A method and system for intelligent detection and security alarm

Abstract

This invention discloses a method and system for intelligent detection and safety alarm. The method first acquires real-time monitoring data from the factory and updates the node features of the factory monitoring graph structure. Then, the graph data is input into a pre-trained improved graph convolutional neural network model. This model propagates features through convolutional layers with integrated attention mechanisms, employs multi-scale learning to fuse local and global features, and utilizes residual connections to ensure training stability. Finally, it outputs a risk score for each node. Finally, differentiated early warnings are triggered based on the comparison between the score and a threshold. The corresponding system includes modules for data processing, analysis, and early warning. This invention also relates to a graphical user interface and storage medium for implementing this method. This invention achieves deep correlation analysis and accurate risk assessment of multi-source data, significantly improving the real-time performance, accuracy, and interpretability of factory safety monitoring.

Description

Technical Field

[0001] This invention relates to the field of industrial automation and safe production technology, and in particular to an intelligent monitoring system and method that integrates Internet of Things, computer vision and artificial intelligence technologies, especially a real-time detection and safety alarm method and system for intelligent factories based on improved graph convolutional networks. Background Technology

[0002] Safety management in modern smart factories relies on comprehensive perception of the state of complex systems. Traditional monitoring solutions have significant limitations: First, data from multiple sources, such as video surveillance and equipment sensors, are independent of each other, forming data silos and making it difficult to conduct overall risk assessments. Second, existing analysis methods (such as rule-based thresholding or traditional image recognition) cannot effectively model the complex spatial and functional relationships between multiple elements such as equipment, areas, and environment within the factory, resulting in insufficient ability to identify systemic and interconnected risks, and high false alarm and false negative rates.

[0003] In recent years, the development of deep learning technology, especially graph neural networks, has made it possible to process non-Euclidean relational data. However, standard graph convolutional networks still have shortcomings when dealing with industrial monitoring scenarios: their convolutional operations treat all related nodes equally and cannot adaptively focus on key or abnormal entities; single-scale feature extraction is difficult to capture both local anomaly details and global risk propagation patterns; and deep network training is prone to performance degradation. Therefore, how to build an intelligent monitoring system that can deeply integrate multi-source heterogeneous data, accurately model complex relationships, and achieve stable and efficient risk perception has become an urgent technical problem to be solved. Summary of the Invention

[0004] This invention aims to overcome the shortcomings of existing technologies and provide a method and system for intelligent detection and safety alarm based on an improved graph convolutional network. This invention abstracts a smart factory into a dynamic graph model and utilizes an improved graph convolutional network integrating attention mechanisms, multi-scale learning, and residual connections for deep analysis. Finally, it achieves risk visualization and interaction through a dedicated graphical user interface, thereby constructing a precise, real-time, and interpretable closed-loop safety monitoring system.

[0005] According to one aspect of the present invention, a method for intelligent detection and security alarm is provided, the method being implemented based on a pre-trained graph convolutional neural network model, comprising: The system acquires real-time monitoring data of the factory and updates the feature vectors of nodes in the factory monitoring graph structure based on this data. The nodes in the factory monitoring graph structure represent equipment, areas, or sensor entities within the factory, and the edges represent physical or functional relationships between the entities. Each node has a feature vector corresponding to the type of entity it represents, and the feature vector contains feature data extracted from the real-time monitoring data to characterize the state of the entity. The updated factory monitoring graph structure is input into a pre-trained graph convolutional neural network model, which outputs the risk score value corresponding to each node in the graph. The graph convolutional neural network model used is a network containing multiple graph convolutional layers. Feature propagation is performed by graph convolutional layers that use an attention mechanism to weight and aggregate the features of neighboring nodes. Residual connections are set between adjacent graph convolutional layers. Furthermore, the network is configured to fuse node features from different levels for multi-scale feature learning. The risk score of each node is compared with a preset threshold, and an early warning is triggered based on the comparison result.

[0006] As a further technical solution, the attention mechanism is implemented by calculating the attention weight between each node in the graph and its neighboring nodes, and by weighting and aggregating the features of the neighboring nodes based on the attention weight.

[0007] As a further technical solution, the fusion of node features from different levels for multi-scale feature learning includes: extracting local neighborhood features of nodes output from shallow graph convolutional layers and global structural features of nodes output from deep graph convolutional layers, and fusing the local neighborhood features and global structural features.

[0008] As a further technical solution, the local neighborhood features are fused with the global structural features, including: Calculate attention weights for the local neighborhood features and the global structural features respectively; Based on the attention weights, the local neighborhood features and global structural features are weighted and summed.

[0009] As a further technical solution, the construction of the factory monitoring diagram structure includes: Physical equipment in the factory is defined as equipment nodes, continuous spatial regions are defined as region nodes, and sensors are defined as environmental nodes. Define the edges between device nodes based on the material flow, data flow, or physical adjacency between devices; Define the edges between environmental nodes and device nodes or area nodes based on the sensor's monitoring range and its subordinate relationship with the device or area.

[0010] As a further technical solution, triggering the warning based on the comparison result includes: Based on the specific node type and / or node location information where the risk score exceeds the threshold, different levels and different content of early warning notifications are triggered.

[0011] As a further technical solution, the method of triggering the early warning includes at least one of the following: visual alarm on the monitoring interface, triggering on-site audible and visual alarm device, and automatically pushing alarm message to the management terminal.

[0012] According to one aspect of the present invention, a system for intelligent detection and security alarm is provided, comprising: The data processing module is used to acquire real-time monitoring data of the factory and update the feature vectors of the nodes in the factory monitoring graph structure based on the data. The nodes in the factory monitoring graph structure represent equipment, areas or sensor entities within the factory, and the edges represent the physical or functional relationships between the entities. Each node has a feature vector corresponding to the type of entity it represents, and the feature vector contains feature data extracted from the real-time monitoring data to characterize the state of the entity. The analysis and early warning module is used to input the updated factory monitoring graph structure into a pre-trained graph convolutional neural network model, output the risk score value corresponding to each node in the graph, and compare the risk score value of each node with a preset threshold, triggering an early warning based on the comparison result. The graph convolutional neural network model used is a network containing multiple graph convolutional layers. It uses an attention mechanism to weighted aggregate the features of neighboring nodes for feature propagation, and residual connections are set between adjacent graph convolutional layers. Furthermore, the network is configured to fuse node features from different levels for multi-scale feature learning.

[0013] According to one aspect of the present invention, a graphical user interface for an intelligent detection and security alarm system is provided, the interface being displayed on a terminal device, the interface comprising: A factory monitoring diagram structure visualization area is used to display factory entities represented by nodes and edges and their relationships; wherein, the interface is configured to dynamically update the visual attributes of the corresponding nodes in the visualization area in response to changes in the risk score values ​​of each node obtained according to the method of claim 1; wherein, the visual attributes include at least one of the following: node color, icon, size or border style; The risk details display panel is used to display detailed status data and historical risk trends related to a node when a user's interactive operation on a node in the visualization area is detected.

[0014] According to one aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the method described thereon.

[0015] Compared with the prior art, the technical solution provided by the present invention brings the following significant beneficial effects: 1. Achieved deep, multi-source data fusion and correlation analysis: Through innovative factory monitoring graph structure modeling, heterogeneous data such as previously isolated video streams, equipment operation data, and environmental sensor readings are unified into a graph reflecting their physical and functional relationships for analysis. This method fundamentally breaks down data silos, enabling the system to identify complex systemic risks caused by the synergistic effects or chain reactions of multiple devices and environmental factors—something that traditional methods of independently analyzing individual sensor data cannot achieve.

[0016] 2. Significantly improved accuracy, early detection, and interpretability of risk identification: By integrating an attention mechanism, GCN can adaptively focus on key nodes and anomaly correlations. Combined with multi-scale feature learning, it can more accurately distinguish between normal fluctuations and real risks, effectively reducing false alarms and missed alarms. Through in-depth mining of global graph structure features, the system can capture early signs of risk propagation before local anomalies spread into global failures, achieving earlier warnings. By outputting node-level risk scores from the model, the system can directly locate specific risky devices or areas and display them intuitively through a GUI. This upgrades alarm information from "what happened" to "where, what device, and why there is a risk," greatly improving the understandability of alarms and the targeted nature of responses.

[0017] 3. Ensures high performance and stability of the algorithm model: The introduction of residual connections ensures that the improved GCN deep network can be effectively trained, avoiding performance degradation and enabling the model to learn and memorize more complex and hidden risk patterns. The entire system forms an efficient closed loop from data perception and intelligent analysis to automatic response, greatly improving the real-time performance and automation level of security monitoring.

[0018] 4. Provides intuitive and efficient human-computer interaction and decision support: The dedicated graphical user interface protected by this invention transforms abstract graphical data and algorithmic scoring into a dynamic, visual, and interactive monitoring dashboard. Administrators can not only see all the risk hotspots in the plant at a glance, but also quickly drill down into details through interaction. This transforms the traditional passive alarm receiving mode into a proactive safety management mode of active monitoring, intuitive analysis, and rapid decision-making, thereby comprehensively improving the plant's safety operation and maintenance efficiency and intelligent management level. Attached Figure Description

[0019] Figure 1 This is a flowchart of an intelligent detection and security alarm method provided in an embodiment of the present invention.

[0020] Figure 2 This is a structural block diagram of an intelligent detection and security alarm system provided in an embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention.

[0022] This invention uses a large-scale, modern prefabrication plant for building components as an example to illustrate the specific implementation of the technical solution of this invention. The plant includes automated production lines, large lifting equipment (such as tower cranes), a concrete mixing plant, curing kilns, raw material warehouses, and various environmental and safety sensors.

[0023] like Figure 2 As shown, this embodiment utilizes an intelligent detection and safety alarm system, which mainly includes a data processing module, an analysis and early warning module, and a user interface module. The system connects to various factory monitoring data sources via the factory network, including: high-definition network cameras, equipment PLC (programmable logic controller) systems, IoT sensors (temperature, humidity, smoke, noise, and wind speed sensors), and personnel positioning systems. The system outputs early warning information to management terminals, such as the central monitoring room screen, safety officer mobile terminals, and on-site audible and visual alarms.

[0024] like Figure 1 As shown, the intelligent detection and security alarm method includes the following steps:

[0025] Step 1: Acquire real-time monitoring data of the factory and update the feature vectors of nodes in the pre-constructed factory monitoring graph structure based on this data. Nodes in the factory monitoring graph structure represent equipment, areas, or sensor entities within the factory, and edges represent physical connections or functional associations between these entities. Each node has a feature vector corresponding to the type of entity it represents, which contains feature data extracted from the real-time monitoring data to characterize the entity's state.

[0026] Step 2 involves inputting the updated factory monitoring graph structure into a pre-trained graph convolutional neural network (GNN) model for processing to obtain the risk score value corresponding to each node in the graph. The GNN model used is a network containing multiple layers of graph convolutional layers. In this network, feature propagation is performed through graph convolutional layers that weight and aggregate neighbor node features using an attention mechanism, and residual connections are established between adjacent graph convolutional layers. Furthermore, the network is configured to fuse node features from different levels for multi-scale feature learning.

[0027] Step 3: Compare the risk score of each node with a preset threshold, and trigger an early warning based on the comparison result. The early warning can be triggered with different levels and content based on the specific node type and / or node location information where the risk score exceeds the threshold. This can be achieved through at least one of the following methods: visual alarm on the monitoring interface, triggering on-site audible and visual alarm devices, or automatically pushing alarm messages to the management terminal.

[0028] Prior to step 1, the method described in this embodiment of the invention further includes: constructing and updating in real time a factory monitoring graph structure G = (V, E) based on the factory's physical layout and operational logic. Here, V represents a set of nodes, each node representing a device or region, and E represents a set of edges, representing connections between devices. The node feature vector X represents the state information of each device or region.

[0029] The node (V) definition and feature vector initialization are as follows:

[0030] Equipment Nodes: For example, specific equipment such as Tower Crane No. 1, Concrete Mixer No. 2, and A-line Formwork Conveyor Belt are defined as equipment nodes. Each equipment node contains data about equipment performance, such as operating status, failure rate, and maintenance records. The initial feature vector of each equipment node can contain static attributes (such as equipment ID, model, and rated load) and dynamic attributes (such as current operating status, real-time power, and last maintenance time), and this data comes from the equipment PLC.

[0031] Area nodes: These represent different areas of the factory. For example, high-altitude work areas, heavy component storage yards, and curing kiln areas can be defined as area nodes. Area nodes can contain comprehensive information from all equipment and sensors within that area. For instance, the feature vector of an area node can integrate personnel density (from video analytics or positioning systems), average environmental data, and a summary of equipment status within the area.

[0032] Environmental nodes: These provide real-time data about the plant environment. For example, wind speed sensor S01 located in the storage yard, temperature and humidity sensor T01 in the curing kiln, and smoke detector F01 in the plant area are defined as environmental nodes. Their feature vectors are their real-time readings.

[0033] An edge exists between two devices if they are physically connected or directly related in a workflow. An edge exists between a device node and its adjacent environmental sensor node, representing the impact of environmental conditions on device performance. Edges between different regions represent the flow of materials, data, or personnel. The specific definition of an edge (E) is as follows:

[0034] Physical connection edge: For example, there is an edge between tower crane No. 1 and the heavy component storage yard, indicating that the tower crane is operating in this area; there is an edge between the A-line formwork conveyor belt and the curing kiln inlet, indicating the material flow direction.

[0035] Functional association edges: For example, there is an edge between a concrete mixer and a concrete pump truck, which represents a supply relationship in the production process; there are edges between the high-voltage distribution cabinet node and all the main electrical equipment nodes in the factory, which represent power dependence relationships.

[0036] Spatial / Subordinate Edge: For example, the wind speed sensor S01 node is connected to the heavy component yard area node, indicating that it monitors that area; the camera node in the tower crane cab is connected to the No. 1 tower crane equipment node.

[0037] According to the above definition, the process of constructing a graph involves determining all nodes and their attributes, as well as defining edges based on the actual factory layout and equipment relationships. Graphs constructed in this way not only reflect the actual layout and status of equipment in the factory, but also capture the complex interactions between equipment and between equipment and the environment, providing a rich information foundation for subsequent graph convolutional network analysis.

[0038] In building a monitoring system for a smart factory, a Graph Convolutional Neural Network (GCN) is used to analyze the complex interactions and environmental states within the factory. The following are the detailed steps for building and training a GCN:

[0039] (1) Selection of convolutional layers

[0040] Choose a spatial graph convolutional layer based on the characteristics of the factory data (such as the complexity of relationships between equipment). This is suitable for processing non-Euclidean data, such as factory layouts. Use the formula... Implement graph convolution, where It is an adjacency matrix with self-connections. It is a degree matrix. It is a layer The weight matrix, It is an activation function.

[0041] (2) Multi-scale feature learning

[0042] First, a multi-layered GCN is constructed, consisting of multiple layers within the graph convolutional network, each corresponding to a different graph convolutional operation. These layers are designed to capture graph structural features from local to global perspectives. Next, inter-layer relationships are defined between consecutive graph convolutional layers to allow information to be transferred between different layers.

[0043] Next, local feature extraction is performed. In the initial graph convolutional layer, a small neighborhood range is set to focus on the direct neighbors of each node. This helps capture the local features of the nodes. Local feature extraction is performed using the graph convolution formula:

[0044]

[0045] Further expanding the neighborhood, in higher-level graph convolutional layers, increases the range to include more distant nodes. This helps in understanding the graph's structure from a global perspective. This is achieved by adjusting the weight matrix in the graph convolutional layers. This enables the network to learn the global characteristics of the nodes.

[0046] The local and global features captured from different levels are then fused. This is achieved by introducing an attention mechanism. The attention mechanism dynamically adjusts the contribution of each feature by calculating its importance weight. These weights are adaptively learned based on the feature itself and task-related contextual information. For each feature vector, an attention score is calculated using the following formula:

[0047]

[0048] in Represents a node Attention score The weight vector representing the attention mechanism is a learned parameter. Represents a node The feature vectors are then used. These scores are then normalized using the softmax function to obtain the final attention weights.

[0049]

[0050] in Represents a node Attention weights Represents a node Attention score Represents a node The attention score is used to sum the denominators. This represents the exponential sum of the attention scores of all nodes, used for normalization. Then, the calculated attention weights are used to weight the features at different scales. The weighted features are represented as:

[0051]

[0052] in Represents a node The final feature vector, This represents the original feature vector of node i. The weighted features are then merged or concatenated to form a fused feature representation. This embodiment uses a weighted average to fuse the features.

[0053] To explain this more clearly, multi-scale feature learning and fusion can also be understood as:

[0054] Shallow graph convolutional layers (such as layers 1-2) have small receptive fields and mainly extract local neighborhood features of nodes, such as abnormal vibration of tower crane nodes and abnormal readings of wind speed sensors directly related to them.

[0055] Deep graph convolutional layers (such as layers 3-4) can aggregate a wider range of neighbor information through multiple information transmissions and extract global structural features of nodes. For example, a chain reaction risk pattern may occur where a concrete mixer stops due to an abnormality in a high-voltage distribution cabinet, which in turn causes the formwork conveyor belt to idle and wait.

[0056] The network is configured to fuse node features from layers 2 and 4. During fusion, an attention weight is calculated for each of the local features from layer 2 and the global features from layer 4, and then these are weighted and summed to obtain the final composite feature representation for each node. This attention-based fusion strategy allows the model to dynamically decide whether to rely more on local details or the global situation in the final decision.

[0057] (3) Residual connection

[0058] To address the vanishing gradient problem in deep network training and improve model training efficiency and performance, residual connections are added between consecutive graph convolutional layers.

[0059]

[0060] A residual connection is introduced between every two consecutive graph convolutional layers. That is, the output of the previous layer is added to the output of the current layer after graph convolution transformation, and then input into the activation function. This ensures that even if the network is deepened, the gradient can be effectively backpropagated, and the model performance remains stable.

[0061] (4) Network training and optimization

[0062] Prepare feature vectors containing information about factory equipment and environmental conditions. and the corresponding tags The model is trained using supervised learning. The labels Y are node-level or graph-level binary (normal / abnormal) or multi-class labels (e.g., normal, equipment failure, environmental exceedance, safety violation). These labels are derived from confirmed safety incident reports and corresponding timestamps in the factory's historical operation and maintenance records. Optimization is performed using the cross-entropy loss function.

[0063]

[0064] The model is trained using backpropagation and gradient descent, and network parameters are adjusted to minimize the loss function. Finally, the number of graph convolutional layers is adjusted based on the complexity of the factory environment and the network's performance. Furthermore, hyperparameters such as the learning rate and kernel size are adjusted according to actual training conditions to optimize the multi-scale feature learning effect.

[0065] The real-time monitoring and early warning process in step 3 is as follows:

[0066] Data Updates and Model Inference: During system operation, the data processing module continuously receives real-time data streams and updates the feature vectors of corresponding nodes in the graph structure. For example, if the video analysis subroutine identifies personnel not wearing safety helmets in the heavy component storage area, it updates the heavy component storage area node with this event as a feature. The updated entire graph structure is then fed into the pre-trained model of the analysis and early warning module in real time.

[0067] Risk score output: The model performs forward calculations on the input graph and outputs a real-time risk score value for each node in the graph. The score range is, for example, 0 to 1, with higher values ​​indicating greater risk.

[0068] Intelligent early warning trigger: The analysis and early warning module compares the risk score of each node with a preset threshold (e.g., 0.7). Once a node's score exceeds the threshold, the system immediately triggers an early warning. The early warning method varies depending on the node type and risk level.

[0069] For equipment nodes (such as tower crane No. 1 with a risk score of 0.85): the node icon will turn red and flash on the user interface (visual alarm); at the same time, a vibration alarm will be sent to the smart wristbands of the tower crane operator and the ground commander (a message will be pushed to the management terminal); if the risk score is extremely high (>0.9), automatic interlocking control can be used to limit the lifting speed of the tower crane.

[0070] For area nodes (such as high-altitude work area with a risk score of 0.78): the area is highlighted on the interface; the surrounding sound and light alarms are triggered to alert personnel in the area (triggering on-site sound and light alarms); and a message containing the specific area and risk type is pushed to the safety inspector.

[0071] For environmental nodes (such as smoke detector F01 with a risk score of 0.95): immediately trigger a plant-wide fire alarm and automatically push information to the fire duty room and key personnel.

[0072] Based on the inventive concept disclosed in the foregoing method embodiment using a prefabrication plant for building components as an example, the present invention also provides corresponding system, user interface, and storage medium embodiments. These embodiments together constitute a complete technical solution implementation system.

[0073] Based on the same inventive concept, an intelligent detection and security alarm system that implements the aforementioned method is constructed. For example... Figure 2 As shown, the system connects to data sources such as high-definition cameras, equipment PLCs, IoT sensors, and personnel positioning systems through the factory network, and outputs early warnings to management terminals such as the central monitoring screen, mobile terminals, and audible and visual alarms.

[0074] The system includes:

[0075] Data Processing Module: This module is configured to perform the graph structure construction and updating functions described in the aforementioned method embodiments. Specifically, it continuously receives real-time monitoring data from various data sources and dynamically maintains a factory monitoring graph structure G = (V, E) according to predefined rules (such as defining tower crane No. 1 as an equipment node and establishing physical connection edges between it and the heavy component storage area node). It is responsible for converting raw data (such as PLC status codes, sensor readings, and video analysis events) into values ​​in the feature vectors of corresponding nodes (equipment nodes, area nodes, and environment nodes), achieving unified data representation and fusion.

[0076] Analysis and Early Warning Module: This module serves as the core analysis engine of the system, loading and running a pre-trained improved Graph Convolutional Neural Network (GCN) model. This model possesses the structural features described previously, including an integrated attention mechanism, multi-scale feature learning and fusion, and residual connections. This module is configured to: receive real-time graph data from the data processing module; perform forward computation using the GCN model, outputting a real-time risk score for each node in the graph; compare these scores with preset thresholds; and generate differentiated early warning instructions based on the comparison results and node type / location information (such as high risk warnings for tower cranes, speed restrictions, personnel violations in storage yards, and on-site audible and visual alarms).

[0077] User Interface Module: This module is configured to generate and provide a human-computer interaction interface for visually presenting the system status and receiving user commands.

[0078] Based on the same inventive concept, the present invention also provides a dedicated graphical user interface for the intelligent detection and security alarm system, which is displayed on the terminal device of the monitoring center.

[0079] The core design concept of this interface is to transform abstract graphical data and risk scores into intuitive, dynamic, and interactive visual information. The interface mainly includes:

[0080] The factory monitoring diagram visualization area renders the factory monitoring diagram structure in real time as a topology graph. Different types of entities (such as tower cranes, mixers, stockpiles, and temperature sensors) are represented by different graphical nodes, and the relationships between entities are connected by edges. Its key feature is the dynamic mapping mechanism: the interface engine continuously receives risk score values ​​for each node from the analysis and early warning module, and updates the visual attributes of each corresponding node in the visualization area in real time and automatically, such as color, size, or icon status, according to preset visual coding rules (e.g., risk values ​​of 0-0.3 are mapped to green, and 0.7-1.0 to red). This allows administrators to have a clear understanding of the overall safety risk situation of the plant.

[0081] Risk Details Display Panel: This panel provides in-depth interactive diagnostic capabilities. When an administrator interacts with a high-risk node in the visualization area (such as the red tower crane icon #1) by clicking or touching, the panel is triggered and displays detailed information related to that node, including its real-time risk score, historical score change curves, current equipment parameters, associated warning logs, and possible handling suggestions, thus achieving a closed loop from macro-level situational awareness to micro-level root cause analysis.

[0082] Based on the same inventive concept, this invention also provides a computer-readable storage medium, such as a server hard drive, a solid-state memory in an industrial control computer, or a distributed cloud storage space. This medium stores a computer program (instructions). When this computer program is loaded and executed by one or more processors (e.g., a server CPU deployed in a factory data center or a processor in an edge computing device), it will drive the processor to fully implement the intelligent detection and safety alarm process described in the foregoing method embodiments. That is, the processor will perform the following operations: construct and dynamically update the factory monitoring graph structure according to predefined rules; call a pre-trained improved graph convolutional neural network model to process the graph data and generate node-level risk scores; execute threshold comparison and early warning logic; and optionally, drive the generation of the graphical user interface and perform dynamic visualization rendering.

[0083] This medium embodiment enables the technical solutions of the present invention to be copied, distributed, and deployed and run on various computing devices in the form of software products, thereby expanding the scope of protection and application of the present invention.

[0084] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, or improvements made by those skilled in the art without departing from the concept of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent detection and security alarm, characterized in that, The method is implemented based on a pre-trained graph convolutional neural network model, including: The system acquires real-time monitoring data of the factory and updates the feature vectors of nodes in the factory monitoring graph structure based on this data. The nodes in the factory monitoring graph structure represent equipment, areas, or sensor entities within the factory, and the edges represent physical or functional relationships between the entities. Each node has a feature vector corresponding to the type of entity it represents, and the feature vector contains feature data extracted from the real-time monitoring data to characterize the state of the entity. The updated factory monitoring graph structure is input into a pre-trained graph convolutional neural network model, which outputs the risk score value corresponding to each node in the graph. The graph convolutional neural network model used is a network containing multiple graph convolutional layers. Feature propagation is performed by graph convolutional layers that use an attention mechanism to weight and aggregate the features of neighboring nodes. Residual connections are set between adjacent graph convolutional layers. Furthermore, the network is configured to fuse node features from different levels for multi-scale feature learning. The risk score of each node is compared with a preset threshold, and an early warning is triggered based on the comparison result.

2. The method for intelligent detection and security alarm according to claim 1, characterized in that, The attention mechanism is implemented by calculating the attention weights between each node in the graph and its neighboring nodes, and then weighting and aggregating the features of the neighboring nodes based on the attention weights.

3. The method for intelligent detection and security alarm according to claim 1, characterized in that, The fusion of node features from different levels for multi-scale feature learning includes: extracting local neighborhood features of nodes output from shallow graph convolutional layers and global structural features of nodes output from deep graph convolutional layers, and fusing the local neighborhood features with the global structural features.

4. The method for intelligent detection and security alarm according to claim 3, characterized in that, The fusion of the local neighborhood features and the global structural features includes: Calculate attention weights for the local neighborhood features and the global structural features respectively; Based on the attention weights, the local neighborhood features and global structural features are weighted and summed.

5. The method for intelligent detection and security alarm according to claim 1, characterized in that, The construction of the factory monitoring diagram structure includes: Physical equipment in the factory is defined as equipment nodes, continuous spatial regions are defined as region nodes, and sensors are defined as environmental nodes. Define the edges between device nodes based on the material flow, data flow, or physical adjacency between devices; Define the edges between environmental nodes and device nodes or area nodes based on the sensor's monitoring range and its subordinate relationship with the device or area.

6. The method for intelligent detection and security alarm according to claim 1, characterized in that, The method of triggering an early warning based on the comparison result includes: Based on the specific node type and / or node location information where the risk score exceeds the threshold, different levels and different content of early warning notifications are triggered.

7. The method for intelligent detection and security alarm according to claim 1, characterized in that, The methods for triggering early warnings include at least one of the following: visual alarm on the monitoring interface, triggering on-site audible and visual alarm devices, and automatically pushing alarm messages to the management terminal.

8. A system for intelligent detection and security alarm, characterized in that, include: The data processing module is used to acquire real-time monitoring data of the factory and update the feature vectors of the nodes in the factory monitoring graph structure based on the data. The nodes in the factory monitoring graph structure represent equipment, areas or sensor entities within the factory, and the edges represent the physical or functional relationships between the entities. Each node has a feature vector corresponding to the type of entity it represents, and the feature vector contains feature data extracted from the real-time monitoring data to characterize the state of the entity. The analysis and early warning module is used to input the updated factory monitoring graph structure into a pre-trained graph convolutional neural network model, output the risk score value corresponding to each node in the graph, and compare the risk score value of each node with a preset threshold, triggering an early warning based on the comparison result. The graph convolutional neural network model used is a network containing multiple graph convolutional layers. It uses an attention mechanism to weighted aggregate the features of neighboring nodes for feature propagation, and residual connections are set between adjacent graph convolutional layers. Furthermore, the network is configured to fuse node features from different levels for multi-scale feature learning.

9. A graphical user interface for an intelligent detection and security alarm system, the interface being displayed on a terminal device, characterized in that, The interface includes: A factory monitoring diagram structure visualization area is used to display factory entities represented by nodes and edges and their relationships; wherein, the interface is configured to dynamically update the visual attributes of the corresponding nodes in the visualization area in response to changes in the risk score values ​​of each node obtained according to the method of claim 1; wherein, the visual attributes include at least one of the following: node color, icon, size or border style; The risk details display panel is used to display detailed status data and historical risk trends related to a node when a user's interactive operation on a node in the visualization area is detected.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 7.

Images