All-domain full-automatic intelligent safety system and control method

Abstract

The invention provides a global full-automatic intelligent safety system and a control method. The global full-automatic intelligent safety system comprises monitoring equipment, a sensor, communication equipment, an alarm device and a server. Wherein the monitoring equipment and the sensor are connected with the server through the communication equipment; the server is provided with an edge server and a cloud server which form a collaborative architecture; the edge server is deployed with multiple containers, and independent image recognition modules are packaged in the containers; the edge server is also connected with the alarm device; the cloud server is provided with a shadow system of an edge server, a training module and an online upgrading module. In combination with the corresponding control method, the system is simple in structure and low in deployment and maintenance cost, local quick response is achieved through the edge and cloud server collaborative architecture, effective and quick updating is carried out through the cloud under the corresponding control method, and therefore stable operation of the safety system and the recognition accuracy of monitoring are guaranteed, and the safety performance of the system is improved. And the requirements of different scenes can be quickly adapted, and the system use difficulty is simplified.

Description

Technical Field

[0001] The present invention relates to the field of intelligent security technologies, and particularly to a full-domain fully automatic intelligent security system and a control method thereof. Background Art

[0002] In common industrial scenarios, safety management is of crucial importance. Especially in areas and time periods without the presence of safety officers, serious safety accidents are likely to occur due to the negligence, violation, or non-compliance of individual personnel; while traditional safety management methods mainly rely on manual inspections and monitoring systems with fixed rules, and there are the following problems:

[0003] 1. The efficiency of manual inspections is low, it is difficult to comprehensively cover all areas in real time, and it is easy to have supervision blind spots, resulting in safety hazards that cannot be discovered and handled in a timely manner.

[0004] 2. The monitoring system with fixed rules lacks flexibility and intelligence, cannot automatically identify and adapt according to different working environments and scenarios, and is prone to false alarms or missed alarms, affecting the accuracy and effectiveness of safety management.

[0005] 3. The existing image recognition technology is not yet mature enough in industrial scenarios, especially in terms of full-domain active perception and warning. There is a lack of effective solutions to meet the increasingly complex safety management requirements. For example, Chinese Patent Application CN118411121A discloses a safety management method based on AI technology, including installing cameras at important positions and near required equipment at the construction site; installing a meteorological module in the dust monitoring equipment at the construction site; configuring positioning cards for objects to be managed at the construction site, and editing all information of the parties in the cards, and installing equipment such as a positioning engine, point beacons, and base stations at the construction site; installing angle sensors on the construction machinery at the construction site; installing a self-developed AI algorithm on the server of the company or group; through fusion, performing AI intelligent management on behaviors such as safety helmet detection, reflective vest detection, smoking detection, falling detection, fire detection, and work uniform recognition. Although the above technical solution can achieve fully automated safety monitoring at the principle level, it can be known that the workload during the initial system setup period is large, and a large number of specifically trained models need to be deployed on the cloud server to achieve corresponding multi-scenario AI recognition. Not only is the training volume large and the cost high; and there are also communication problems such as network latency and bandwidth congestion in the cloud deployment scenario, affecting the actual use experience and prone to lag problems. Further, the above-mentioned existing technologies usually lack corresponding means of autonomous learning and upgrade, and it is difficult to adapt to the problem of scenario changes. Especially in some construction site scenarios, the on-site scenario images and safety rules change over time. If manual re-annotation, re-training, and temporary stop and redeployment of the cloud model are required during the change, it will waste a lot of manpower and create security vulnerabilities.

[0006] In summary, it is still necessary to propose a security system and corresponding implementation and control methods to achieve full-domain lightweight and non-intrusive deployment, automatically adapt to multi-scenario requirements, reduce deployment and later maintenance costs, and improve the efficiency of security monitoring. Summary of the Invention

[0007] In view of the problems existing in the prior art, the present invention provides a technical solution with a simple system structure and low implementation cost, which can achieve full-domain fully automatic security monitoring and alarming.

[0008] To achieve the above object, the technical solution adopted by the present invention is as follows:

[0009] On the one hand, the present invention provides a full-domain fully automatic intelligent security system, which mainly includes monitoring devices, sensors, communication devices, alarm devices, and servers;

[0010] The monitoring devices are distributed at multiple points and are connected to the server through the communication devices;

[0011] The sensors are correspondingly installed at detection positions and are connected to the server through the communication devices;

[0012] The server is provided with an edge server and a cloud server; the edge server and the cloud server form a collaborative architecture;

[0013] Multiple containers are deployed on the edge server, and at least an independent image recognition module is encapsulated in each container; the edge server is connected to the alarm device;

[0014] The cloud server is deployed with a shadow system, a training module, and an online upgrade module of the edge server; the training module is at least used for retraining the image recognition model in the image recognition module; the online upgrade module is at least used for online upgrading of the image recognition module.

[0015] Optionally, the sensors include environmental parameter sensors and device status and operation parameter sensors.

[0016] Optionally, the communication devices include cables, switches, data collectors, gateways, and wireless communication modules.

[0017] Optionally, the alarm device includes a sound and light alarm or a sound amplification device.

[0018] On the other hand, the present invention also provides a control method for a full-domain fully automatic intelligent security system, which mainly includes the following steps:

[0019] Build an automated interconnected full link of sensors and monitoring devices covering the application scenario based on the industrial Internet of Things architecture;

[0020] The sensor and the monitoring device transmit the collected information to the server through the automated interconnected full link;

[0021] The edge server in the server is configured with an image recognition module and a rule set. The image recognition module obtains the monitoring scenario corresponding to the monitoring device based on the scenario recognition model; the image recognition module also constructs a three-dimensional model of the monitoring scenario based on the image recognition algorithm and the position calculation algorithm, and obtains the motion states and motion positions of people and objects in the monitoring scenario; filters the rule set based on the monitoring scenario to obtain the alarm judgment rule for the monitoring scenario; compares the motion states and motion positions of people and objects in the monitoring scenario by combining the alarm judgment rule of the monitoring scenario to obtain the alarm information in real time;

[0022] Send the alarm information to the alarm device, and the alarm device performs an alarm operation based on the alarm information;

[0023] The cloud server in the server is configured with a shadow system of the edge server based on the digital twin technology, and pulls the image data of the monitoring device; the training module of the cloud server retrains the scenario recognition model in the cloud server based on the image data of the monitoring device, and obtains an updated model; the online upgrade module of the cloud server upgrades the scenario recognition model of the edge server based on the updated model.

[0024] Optionally, the construction method of the automated interconnected full link includes:

[0025] After the monitoring device and the sensor are powered on, they broadcast autonomously or through a communication device;

[0026] The edge computing gateway in the communication device scans the devices in the subnet and automatically groups them according to the device tags of the devices in the subnet to establish the automated interconnected full link.

[0027] Optionally, the method for obtaining the monitoring scenario includes the following steps:

[0028] Collect the image information and scenario tags of the preset scenario, and form a scenario training set;

[0029] Based on the scenario training set, train and generate the scenario recognition model through a deep learning algorithm;

[0030] Convert the video stream of the monitoring device into a first monitoring image set, and preprocess the images in the first monitoring image set to obtain a judgment image set;

[0031] Input the judgment image set into the scenario recognition model and output the monitoring scenario.

[0032] Optionally, the method for obtaining the updated model includes the following steps:

[0033] Convert the monitoring device video stream pulled from the shadow system into a second set of monitoring images, and preprocess the images in the second set of monitoring images to obtain a retraining image set;

[0034] Based on the retraining image set, retrain the scene recognition model of the previous time period through distillation and quantization techniques, and obtain the updated model.

[0035] Optionally, the method for obtaining the updated model further includes the following steps:

[0036] Pre-train a teacher model;

[0037] Automatically annotate the retraining image set in batches through the teacher model, and screen out the images with confidence lower than the threshold for manual annotation to obtain a retraining image set containing annotations;

[0038] Based on the distillation technique, train a student model through the intermediate layer features of the retraining image set containing annotations and the teacher model;

[0039] Based on the quantization technique, convert the weights and activation values of the student model from high-precision floating-point numbers to low-precision numerical values, and obtain a quantized student model;

[0040] Verify the metrics of the quantized student model. When the metrics of the quantized student model are improved compared to the scene recognition model of the previous time period, output the quantized student model as the updated model.

[0041] Optionally, the image preprocessing includes the following steps:

[0042] Extract the features of the image, and filter the still-frame images based on the eigenvalue comparison method.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] The system structure of the present invention is simple, and the deployment and maintenance costs are low. Through the collaborative architecture of the edge and cloud servers, local rapid response is achieved, and effective and rapid updates are carried out relying on the cloud under the corresponding control method, thereby ensuring the stable operation of the security system and the recognition accuracy of monitoring, and can quickly adapt to the needs of different scenarios and simplify the difficulty of system use. Description of the Drawings

[0045] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0046] Figure 1 is the system diagram in a specific embodiment of the present invention;

[0047] Figure 2 is the flowchart of the control method in a specific embodiment of the present invention;

[0048] Figure 3 is the flowchart of distillation quantization in a specific embodiment of the present invention;

[0049] Figure 4 is the flowchart of student model quantization in a specific embodiment of the present invention. Detailed implementation manners

[0050] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art without creative efforts based on the embodiments of the present invention belong to the scope of protection of the present invention.

[0051] It should be noted that similar reference numerals and letters denote similar items in the following drawings. Therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings.

[0052] In the description of the present invention, "a plurality of" means two or more unless otherwise specifically defined.

[0053] In the present invention, unless otherwise clearly defined and limited, terms such as "installation" and "connection" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or integrated; it can be directly connected, or indirectly connected through an intermediate medium, and can be the internal connection of two components or the interaction relationship between two components. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0054] It is worth noting that the methods used in the present invention are all conventional methods unless otherwise specified; the raw materials and devices used are all conventional commercially available products, and their sources are not specifically limited.

[0055] On the one hand, as Figure 1 shown, this embodiment provides a system including a monitoring device, a sensor, a communication device, an alarm device, and a server.

[0056] Among them, the monitoring devices are distributed at multiple points and are connected to the server through the communication device. Further, in this embodiment, the original monitoring cameras in the application scenario can be utilized. When the original monitoring cameras cannot achieve full coverage of the entire area, monitoring devices can be separately added for the uncovered areas. For example, multispectral cameras can be adopted, and a combination of visible light and infrared can be considered, supporting 360° panoramic monitoring, with built-in distortion correction algorithms; the deployment density is dynamically configured according to the risk level of the operation area. Specifically, it is recommended that the monitoring coverage radius in high-risk areas does not exceed 25m.

[0057] The sensors include environmental parameter sensors and device status and operation parameter sensors. Taking the hydropower station scenario as an example, the specific types of sensors include the original water level sensors, temperature and humidity sensors, generator set power detection sensors, and unit temperature sensors in the station. Optionally, according to on-site requirements, UWB positioning tags can be integrated into the sensors to achieve centimeter-level positioning of personnel / equipment. Each sensor is correspondingly installed at the detection position and is connected to the server through the communication device; the communication device includes cables, switches, data collectors, gateways, and wireless communication modules. Further, edge computing gateways are selected from various gateways, and industrial-grade 5G routers are used to support Mesh networking; the data collector integrates the LoRaWAN protocol.

[0058] The above-mentioned monitoring devices and sensors are respectively located in different scenarios, such as Scenario 1 - Scenario n, and the scenarios correspondingly include high altitudes, waters, machine rooms, etc., and each scenario is not limited to a single monitoring device or a single sensor.

[0059] The server is provided with an edge server and a cloud server; the edge server and the cloud server form a collaborative architecture. Optionally, the edge server in this embodiment is an edge server deployed with a neural network processor NPU, and the cloud server is a high-performance inference server accelerated by a GPU.

[0060] The edge server is connected to the alarm device; the alarm device includes an audible and visual alarm and / or a sound amplification device. Specifically, a loudspeaker can be configured in each monitoring area to broadcast a pre-set alarm audio according to the violations or irregularities detected by the system, such as "Please wear a safety helmet". In another embodiment, the security system is also connected to the handheld communication devices of safety officers or patrol officers through the internal network and conveys the alarm information; it can also be connected to the mobile communication devices of the responsible persons through the external network and convey the alarm information, so as to achieve full coverage of safety-related alarm information.

[0061] The cloud server deploys a shadow system, a training module, and an online upgrade module of the edge server; among them, the shadow system is based on the lightweight VGG-16 architecture and synchronizes the production environment video stream in real time. The training module is at least used to retrain the image recognition model in the image recognition module; the online upgrade module is at least used to perform an online upgrade on the image recognition module.

[0062] On the other hand, as Figure 2 shown, based on the above-mentioned global fully automatic intelligent security system, this embodiment also provides a corresponding control method, which mainly includes the following steps:

[0063] 1. System deployment;

[0064] Build an automated interconnected full link covering sensors and monitoring devices in the application scenario based on the industrial Internet of Things architecture; among them, the construction method of the automated interconnected full link includes:

[0065] After the monitoring device and the sensor are powered on, they broadcast autonomously or through a communication device, and publish the device ID and type label using the MQTT protocol, such as "camera 00xx - high-altitude operation area";

[0066] The edge computing gateway in the communication device scans the devices within the subnet, automatically groups them according to the device labels of the devices within the subnet to establish an automated interconnected full link, and dynamically generates an optimal transmission path based on the Dijkstra algorithm, supporting automatic detouring in case of failure.

[0067] The above initial deployment method can achieve a non-intrusive connection. When the system is connected to the scenario, it automatically associates with the original devices, so that there is no need for operators to set them individually one by one.

[0068] 2. Information collection;

[0069] After the system runs, the sensor collects data according to the sampling frequency, converts the collected data into digital signals, and transmits them to the edge server through the automated interconnected full link; the monitoring device records continuously and transmits the video stream to the edge server through the automated interconnected full link.

[0070] 3. Alarm judgment;

[0071] In the edge server system, container technology is also used to batch-configure image recognition modules. Each container encapsulates at least an independent image recognition module, and the TensorRT framework is used for model acceleration, and operator fusion optimization is carried out for the Jetson Xavier NX device. Further, for the NPU characteristics in the edge server, such as Huawei Ascend or NVIDIA Jetson, a customized quantization-aware training scheme is developed, and the model structure is adapted to the operator combination supported by the hardware, and highly optimized machine code is generated through the TVM compiler. The edge server is configured with a preset rule set. The image recognition module obtains the monitoring scenario corresponding to the monitoring device based on the scenario recognition model. The scenario recognition model uses the improved YOLOv5s and adds the CBAM attention mechanism to improve the small target detection ability. Among them, the scenario recognition in this embodiment is not only carried out once at the initial stage of system deployment, but is carried out periodically, so as to avoid the problem that the scenario changes over time affecting the rule judgment.

[0072] Further, the method for obtaining the monitoring scenario includes the following steps:

[0073] First, pre-training is carried out, that is, the image information and scenario labels of typical scenarios are collected, and a scenario training set is formed. For example, the label specification defines that "high-altitude operation" needs to include key features such as safety belts and scaffolding;

[0074] Based on the scenario training set, a scenario recognition model is trained through a deep learning algorithm. Data augmentation uses Mosaic and CutOut technologies to improve the generalization of the model;

[0075] The video stream of the monitoring device is transcoded as necessary, and multiple frames of pictures within a time period are captured according to the preset acquisition frequency, so as to be converted into a first monitoring image set, and the images in the first monitoring image set are preprocessed to obtain a judgment image set. Optionally, in order to save system resource consumption, in image preprocessing, the features of the image are first extracted, and the still-frame images are filtered based on the eigenvalue comparison method. Specifically, the Gaussian mixture model GMM background modeling and the still-frame filtering algorithm based on the Mahalanobis distance can be used;

[0076] The judgment image set is input into the scenario recognition model, and the monitoring scenario is output.

[0077] The above design can automatically match the monitoring scenario of the area where each monitoring camera is located, such as "high-altitude operation", "water area", etc. through scenario recognition, so as to provide a basis for subsequent rule selection.

[0078] The image recognition module also constructs a three-dimensional model of the monitoring scene based on image recognition algorithms and position calculation algorithms, and obtains the motion states and motion positions of people and objects in the monitoring scene. Among them, according to the number of monitoring devices existing in a single scene, the corresponding three-dimensional model construction strategy is selected. For example, in the case of a single camera, the sequential three-dimensional reconstruction method can be used. The ORB feature is used to track the inter-frame motion, and the CNN is used to complement the texture details. VideoNeRF inputs multiple frames from the same perspective and optimizes the time dimension parameters. The dense reconstruction PatchMatchNet is performed through the sparse point cloud COLMAP. In the case of multiple cameras, the multi-view three-dimensional reconstruction method can be used. MVSNet directly generates a dense depth map from multi-view images, uses COLMAP to generate a sparse point cloud, then converts it into a dense point cloud through MVSNet, and then uses NeRF + multi-camera perspectives to generate a photo-realistic texture model.

[0079] Correspondingly, in the process of constructing the above three-dimensional model, by extracting features from the image, the personnel and items in the monitoring screen are identified, and the motion direction and speed are calculated based on the position changes between two frames, so that the events occurring in the scene can be known. Therefore, there is no need to rely on a large number of complex rules and high computing power support set in advance, and only simple rules are required. For example, in the traditional technology, to detect whether a safety helmet is worn during high-altitude operation, it is necessary to identify the position relationship between the safety helmet and the personnel, and continuously judge whether it is on the head of the personnel in real time. In this embodiment, whether there is a safety helmet on the head of the personnel is initially identified by image feature recognition, and then whether the safety helmet is in a worn state can be known according to the speed and position relationship between the safety helmet and the personnel, thus reducing the occupation of system resources and being more suitable for running on edge servers. Another example is that if there is no operation plan in this area at the current time in the system, and the monitoring recognizes that someone is operating in this area, it is a violation. The above scheme can actively calculate the irradiation coordinate range of the camera and compare it with the three-dimensional model without configuring the mapping relationship between the camera and the area, so as to identify violations or irregularities.

[0080] A dataset of monitoring scene - operation plan area - safety rules is established in advance, and then the rule set is screened based on the monitoring scene to obtain the alarm judgment rule of the monitoring scene. Among them, the rule set consists of safety rules for the operation plan area, such as generally set alarm judgment rules, etc. According to the monitoring scene identified by the aforementioned monitoring device, the operation plan area to which it belongs is determined, so as to obtain the safety rule. Further, by comparing the motion states and motion positions of people and objects in the monitoring scene with the alarm judgment rule of the monitoring scene, alarm information can be obtained in real time. For example, in the monitoring scene of "water area", the safety rule requires that the personnel keep a distance of 1m from the water surface. If the edge server calculates that the distance between the personnel and the water surface is 0.9m, the alarm rule is triggered, thus generating alarm information.

[0081] 4. Broadcast the alarm;

[0082] Send the alarm information to the alarm device, and the alarm device performs an alarm operation based on the alarm information, such as playing "Please stay away from the water surface for on-site personnel" through the on-site loudspeaker.

[0083] 5. System update;

[0084] The cloud server configures a shadow system of the edge server based on digital twin technology and pulls the image data of the monitoring device. In this embodiment, the video stream of the monitoring device is selected.

[0085] The cloud server also deploys mixed-precision inference, with optional combinations of two data precision formats, FP16 and INT8. Combined with the layer fusion technology of TensorRT to eliminate redundant memory copies, and using CUDAGraph to capture the computational flow to reduce kernel startup overhead. Its training module retrains the scene recognition model in the cloud server based on the image data of the monitoring device and obtains an updated model; among them, as Figure 3 、 4 shown, the method for obtaining the updated model includes the following steps:

[0086] Convert the video stream of the monitoring device pulled in the shadow system into a second set of monitoring images, and preprocess the images in the second set of monitoring images to obtain a retraining image set; optionally, as described above, in image preprocessing, first extract the features of the images and filter the still-frame images based on the eigenvalue comparison method.

[0087] Based on the retraining image set, retrain the scene recognition model of the previous time period through distillation and quantization techniques and obtain an updated model; specifically, pre-train a teacher model, and in this embodiment, select the teacher model as ResNet101;

[0088] Automatically annotate the retraining image set in batches through the teacher model, and screen out the images with a confidence level lower than the threshold for manual annotation to obtain a retraining image set containing annotations;

[0089] Based on the distillation technology, train a student model through the intermediate layer features of the retraining image set containing annotations and the teacher model. The corresponding student model is MobileNetV3;

[0090] Based on the quantization technology, convert the weights and activation values of the student model from high-precision floating-point numbers to low-precision values and obtain a quantized student model;

[0091] Verify the metrics of the quantized student model. When the metrics of the quantized student model are improved compared to the scene recognition model of the previous time period, output the quantized student model as the updated model.

[0092] The online upgrade module of the cloud server upgrades and updates the scenario recognition model of the edge server based on OTA technology. The OTA update uses TLS1.3 for encrypted transmission and SHA-256 for hash integrity verification. Optionally, in this embodiment, in response to the problem of long downtime of the edge server caused by the upgrade process, targeted improvements are also made. The cloud server automatically calculates the parameter differences between the new and old models based on the updated model generated by model distillation and generates a lightweight patch package, the size of which is only 15%-20% of the full model. Specifically, model pruning combined with quantization-aware training technology is used to compress the volume of the patch package. MD5 hash verification is used to ensure the integrity of the patch, and AES-256 encryption is combined to ensure transmission security. The above work can compress the YOLOv5s model in this embodiment from 50MB to 8MB, reducing the transmission time by 84%. Thus, on the edge server, container orchestration technology is adopted to split the image recognition module into independent microservices, so as to achieve independent upgrade of each service, reduce the upgrade risk and improve the upgrade speed. For example, image recognition is split into three microservices: preprocessing - feature extraction - postprocessing. When the feature extraction microservice is upgraded, only the corresponding container needs to be restarted, and the others remain running.

[0093] Such an update strategy can not only adapt to changes in the monitoring scenario, but also effectively reduce the model scale of the edge server by means of distillation and quantization while ensuring the accuracy, thereby saving system resources, reducing the requirements for setup and configuration, and controlling the overall cost.

[0094] The solution of this embodiment first dynamically binds three-dimensional space calculation and security rules to achieve a closed-loop intelligent decision-making of "environment - rules - alarm", and for the first time integrates technologies such as cloud-edge collaboration, three-dimensional modeling, dynamic rule generation, and automated deployment into the field of industrial safety monitoring to form a complete intelligent security closed-loop. Thus, in response to the requirements of high concurrency, low latency, and dynamic environment in industrial scenarios, a practical solution is proposed, filling the gaps in the existing technology.

[0095] Finally, it should be noted that the above content is only used to illustrate the technical solution of the present invention, rather than a limitation on the protection scope of the present invention. Any simple modification or equivalent replacement of the technical solution of the present invention by those of ordinary skill in the art shall not depart from the essence and scope of the technical solution of the present invention.

Claims

1. An all-region fully automatic intelligent security system, characterized in that: It includes monitoring devices, sensors, communication devices, warning devices and servers; The monitoring devices are distributed at multiple points and are connected to the server through the communication devices; The sensors are correspondingly installed at the detection positions and are connected to the server through the communication devices; The server is provided with an edge server and a cloud server; the edge server and the cloud server form a collaborative architecture; The edge server deploys multiple containers, and at least one independent image recognition module is encapsulated in each container; the edge server is connected to the warning device; The cloud server deploys a shadow system, a training module and an online upgrade module of the edge server; the training module is at least used to retrain the image recognition model in the image recognition module; the online upgrade module is at least used to perform online upgrade on the image recognition module.

2. The full-domain fully automatic intelligent security system according to claim 1, characterized in that: The sensors include environmental parameter sensors and device status and operation parameter sensors.

3. The full-region fully automatic intelligent security system according to claim 1, characterized in that: The communication devices include cables, switches, data collectors, gateways and wireless communication modules.

4. The full-domain fully automatic intelligent security system according to claim 1, wherein: The warning device includes a sound and light alarm or a sound amplification device.

5. A control method for a global fully automatic intelligent security system, characterized in that: It includes the following steps: Based on the industrial Internet of Things architecture, construct an automated interconnected full link of sensors and monitoring devices covering the application scenario; The sensors and the monitoring devices transmit the collected information to the server through the automated interconnected full link; The edge server in the server is configured with an image recognition module and a rule set. The image recognition module obtains the monitoring scenario corresponding to the monitoring device based on the scenario recognition model; the image recognition module also constructs a three-dimensional model of the monitoring scenario based on the image recognition algorithm and the position calculation algorithm, and obtains the motion states and motion positions of people and objects in the monitoring scenario; filter the rule set based on the monitoring scenario to obtain the alarm judgment rule of the monitoring scenario; compare the motion states and motion positions of people and objects in the monitoring scenario by combining the alarm judgment rule of the monitoring scenario to obtain alarm information in real time; Send the alarm information to the warning device, and the warning device performs an alarm operation based on the alarm information; The cloud server in the server is configured with a shadow system of the edge server based on the digital twin technology, and pulls the image data of the monitoring device; The training module of the cloud server retrains the scenario recognition model in the cloud server based on the image data of the monitoring device and obtains an updated model; the online upgrade module of the cloud server upgrades the scenario recognition model of the edge server based on the updated model.

6. The control method of the global fully automatic intelligent security system according to claim 5, characterized in that: The construction method of the automated interconnected full link includes: After the monitoring devices and the sensors are powered on, they broadcast autonomously or through the communication devices; The edge computing gateway in the communication device scans the devices in the subnet and automatically groups them according to the device tags of the devices in the subnet to establish the automated interconnected full link.

7. The control method of the global fully automatic intelligent security system according to claim 5, characterized in that: The method for obtaining the monitoring scenario includes the following steps: Collect the image information and scenario tags of the preset scenario and form a scenario training set; Based on the scenario training set, train and generate the scenario recognition model through the deep learning algorithm; Convert the video stream of the monitoring device into a first set of monitoring images, and preprocess the images in the first set of monitoring images to obtain a set of judgment images; Input the set of judgment images into the scene recognition model and output the monitoring scene.

8. The control method of the global fully automatic intelligent security system according to claim 5, characterized in that: The method for obtaining the updated model includes the following steps: Convert the video stream of the monitoring device pulled from the shadow system into a second set of monitoring images, and preprocess the images in the second set of monitoring images to obtain a set of retraining images; Based on the set of retraining images, retrain the scene recognition model of the previous time period through distillation and quantization techniques, and obtain the updated model.

9. The control method of the global fully automatic intelligent security system according to claim 8, characterized in that: The method for obtaining the updated model further includes the following steps: Pre-train a teacher model; Automatically annotate the set of retraining images in batches through the teacher model, and screen out the images with confidence lower than the threshold for manual annotation to obtain a set of retraining images with annotations; Based on the distillation technique, train and obtain a student model through the intermediate layer features of the set of retraining images with annotations and the teacher model; Based on the quantization technique, convert the weights and activation values of the student model from high-precision floating-point numbers to low-precision numerical values, and obtain a quantized student model; Verify the metrics of the quantized student model. When the metrics of the quantized student model are improved compared with the scene recognition model of the previous time period, output the quantized student model as the updated model.

10. The control method of the global fully automatic intelligent security system according to claim 7 or 8, characterized in that: Image preprocessing includes the following steps: Extract the features of the image, and filter the still-frame images based on the feature value comparison method.

Images