightweight design. The CIoU loss function is employed, considering not only the overlap area between the predicted and ground truth boxes during training, but also the distance between their center points and aspect ratio, resulting in more accurate bounding boxes predicted by the model.
[0009] In step S4, the threat level determination based on preset risk assessment rules specifically includes: dynamically adjusting the risk coefficient according to the distance between the intrusion target and the critical electrical equipment; the closer the distance, the higher the risk coefficient; predicting the target's intent based on its movement trajectory; if the trajectory points to core equipment such as transformers or circuit breakers, the threat level is increased; and using a historical intrusion event database for similarity matching; if the current target characteristics match historical high-risk events, a high-level warning is triggered. The principle is to mimic the judgment logic of human security experts, conducting a multi-factor, dynamic, comprehensive risk assessment of the intrusion target from three dimensions: spatial distance, behavioral intent, and historical experience. During implementation: Distance risk calculation: Real-time calculation of the Euclidean distance between the target and the nearest critical equipment (such as a transformer); the closer the distance, the non-linearly increasing risk coefficient (e.g., set according to a reciprocal or exponential function). Intent risk analysis: By analyzing the direction and speed changes of the target's movement trajectory (e.g., hovering, rapid approach), it is determined whether its behavior has malicious or high-risk tendencies. If the trajectory continuously points to core equipment, it is determined to be a high-risk intent. Historical Matching and Comparison: The current target's type, size, speed, and other characteristics are compared with historical alarms / incident databases in the cloud. If the characteristics are highly similar to historical cases that caused faults, the risk level is directly increased. This avoids ineffective alarms caused by overreacting (e.g., a plastic bag floating in the distance being judged as a low-level threat) and allows for targeted warnings of truly high-risk behaviors (e.g., a drone approaching the main line). This significantly improves the accuracy and usability of alarms, reduces the burden on maintenance personnel, and makes security responses more precise and efficient.
[0010] In step S5, the tiered early warning includes: low-level threats: recording intrusion events through system logs and saving relevant video clips; medium-level threats: sending alarm information to the monitoring center and automatically controlling the PTZ camera to track and zoom in on the target; high-level threats: immediately activating the audible and visual alarm device and linking with the substation automation system to perform power outage isolation operations on the intruded area. This tiered model of "low-level recording, medium-level tracking, and high-level linkage" avoids interference from low-level events, ensuring that personnel can concentrate on dealing with truly medium- and high-level threats. In particular, the linkage with the automation system elevates security from "auxiliary monitoring" to a "proactive defense" level that can work in conjunction with the production control system, greatly shortening the critical time from hazard detection to response and improving the inherent safety level of the substation.
[0011] A substation airspace intrusion security sensing system based on the SSD algorithm includes: a data acquisition module, comprising a multi-source sensor cluster deployed around the substation perimeter, signal conditioning circuitry, and a data acquisition card; an edge computing node module, equipped with a dual-core FPGA and DSP processor, for running an improved SSD target detection model to achieve real-time identification and tracking of intrusion targets; a cloud analysis platform for storing historical data, training deep learning models, and performing big data trend analysis; and an early warning response module, including an alarm information push unit, an automatic control command generation unit, and a human-machine interface. This system architecture achieves reasonable allocation of computational load, with edge computing ensuring real-time performance and reliability (network-independent); and the cloud platform providing powerful computational intelligence and centralized management capabilities. This architecture makes the system easy to deploy on a large scale across multiple substations, and the failure of a single substation does not affect the cloud or other substations. The system as a whole is highly scalable and easy to maintain.
[0012] Furthermore, the data acquisition module also integrates an environmental adaptability unit, used to dynamically adjust sensor operating parameters based on real-time meteorological data; automatically switch to infrared thermal imaging mode and enhance image contrast in hazy weather; and activate supplementary lighting at night, fusing lidar point cloud data to assist visual recognition. Its principle is to dynamically perceive changes in the external environment by integrating environmental sensors or acquiring meteorological data, and automatically adjust the operating mode of the front-end sensors and the back-end processing strategy accordingly to maintain optimal perception at all times. Through built-in meteorological sensors or access to meteorological services, it acquires real-time information such as light intensity, visibility, and weather conditions (sunny, foggy, rainy, snowy, night). This solves the pain point of traditional video surveillance being dependent on weather conditions, achieving stable performance output under all-weather, all-climate conditions. The system no longer passively receives images of varying quality but actively adapts to the environment, selecting the optimal combination of "sensory" inputs and processing parameters, ensuring that the system's perception capability does not decline under any adverse weather conditions.
[0013] Furthermore, the edge computing node module adopts an FPGA+DSP coprocessing architecture. The FPGA is responsible for parallel preprocessing of sensor data and accelerating convolution operations in the SSD model, while the DSP is responsible for executing complex post-processing logic and early warning judgments. The principle is to leverage the parallel pipeline processing capabilities of the FPGA (Field-Programmable Gate Array) and the flexible scalar computing capabilities of the DSP (Digital Signal Processor) for collaborative computing (coprocessing), assigning different tasks to the hardware most suited to achieve optimal efficiency.
[0014] Furthermore, it includes a self-learning optimization module, which adds false alarms or missed alarms to the training set and adaptively adjusts the weight coefficients of the risk assessment rules based on the early warning feedback results. The principle is to establish a closed-loop feedback system that collects "errors" (false alarms, missed alarms) generated during actual system operation, as well as human feedback. This data is used to automatically optimize the AI model and judgment rules, allowing system performance to continuously improve over time. In implementation, the system automatically collects scenario data (fusion data stream and target labels) marked as "false alarms" (system alerts but no actual threat) or "missed alarms" (an incident occurs but the system does not alert). After periodically accumulating a certain amount of new samples or in the cloud, these challenging samples are added to the training set for fine-tuning or incremental learning of the improved SSD model, enabling it to better identify targets previously misjudged. The system records the feedback from operations personnel on each alert (e.g., confirmation of validity, marking as a false alarm). By analyzing these feedbacks through algorithms (such as reinforcement learning), the weight coefficients (α, β, γ) of the risk assessment rules can be adaptively adjusted to make the risk judgment more consistent with the actual situation on site and the experience of the operation and maintenance personnel.
[0015] The beneficial effects of this invention are as follows: By fusing multi-source data from visible light, infrared, and lidar, combined with an improved SSD model, and introducing attention mechanisms, multi-scale enhancement, and a CIoU loss function, the invention significantly improves environmental adaptability under adverse weather conditions and varying lighting conditions, as well as high-precision detection capabilities for small and occluded targets. Utilizing edge computing nodes (FPGA+DSP architecture) to deploy a lightweight improved SSD model, real-time inference is performed at the substation site, greatly reducing data transmission latency and cloud computing load, meeting the real-time requirements of security response. Beyond intrusion detection, the invention further performs intelligent risk assessment through target feature extraction, trajectory prediction, and historical data matching, triggering tiered responses from log recording and camera tracking to audible and visual alarms and even power outages based on different threat levels, achieving all-weather, all-time perception. Attached Figure Description
[0016] Figure 1 This is a schematic diagram of the method steps of the present invention;
[0017] Figure 2 This is a schematic diagram of the system architecture of the present invention. Detailed Implementation
[0018] The present invention will now be described in further detail with reference to the accompanying drawings. It should be noted that this is only for the purpose of more clearly illustrating and explaining the present invention.
[0019] Example 1
[0020] like Figure 1-2As shown, this embodiment discloses a substation airspace intrusion security perception method based on the SSD algorithm, including the following steps: S1, real-time collection of environmental data through a multi-source sensor cluster deployed at the substation perimeter; the multi-source sensor cluster includes at least a high-definition visible light camera, an infrared thermal imager, and a lidar; S2, preprocessing and spatiotemporal registration of the collected environmental data to form a fused data stream; S3, inputting the fused data stream into an improved SSD target detection model for airspace intrusion target identification; the improved SSD model optimizes the detection capability for small targets and occluded targets by introducing an attention mechanism and a multi-scale feature enhancement module; S4, if an intrusion target is identified, its motion trajectory, speed, and morphological features are extracted, and the threat level is determined based on preset risk assessment rules; S5, a graded early warning is triggered according to the threat level, and the early warning information is sent to the substation security system.
[0021] By constructing a complete intelligent security closed loop encompassing perception, judgment, and response, the principle is to overcome the limitations of single sensors by utilizing multi-source sensors. An improved SSD algorithm enables accurate and robust intrusion target identification, and intelligent risk assessment is performed based on target behavior, ultimately driving a tiered automated response. In implementation, high-definition visible light cameras, infrared thermal imagers, and lidar are deployed around the substation perimeter to form a collaborative observation network, capturing target texture, thermal radiation, and 3D point cloud information. Heterogeneous data is synchronized in time, aligned spatially, and fused at the feature level to form a more comprehensive and complementary "fused data stream," providing high-quality input for subsequent identification. This fused data stream is then input into a pre-trained, specially optimized SSD target detection model. This model quickly and accurately selects "out-of-air" targets, such as drones, balloons, and birds, from complex backgrounds. The identified targets are tracked, and their trajectory, speed, size, and other characteristics are analyzed. Combined with their distance from critical equipment, behavioral intent (such as whether they are pointing at core equipment), and historical case matching, a quantified threat level is calculated. Based on the determined threat level, the system automatically executes different levels of response measures, ranging from "recording only" to "audio-visual alarm + device linkage," and pushes the early warning information to the security system, thus solving the technical problems of poor environmental adaptability, low level of intelligence, and delayed response in existing technologies.
[0022] Example 2
[0023] like Figure 1-2As shown, this embodiment discloses a substation airspace intrusion security perception method based on the SSD algorithm, including the following steps: S1, real-time collection of environmental data through a multi-source sensor cluster deployed at the substation perimeter; the multi-source sensor cluster includes at least a high-definition visible light camera, an infrared thermal imager, and a lidar; S2, preprocessing and spatiotemporal registration of the collected environmental data to form a fused data stream; S3, inputting the fused data stream into an improved SSD target detection model for airspace intrusion target identification; the improved SSD model optimizes the detection capability for small targets and occluded targets by introducing an attention mechanism and a multi-scale feature enhancement module; S4, if an intrusion target is identified, its motion trajectory, speed, and morphological features are extracted, and the threat level is determined based on preset risk assessment rules; S5, a graded early warning is triggered according to the threat level, and the early warning information is sent to the substation security system.
[0024] By constructing a complete intelligent security closed loop encompassing perception, judgment, and response, the principle is to overcome the limitations of single sensors by utilizing multi-source sensors. An improved SSD algorithm enables accurate and robust intrusion target identification, and intelligent risk assessment is performed based on target behavior, ultimately driving a tiered automated response. In implementation, high-definition visible light cameras, infrared thermal imagers, and lidar are deployed around the substation perimeter to form a collaborative observation network, capturing target texture, thermal radiation, and 3D point cloud information. Heterogeneous data is synchronized in time, aligned spatially, and fused at the feature level to form a more comprehensive and complementary "fused data stream," providing high-quality input for subsequent identification. This fused data stream is then input into a pre-trained, specially optimized SSD target detection model. This model quickly and accurately selects "out-of-air" targets, such as drones, balloons, and birds, from complex backgrounds. The identified targets are tracked, and their trajectory, speed, size, and other characteristics are analyzed. Combined with their distance from critical equipment, behavioral intent (such as whether they are pointing at core equipment), and historical case matching, a quantified threat level is calculated. Based on the determined threat level, the system automatically executes different levels of response measures, ranging from "recording only" to "audio-visual alarm + device linkage," and pushes the early warning information to the security system, thus solving the technical problems of poor environmental adaptability, low level of intelligence, and delayed response in existing technologies.
[0025] In step S2, the preprocessing and spatiotemporal registration of the collected environmental data specifically includes: performing Gaussian filtering and histogram equalization on the visible light video stream to enhance image quality; using a temperature threshold segmentation method to extract abnormal heat source areas from the infrared thermal imaging data; generating a depth map based on the lidar point cloud data and unifying the multi-source data to the same spatiotemporal coordinate system through coordinate transformation; and using an adaptive weighted algorithm to fuse visible light features, thermal infrared features, and depth features to output the fused data stream. The principle is to extract effective information using the most suitable preprocessing method for the data characteristics of different sensors, and then integrate multi-dimensional information into a coherent and enhanced data representation through coordinate unification and weighted fusion.
[0026] During implementation, Gaussian filtering and histogram equalization are used to improve image quality in visible light video; temperature threshold segmentation is used to directly extract potential thermal targets from infrared data; and depth maps are generated from LiDAR point clouds to provide distance information. Using calibration parameters, images and point cloud data from different sensors and coordinate systems are mapped to the same unified world coordinate system and time series, ensuring that information from the same spatiotemporal point is "seen." This is not a simple superposition, but rather a dynamic adjustment of the weights of visible light, infrared, and depth features based on the environment (e.g., day / night, clear / foggy). For example, the contribution of infrared and depth features is enhanced in foggy weather to generate an optimal "fused data stream." This solves the blind spot problem of single sensors (e.g., visible light cameras failing at night), achieving all-weather, all-time perception.
[0027] In step S3, the improved SSD object detection model specifically includes: introducing a dual mechanism of channel attention and spatial attention between the convolutional feature layers of the SSD model; the channel attention module generates a channel weight matrix through average pooling and max pooling operations, and the spatial attention module generates a spatial weight matrix through convolutional dimensionality reduction and activation functions; adding an ultra-small object detection layer to the feature pyramid network to detect intrusion targets with pixel areas below a set threshold; replacing ordinary convolution with an MSConv module combining depthwise separable convolution and shuffling operations to reduce the number of model parameters; and using the CIoU loss function instead of the traditional Smooth L1 loss function to comprehensively consider factors such as the distance between the center points of the predicted box and the ground truth box and the aspect ratio, thereby improving localization accuracy. In implementation, a channel attention module (focusing on "what are the important feature channels") and a spatial attention module (focusing on "where are the important regions in the feature map") are added after the feature layer, enabling the model to adaptively strengthen features related to the target. A dedicated detection head is added to the shallow layers (high resolution, many details) of the feature pyramid network (FPN) to detect targets with very small pixel areas (such as distant birds). The lightweight MSConv module replaces some standard convolutions with depthwise separable convolutions, significantly reducing computation and parameters. Combined with channel shuffling, it promotes information exchange between channels, maintaining model performance while achieving lightweight design. The CIoU loss function is employed, considering not only the overlap area between the predicted and ground truth boxes during training, but also the distance between their center points and aspect ratio, resulting in more accurate bounding boxes predicted by the model.
[0028] In step S4, the threat level determination based on preset risk assessment rules specifically includes: dynamically adjusting the risk coefficient according to the distance between the intrusion target and the critical electrical equipment; the closer the distance, the higher the risk coefficient; predicting the target's intent based on its movement trajectory; if the trajectory points to core equipment such as transformers or circuit breakers, the threat level is increased; and using a historical intrusion event database for similarity matching; if the current target characteristics match historical high-risk events, a high-level warning is triggered. The principle is to mimic the judgment logic of human security experts, conducting a multi-factor, dynamic, comprehensive risk assessment of the intrusion target from three dimensions: spatial distance, behavioral intent, and historical experience. During implementation: Distance risk calculation: Real-time calculation of the Euclidean distance between the target and the nearest critical equipment (such as a transformer); the closer the distance, the non-linearly increasing risk coefficient (e.g., set according to a reciprocal or exponential function). Intent risk analysis: By analyzing the direction and speed changes of the target's movement trajectory (e.g., hovering, rapid approach), it is determined whether its behavior has malicious or high-risk tendencies. If the trajectory continuously points to core equipment, it is determined to be a high-risk intent. Historical Matching and Comparison: The current target's type, size, speed, and other characteristics are compared with historical alarms / incident databases in the cloud. If the characteristics are highly similar to historical cases that caused faults, the risk level is directly increased. This avoids ineffective alarms caused by overreacting (e.g., a plastic bag floating in the distance being judged as a low-level threat) and allows for targeted warnings of truly high-risk behaviors (e.g., a drone approaching the main line). This significantly improves the accuracy and usability of alarms, reduces the burden on maintenance personnel, and makes security responses more precise and efficient.
[0029] In step S5, the tiered early warning includes: low-level threats: recording intrusion events through system logs and saving relevant video clips; medium-level threats: sending alarm information to the monitoring center and automatically controlling the PTZ camera to track and zoom in on the target; high-level threats: immediately activating the audible and visual alarm device and linking with the substation automation system to perform power outage isolation operations on the intruded area. This tiered model of "low-level recording, medium-level tracking, and high-level linkage" avoids interference from low-level events, ensuring that personnel can concentrate on dealing with truly medium- and high-level threats. In particular, the linkage with the automation system elevates security from "auxiliary monitoring" to a "proactive defense" level that can work in conjunction with the production control system, greatly shortening the critical time from hazard detection to response and improving the inherent safety level of the substation.
[0030] A substation airspace intrusion security sensing system based on the SSD algorithm includes: a data acquisition module, comprising a multi-source sensor cluster deployed around the substation perimeter, signal conditioning circuitry, and a data acquisition card; an edge computing node module, equipped with a dual-core FPGA and DSP processor, for running an improved SSD target detection model to achieve real-time identification and tracking of intrusion targets; a cloud analysis platform for storing historical data, training deep learning models, and performing big data trend analysis; and an early warning response module, including an alarm information push unit, an automatic control command generation unit, and a human-machine interface. This system architecture achieves reasonable allocation of computational load, with edge computing ensuring real-time performance and reliability (network-independent); and the cloud platform providing powerful computational intelligence and centralized management capabilities. This architecture makes the system easy to deploy on a large scale across multiple substations, and the failure of a single substation does not affect the cloud or other substations. The system as a whole is highly scalable and easy to maintain.
[0031] The data acquisition module also integrates an environmental adaptability unit, used to dynamically adjust sensor operating parameters based on real-time meteorological data; automatically switch to infrared thermal imaging mode and enhance image contrast in foggy or hazy weather; and activate supplementary lighting at night, integrating lidar point cloud data to assist visual recognition. Its principle is to dynamically perceive changes in the external environment by integrating environmental sensors or acquiring meteorological data, and automatically adjust the operating mode of the front-end sensors and the back-end processing strategy accordingly to maintain optimal perception at all times. Through built-in meteorological sensors or access to meteorological services, it acquires real-time information such as light intensity, visibility, and weather conditions (sunny, foggy, rainy, snowy, night). This solves the pain point of traditional video surveillance being dependent on weather conditions, achieving stable performance output under all-weather, all-climate conditions. The system no longer passively receives images of varying quality but actively adapts to the environment, selecting the optimal combination of "sensory" inputs and processing parameters, ensuring that the system's perception capability does not decline under any adverse weather conditions.
[0032] The edge computing node module adopts an FPGA+DSP coprocessing architecture. The FPGA is responsible for parallel preprocessing of sensor data and accelerating convolution operations in the SSD model, while the DSP is responsible for executing complex post-processing logic and early warning judgments. The principle is to leverage the parallel pipeline processing capabilities of the FPGA (Field-Programmable Gate Array) and the flexible scalar computing capabilities of the DSP (Digital Signal Processor) for collaborative computing (coprocessing), assigning different tasks to the hardware best suited for them to achieve optimal efficiency.
[0033] In this embodiment, the system also includes a self-learning optimization module. This module is used to add false alarms or missed alarms to the training set and adaptively adjust the weight coefficients of the risk assessment rules based on the early warning feedback results. The principle is to establish a closed-loop feedback system that collects "errors" (false alarms and missed alarms) generated during actual system operation, as well as human feedback. This data is used to automatically optimize the AI model and judgment rules, allowing system performance to continuously improve over time. In implementation, the system automatically collects scenario data (fusion data stream and target labels) marked as "false alarms" (system alerts but no actual threat) or "missed alarms" (an incident occurs but the system does not alert). After periodically accumulating a certain amount of new samples or in the cloud, these challenging samples are added to the training set for fine-tuning or incremental learning of the improved SSD model, enabling the model to better identify targets previously misjudged. The system records the feedback from operations personnel on each alert (e.g., confirmation of validity, marking as a false alarm). By analyzing these feedbacks through algorithms (such as reinforcement learning), the weight coefficients (α, β, γ) of the risk assessment rules can be adaptively adjusted to make the risk judgment more consistent with the actual situation on site and the experience of the operation and maintenance personnel.
[0034] During implementation, a sensor cluster is deployed at high points and key areas around the substation perimeter. This cluster includes: high-resolution visible light network cameras for acquiring color video streams during the day and under good lighting conditions; uncooled infrared thermal imagers for acquiring thermal radiation images at night and in foggy or hazy conditions, unaffected by visible light; and 3D lidar for acquiring precise distance, contour, and location point cloud information of intruding targets. All sensors are time-synchronized and transmit data streams to edge computing nodes via the network.
[0035] Risk assessment and threat level determination involve associating identical targets identified in consecutive frames and extracting morphological features such as motion trajectory, instantaneous velocity, and target size using algorithms like Kalman filtering. Subsequently, threat determination is performed based on pre-defined risk assessment rules.
[0036] Distance risk: Calculate the real-time distance D between the target and the nearest critical electrical equipment (such as transformers or circuit breakers). The closer the distance, the higher the risk factor R_dist (e.g., R_dist = 1 / D).
[0037] Intent Risk: Analyze the target's movement trajectory. If the trajectory direction continuously points towards the core equipment area, or exhibits suspicious behaviors such as hovering or circling, the intent risk coefficient R_intent is increased.
[0038] Historical Matching Risk: Matches the current target's characteristics (type, speed, size) with the cloud-based historical intrusion event database based on similarity. If the characteristics are highly similar to those of events that have historically caused alerts or incidents, the matching risk coefficient R_hist is triggered.
[0039] Calculate the total risk value R_total = αR_dist + βR_intent + γR_hist (α, β, γ are adjustable weights). Based on the value of R_total, the threat is divided into three levels: low, medium, and high.
[0040] Different response procedures are triggered based on the determined threat level:
[0041] Low-level threats (such as a harmless plastic bag passing by in the distance): The system automatically records the event in the log and saves a 30-second video clip before and after the event to local storage for later review, without generating an active alarm.
[0042] Intermediate threat (e.g., birds circling near equipment areas): Sends audible and visual alarms to the substation monitoring center's security platform, displaying a real-time image of the intrusion target. Simultaneously, automatically controls the nearest PTZ camera to track and optically magnify the target, providing maintenance personnel with a clearer viewing angle.
[0043] Advanced threats (such as drones approaching high-voltage busbars or small animals climbing equipment): Immediately activate the on-site high-decibel audible and visual alarms to drive them away. Simultaneously, coordinate with the substation automation system to send warning signals to the associated switching equipment in the intrusion area. If necessary (such as when the target has already touched the equipment), automatically perform power outage isolation operations and push the highest-level alarm information to the mobile APP of the relevant personnel.
[0044] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A substation airspace intrusion security sensing method based on SSD algorithm, characterized in that, Includes the following steps: S1. Real-time environmental data is collected by a multi-source sensor cluster deployed around the substation perimeter; the multi-source sensor cluster includes at least a high-definition visible light camera, an infrared thermal imager, and a lidar. S2. The collected environmental data is preprocessed and spatiotemporally registered to form a fused data stream; S3. Input the fused data stream into the improved SSD target detection model to identify intrusive targets in the airspace; the improved SSD model optimizes the detection capability of small targets and occluded targets by introducing an attention mechanism and a multi-scale feature enhancement module; S4. If an intrusion target is identified, its movement trajectory, speed and morphological characteristics are extracted, and the threat level is determined based on the preset risk assessment rules. S5. Trigger a graded early warning based on the threat level and send the warning information to the substation security system.
2. The substation airspace intrusion security sensing method based on the SSD algorithm according to claim 1, characterized in that, In step S2, the preprocessing and spatiotemporal registration of the collected environmental data specifically includes: performing Gaussian filtering and histogram equalization on the visible light video stream to enhance image quality; using a temperature threshold segmentation method to extract abnormal heat source areas from the infrared thermal imaging data; generating a depth map based on the lidar point cloud data and unifying the multi-source data to the same spatiotemporal coordinate system through coordinate transformation; and using an adaptive weighting algorithm to fuse visible light features, thermal infrared features, and depth features to output the fused data stream.
3. The substation airspace intrusion security sensing method based on the SSD algorithm according to claim 1, characterized in that, In step S3, the improved SSD target detection model specifically includes: introducing a dual mechanism of channel attention and spatial attention between the convolutional feature layers of the SSD model; the channel attention module generates a channel weight matrix through average pooling and max pooling operations, and the spatial attention module generates a spatial weight matrix through convolutional dimensionality reduction and activation functions; adding an ultra-small target detection layer to the feature pyramid network to detect intrusion targets with pixel areas below a set threshold; replacing ordinary convolution with an MSConv module that combines depthwise separable convolution with shuffling operations to reduce the number of model parameters; and using the CIoU loss function instead of the traditional Smooth L1 loss function to comprehensively consider the center point distance and aspect ratio of the predicted box and the ground truth box to improve localization accuracy.
4. The substation airspace intrusion security sensing method based on the SSD algorithm according to claim 1, characterized in that, In step S4, the determination of the threat level based on the preset risk assessment rules specifically includes: dynamically adjusting the risk coefficient according to the distance between the intrusion target and the critical electrical equipment, with the risk coefficient being higher the closer the distance; predicting the target's intention by combining the target's movement trajectory, and increasing the threat level if the trajectory points to the core equipment of the transformer or circuit breaker; and performing similarity matching using a historical intrusion event database, and triggering a high-level warning if the current target characteristics match historical high-risk events.
5. The substation airspace intrusion security sensing method based on the SSD algorithm according to claim 1, characterized in that, In step S5, the tiered early warning includes: Low-level threats: Intrusion events are logged through system logs, and relevant video clips are saved; Intermediate threat: Sends alarm information to the monitoring center and automatically controls the PTZ camera to track and zoom in on the target; Advanced threat: Immediately activate the audible and visual alarm devices and link with the substation automation system to perform power outage and isolation operations on the intruded area.
6. A substation airspace intrusion security sensing system based on the SSD algorithm, characterized in that, include: The data acquisition module includes a multi-source sensor cluster deployed around the perimeter of the substation, a signal conditioning circuit, and a data acquisition card. The edge computing node module has a built-in dual-core FPGA and DSP processor to run an improved SSD target detection model, enabling real-time identification and tracking of intrusion targets; A cloud-based analytics platform used to store historical data, train deep learning models, and perform big data trend analysis. The early warning response module includes an alarm information push unit, an automatic control command generation unit, and a human-machine interface.
7. The substation airspace intrusion security sensing system based on the SSD algorithm according to claim 6, characterized in that, The data acquisition module also integrates an environmental adaptability unit, which dynamically adjusts the sensor's operating parameters based on real-time meteorological data; automatically switches to infrared thermal imaging mode in hazy weather and enhances image contrast; and activates a supplementary lighting device at night and integrates lidar point cloud data to assist visual recognition.
8. The substation airspace intrusion security sensing system based on the SSD algorithm according to claim 6, characterized in that, The edge computing node module adopts an FPGA+DSP coprocessing architecture, in which the FPGA is responsible for parallel preprocessing of sensor data and acceleration of convolution operations in the SSD model, and the DSP is responsible for executing complex post-processing logic and early warning judgment.
9. The substation airspace intrusion security sensing system based on the SSD algorithm according to claim 6, 7, or 8, characterized in that, It also includes a self-learning optimization module, which is used to add false alarm or missed alarm samples to the training set and adaptively adjust the weight coefficients of the risk assessment rules based on the early warning feedback results.