Industrial tourism safety management system and method for automobile industrial park

Abstract

The invention provides an industrial tourism safety management system and method for an automobile industrial park, and the method comprises the steps: collecting the number of tourists, the flow density and the average speed in real time at a visiting path segment, collecting the temperature, pressure, rotating speed, dust, carbon dioxide, hydrogen and other combustible gas concentrations of equipment, combining the tourist and equipment data with the distance from the path segment to the equipment, and obtaining the distance of the path segment to the equipment; and constructing a heterogeneous space-time diagram, inputting the heterogeneous space-time diagram into a diagram neural network model, outputting a recommended route and risk indexes of each path section, pushing the recommended route and the risk indexes to a tourist end by a platform layer to push guidance in an augmented reality form, and when the risk indexes reach a grading threshold value, automatically triggering a production line to suddenly stop and starting fire-fighting spraying by a system. According to the scheme, sensing, prediction, decision making and response closed loop are realized, the path pushing accuracy can be greatly improved, the accident rate is greatly reduced, the evacuation efficiency is improved, and the method is suitable for automobile manufacturing and other industrial tourism scenes.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent recommendation, in particular to a system and method for industrial tourism safety management in an automobile industrial park. BACKGROUND

[0002] In recent years, "industrial tourism" has gradually become a new hotspot of cultural and leisure consumption. Many automobile manufacturing enterprises have begun to set up visiting channels in production parks and open to the public based on the need for brand display and popular science education. Unlike traditional scenic spots, automobile industrial parks are filled with stamping machines, welding robots, spraying workshops, power battery warehouses and other types of equipment. The operating scene is often accompanied by multiple hazards such as high temperature, high pressure, dust, flammable and explosive gases, and high-speed moving mechanical arms. At the same time, logistics vehicles and automated guided vehicles (AGV) frequently shuttle, and the overall safety risk is significantly higher than that of ordinary tourist sites.

[0003] At present, the safety management of the park mainly relies on two measures: one is to restrict the tourist activity range through ground signs, guardrails and manual explanation; the other is to use closed-circuit television and ordinary environmental sensors for passive monitoring. This mode has the following shortcomings: first, the route planning is statically fixed and it is difficult to dynamically adjust according to real-time passenger flow and equipment state, and congestion or personnel breakthrough isolation zone may occur during peak periods; second, video monitoring alone cannot prospectively assess the risk of high-risk gases, dust explosions or robot collisions, and often only reacts passively after an accident occurs; third, safety information is scattered among production, security, fire protection and other systems, lacking a unified data fusion and linkage response mechanism, resulting in delayed emergency decision-making; fourth, traditional radio or text prompts are not intuitive enough for tourists, and the evacuation efficiency is low in the event of an emergency.

[0004] With the development of Internet of Things, digital twin and artificial intelligence technologies, multi-modal sensing, graph neural network prediction and augmented reality navigation have been gradually applied to smart factories. However, in the "industrial tourism" scenario, these technologies have not yet formed a system solution that takes tourist safety as the core and covers the "sensing-prediction-decision-response-audit" closed loop. There is a lack of targeted computing for the coupling information of path segments and dangerous equipment spaces, and there is also a lack of automatic hierarchical disposal mechanism across departments and systems, resulting in the inability to provide fine and personalized safety management experience while ensuring production rhythm. Moreover, in the existing video image face recognition and dangerous intelligent automatic judgment, the spatial distance relationship between the crowd and the equipment is not considered as a factor, and the existing algorithm does not consider the distance-related dangerous factors as an activation function or network model path recommendation factor, resulting in low path recommendation accuracy and low risk judgment accuracy.

[0005] Therefore, it is urgent to propose an automobile industrial park industrial tourism safety management system and method fusing multi-source sensing guidance to realize real-time risk prediction of tourist flow and hazard source, dynamic path optimization and rapid linkage emergency, and fundamentally improve the safety guarantee level of industrial tourism. SUMMARY

[0006] In view of the above problems mentioned in the prior art, the present application proposes an automobile industrial park industrial tourism safety management system and method, which combines the data of tourists and equipment with the distance from the path segment to the equipment by real-time collection of the number of tourists, the density of people flow, the average speed, the age structure, and the collection equipment temperature, pressure, speed, dust, carbon dioxide and hydrogen gas concentration, constructs a heterogeneous space-time graph and inputs a distance gate graph neural network, outputs a recommended route and a risk index of each path segment, and pushes the platform layer to the tourist end in the form of augmented reality to enhance the guidance, and when the risk index reaches the classification threshold, the system automatically triggers the production line to stop and start the fire sprinkler. This scheme realizes the closed loop of perception, prediction, decision and response, greatly reduces the accident rate and improves the evacuation efficiency, and is suitable for automobile manufacturing and other industrial tourism scenes.

[0007] The present application provides an automobile industrial park industrial tourism safety management method, comprising the steps of:

[0008] S1: target detection and face recognition are performed on real-time video images by cameras deployed in each path segment, the number of tourists, the density of people flow, the average speed of travel and the age distribution of the path segment are obtained, and a tourist feature vector is generated based on the center point coordinates of the path segment;

[0009] S2: the environmental parameters of each industrial equipment E j running in the automobile industrial park are obtained, the environmental parameters include temperature, pressure, speed, dust concentration, flammable gas concentration and carbon dioxide concentration, and the spatial coordinates of the equipment and the hazard level coefficient H j are recorded to generate a device hazard source feature vector;

[0010] S3: the Euclidean distance d i，j between each path segment and each running industrial equipment is calculated, and the hazard factor D i of the i-th path segment is calculated according to the Euclidean distance:

[0011] D i =min(H j / (d i，j +ε))

[0012] Wherein, min() represents the minimum value of each device, i is the path segment index, j is the running industrial equipment index; ε represents the adjustment parameter value, which is a positive number greater than 0 and less than 1;

[0013] S4: fuse the tourist feature vector, the equipment hazard source feature vector and the hazard factors of each path segment to construct a park state space graph, input the state space graph into the trained graph neural network model, and the graph neural network model outputs the recommended visiting route and risk index of each path segment;

[0014] S5: according to the visiting route and the risk index, the platform layer pushes the AR path guide to the tourist APP, and automatically triggers the industrial equipment production line emergency stop, starts the fire sprinkling and real-time alarm when detecting that the risk index exceeds the threshold.

[0015] Preferably, the hidden layer of the graph neural network model adopts an improved activation function f(x i , D i ):

[0016] f(x, D i ) = sigma ((1 + tanh(mu * D i )) * x)

[0017] Wherein, x i is the feature vector of the previous network layer output of the i-th path segment, D i is the hazard factor of the i-th path segment, sigma() represents the ReLU nonlinear operation, tanh() represents the hyperbolic tangent function, and mu is a learnable positive parameter.

[0018] Preferably, the crowd density is obtained by dividing the number of tourists entering the path segment per unit time by the effective area of the path segment; generating the tourist feature vector includes reading the center point coordinates of the current path segment in the three-dimensional Cartesian coordinate system of the park, and sequentially splicing the center point coordinates with the corresponding number of tourists, crowd density, average speed and age distribution information of the path segment to generate a tourist feature vector for subsequent graph neural network processing.

[0019] Preferably, the age distribution includes first detecting the face in the real-time video image, then calling a pre-trained deep learning age estimation model to give an age prediction value for each face; all prediction values are classified according to pre-set age groups, at least including a juvenile group of fourteen years old and below, an adult group of fifteen to sixty years old, and a senior group of sixty years old and above, the number of tourists in each age group is counted, and the proportion of each age group is calculated based on the total number of tourists to form the age distribution of the path segment.

[0020] Preferably, the combustible gas concentration includes the concentration of hydrogen, methane, propane, butane, acetylene, toluene, xylene; the device hazard source feature vector generation comprises: normalizing the temperature, pressure, speed, dust concentration, carbon dioxide concentration and combustible gas concentration data collected in real time for each industrial device, and splicing the device spatial coordinates, device type identification and preset hazard level coefficient together to form a device hazard source feature vector for subsequent graph neural network processing.

[0021] Preferably, the step S4 of performing feature fusion on the tourist feature vector, the device hazard source feature vector and the path segment risk factor to construct a park state space graph comprises:

[0022] S41: taking each path segment as a first type of node, splicing the corresponding tourist feature vector and the path segment risk factor according to a fixed field order to form a complete path segment node attribute;

[0023] S42: taking each industrial device as a second type of node, directly using the device hazard source feature vector as a device node attribute; spatial topology connection, between the path segment nodes, according to the geometric adjacency relationship in the park digital twin model, establishing adjacent walking channel edges to describe the continuity of the walkable path for tourists;

[0024] S43: between the path segment node and the device node, when the spatial distance is lower than a preset threshold, establishing a radiation influence edge, and recording the distance, device hazard level and other attributes on the edge to represent the potential risk effect of the device on the path segment;

[0025] S44: the above two types of nodes and two types of edges jointly form a heterogeneous state space graph; by introducing a time stamp or a sliding window method, the continuous frame graph structure is stacked to form a space-time graph sequence, providing dynamic evolution information for the graph neural network.

[0026] Preferably, the graph neural network model adopts the following hierarchical structure:

[0027] The input embedding layer respectively performs linear mapping on the path segment node attribute and the industrial device node attribute, so that the feature dimensions of different node types are consistent;

[0028] The space-time graph attention layer performs multi-head attention aggregation on the heterogeneous state space graph at each time step, respectively learns weights according to node-edge types, and captures the walkable relationship between path segments and the risk radiation relationship of devices to path segments;

[0029] The distance gate activation layer uses a distance gate activation function containing a path segment risk factor after each hidden layer to adaptively amplify or suppress the features of each node;

[0030] The time sequence convolution layer performs one-dimensional convolution on the historical feature sequence of the same node along the time dimension to model the dynamic change trend of the risk;

[0031] The global readout layer performs average pooling and maximum pooling splicing on the node features output by the plurality of spatio-temporal graph layers to generate a comprehensive representation for each path segment;

[0032] The multi-task output head sets the risk index regression branch and the optimal route classification branch in parallel, and simultaneously optimizes them through a joint loss function during the training stage.

[0033] The application also provides an automobile industrial park industrial tourism safety management system, comprising:

[0034] The tourist multi-modal feature acquisition module acquires the number of tourists, the crowd density, the average speed of travel and the age distribution of the path segment through target detection and face recognition on real-time video images by deploying cameras at each path segment, and generates a tourist feature vector based on the center point coordinates of the path segment;

[0035] The industrial equipment hazard source acquisition module acquires the environmental parameters of each industrial equipment E j running in the automobile industrial park, the environmental parameters including temperature, pressure, speed, dust concentration, flammable gas concentration and carbon dioxide concentration, and records the device space coordinates and the hazard level coefficient H j , and generates a device hazard source feature vector;

[0036] The distance-enhanced feature calculation module calculates the Euclidean distance d i，j between each path segment and each running industrial equipment, and calculates the hazard factor D i of the i th path segment according to the Euclidean distance:

[0037] D i =min(H j / (d i，j +ε))

[0038] Wherein, min() represents the minimum value solved for each device, i is the path segment index, j is the running industrial equipment index; ε represents an adjustment parameter value, which is a positive number greater than 0 and less than 1;

[0039] The graph neural network model prediction module fuses the tourist feature vector, the device hazard source feature vector and the hazard factor of each path segment, constructs a park state space graph, inputs the state space graph into the trained graph neural network model, and the graph neural network model outputs the recommended visiting route and the risk index of each path segment;

[0040] A response module, the platform layer pushes an AR path guide to the tourist APP according to the visiting route and the risk index, and automatically triggers the industrial equipment production line emergency stop, starts the fire sprinkling and real-time alarm when detecting that the risk index exceeds the threshold.

[0041] Preferably, the hidden layer of the graph neural network model adopts an improved activation function f(x i , D i ):

[0042] f(x, D i ) = sigma ((1 + tanh(mu * D i )) * x)

[0043] Wherein, x i is the feature vector of the previous network layer output i th path segment, D i is the dangerous factor of the i th path segment, sigma () represents ReLU nonlinear operation, tanh () represents hyperbolic tangent function, and mu is a learnable positive parameter.

[0044] Preferably, the crowd density is obtained by dividing the number of tourists entering the path segment per unit time by the effective area of the path segment; generating a tourist feature vector includes reading the center point coordinates of the current path segment in the three-dimensional Cartesian coordinate system of the park, and splicing the center point coordinates with the corresponding number of tourists, crowd density, average speed and age distribution information of the path segment in turn, to generate a tourist for subsequent graph neural network processing.

[0045] The application provides an automobile industry park industrial tourism safety management system and method, which can realize the following beneficial technical effects:

[0046] 1. The application unifies the tourist feature vector, the equipment dangerous source feature vector and the distance-weighted dangerous factor of each path segment according to the heterogeneous node-edge norm, constructs a park state space graph, realizes the deep integration of the three-dimensional elements of "people-road-machine", and compared with the existing safety management scheme which only depends on video or a single sensor, the application can simultaneously capture the changes of passenger flow, equipment working conditions and spatial coupling relationship, so that the risk distribution presents a centimeter-level spatial resolution and a second-level time resolution, and the predictability is significantly improved.

[0047] 2. The graph neural network hidden layer of the application adopts an improved activation function that fuses the dangerous factors of the path segment, the dangerous factors of the path segment are added to the calculation process of the activation function as consideration factors, the distance between the equipment and the crowd path segment and the dangerous coefficient factors are considered, the judgment accuracy and precision are greatly improved, and the path recommendation and risk degree judgment accuracy are improved.

[0048] 3.The system can trigger sound-light alarm at the edge side in seconds, and through the industrial control interface to link the production line emergency stop and fire sprinkler; AR guidance pushes the evacuation path to the tourist terminal at the same time. Field verification shows that the potential accident response time is shortened from 90 seconds in the traditional scheme to 35 seconds, avoiding the risk of stopping production of equipment worth millions of yuan. BRIEF DESCRIPTION OF DRAWINGS

[0049] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can be obtained without creative labor on the basis of these drawings.

[0050] Figure 1 is a step flow chart of a safety management method of an automobile industrial park industrial tourism of the present application;

[0051] Figure 2 is a schematic diagram of a safety management system of an automobile industrial park industrial tourism of the present application. DETAILED DESCRIPTION

[0052] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0053] Embodiment 1:

[0054] In view of the above problems mentioned in the prior art, in order to solve the above technical problems, as shown in the accompanying drawings: Figure 1 The present application provides a safety management method of an automobile industrial park industrial tourism, comprising the steps of:

[0055] S1: Real-time video images are detected and face recognition is performed by cameras deployed on each path segment to obtain the number of tourists, the flow density, the average speed of travel and the age distribution of the path segment, and a tourist feature vector is generated based on the path segment center point coordinates;

[0056] In one embodiment, the path segment P12 in the welding workshop visit channel is selected to illustrate how to obtain visitor data from real-time video streams and generate corresponding feature vectors. A 4K network camera (25 fps, 120° field of view) is installed 3 m above the path segment P12, and an edge computing box (NVIDIA Jetson AGX Orin) is connected. The center point coordinates of the path segment are recorded as (24.6 m, 15.2 m, 0 m) in the digital twin platform. The edge box runs the YOLOv9-Pose model to detect human boxes and skeletal key points, and generates unique pedestrian IDs by combining DeepSORT cross-frame tracking. The number of active IDs in the last 5-second window is counted to obtain the number of visitors N = 18. The pedestrian density is calculated by measuring the actual available area of the path segment 50 m 2 using a depth camera binocular ranging, and the pedestrian density p = 18 ÷ 50 ≈ 0.36 person / m2 is calculated. The average speed of travel is measured by taking the displacement of each pedestrian ID in 5 seconds and dividing by the time difference, and the average is 1.2 m / s. The age distribution is estimated by using the ArcFace feature + lightweight age regression network to infer the age value for the detected clear face; classification: 3 children (<14 years old), 13 adults (15-60 years old), 2 old people (>60 years old). The age distribution ratio is: 0.17 / 0.72 / 0.11 (young / adult / old).

[0057] The visitor feature vector is spliced in the preset order: path segment center point x, y, z coordinates, number of visitors N, pedestrian density p, average speed of travel, and three age ratios. Therefore, the visitor feature vector of the path segment P12 is: [24.6, 15.2, 0, 18, 0.36, 1.2, 0.17, 0.72, 0.11]. The data synchronization feature vector is updated every second, sent to the park data bus via MQTT, and written to the attribute field of the path segment node for subsequent spatio-temporal graph neural network inference. Through the above process, second-level, fine-grained passenger flow perception can be achieved on the edge side, providing accurate input for downstream risk prediction and dynamic route planning.

[0058] S2: Obtain the environmental parameters of each industrial equipment E j currently running in the automobile industry park, including temperature, pressure, speed, dust concentration, flammable gas concentration, and carbon dioxide concentration, and record the device space coordinates and hazard level coefficient H j, and generates a device hazard source feature vector; in one embodiment, a high-power spot welding robot E21 in the welding workshop is selected to illustrate how to collect environmental parameters and generate a device hazard source feature vector. Device and coordinate information Device type: six-axis spot welding robot, model IRB-6700; its base geometric center coordinates are recorded as (42.3m, 8.4m, 0.0m) in the digital twin platform; according to the enterprise "Safety Risk Classification and Control Standard", the hazard level coefficient H = 8 (10-level system) is evaluated by considering factors such as welding flow, arc heat, and surrounding flammable gas.

[0059] Table 1: Sensor layout and parameter collection

[0060]

[0061]

[0062] Data preprocessing Linear normalization is performed on continuous quantities such as temperature, pressure, and speed according to the device specification range; dust and gas concentrations are normalized according to the upper limit of "GB50073-2023 Industrial Enterprise Health Standard"; if there is short-time sampling loss, sliding window mean interpolation is used. Feature vector splicing is written in the order of the specified fields: device coordinates x, y, z; device type identifier (such as the enumeration index of "WeldRobot"); normalized temperature, pressure, speed, dust concentration; normalized hydrogen, methane, propane, and carbon dioxide concentration; hazard level coefficient H.

[0063] Splicing result: [42.3, 8.4, 0.0, 4, 0.79, 0.65, 0.60, 0.23, 0.26, 0.12, 0.31, 8], where "4" represents the enumeration value of the device type, and the subsequent decimals are the normalized parameters. Data reporting and caching, feature vectors are packaged as JSON and pushed to the industrial data lake through the OPC UA gateway every 1 second; at the same time, the attributes of the E21 node in the welding workshop state diagram are written, providing real-time reference for the graph neural network. Through this process, multi-parameter sensing and feature extraction of high-risk equipment can be completed at the millisecond-second level, ensuring that the downstream risk prediction model obtains accurate and consistent device hazard source information.

[0064] S3: Calculate the Euclidean distance d between each path segment and each running industrial equipment i，j , and calculate the hazard factor D of the ith path segment according to the Euclidean distance i :

[0065] D i = min(H j / (d i，j + ε))

[0066] Wherein, min() represents the minimum value solved for each device, i is the path segment index, j is the index of the industrial equipment being run; ε represents an adjustment parameter value, which is a positive number greater than 0 and less than 1;

[0067] S4: The visitor feature vector, the equipment hazard source feature vector, and the hazard factors of each path segment are fused, a park state space graph is constructed, the state space graph is input into the trained graph neural network model, and the graph neural network model outputs the recommended visiting route and risk index of each path segment; in some embodiments, based on the aforementioned path segment P12 (the visitor feature vector has been generated) and the spot welding robot E21 (the equipment hazard source feature vector has been generated), and additionally selecting the adjacent path segments P11 and P13 and another robot E22 in the same workshop, the overall process and model working mechanism are demonstrated. Feature fusion and node assembly take P11, P12, and P13 as the first type of node, respectively mounting their own visitor feature vectors and path segment hazard factors. E21 and E22 are taken as the second type of node, mounting the corresponding equipment hazard source feature vectors. Each node records a unified 3D coordinate, which is convenient for subsequent distance calculation and visualization. The generation and labeling of edges, walking connected edges: the digital twin model shows that P11-P12 and P12-P13 are directly adjacent channels, and the system establishes undirected edges between the two pairs of path segment nodes, and records the channel width, floor, and other attributes. Risk radiation edges: P12 is 7m away from E21 (lower than the threshold of 10m), generating a directed edge E21→P12; P13 is 6m away from E22, also generating E22→P13; the edge attributes include the real distance and the equipment hazard level coefficient. Thus, a small heterogeneous graph containing 5 nodes and 4 edges is obtained.

[0068] Table 2: Graph neural network model structure

[0069]

[0070] The model is pre-trained offline with 30-day historical data, and is fine-tuned once a night online. The data infusion system writes the latest node features into the graph at a frequency of 1Hz, and fills the recent 6-second spatio-temporal sequence. Message passing is in the first round of attention aggregation, and the high temperature and hydrogen concentration information of E21 is transmitted to P12 through the risk radiation edge; the walking connected edge between path segments transmits the crowd density gradient, helping the model to evaluate the congestion trend. Distance gating activation: since the risk factor D 12= 0.84, higher than 0.32 and 0.46 of P11 and P13, the node feature is amplified to about 1.6 times of the original, and the subsequent convolution pays more attention to P12. Timing convolution: the model detects that the P12-E21 edge maintains a high attention weight for 4 seconds in a row, judging that the high risk persists. Output: risk index P11: 0.28, P12: 0.81, P13: 0.35 (0-1 normalization). It is recommended to visit the route "P11→P13" (skip P12).

[0071] Platform layer action AR guide: the tourist terminal displays a yellow arrow in real time to guide across P12. The risk index of P12 is higher than the second threshold 0.75, the platform layer immediately sends an emergency stop instruction to the welding line PLC and starts the top spray, and reports the event to the safety operation center. After the emergency stop lasts for 90 seconds, the equipment temperature decreases, the model risk index falls to 0.38, and the production is resumed after the platform layer confirms. It can be seen from the embodiment that the application forms a closed loop in the aspects of node-level feature fusion, space-time relationship modeling, gated risk reinforcement, and real-time decision-linkage disposal, which significantly improves the safety management accuracy and timeliness of the industrial tourism scene.

[0072] S5: The platform layer pushes AR path guidance to the visitor's APP according to the tour route and risk index, and automatically triggers the industrial equipment production line to stop, starts the fire sprinkler, and reports in real time when the risk index exceeds the threshold. In some embodiments, the foregoing graph neural network inference result is continued to explain how the platform layer pushes AR guidance to the visitors in high-risk scenarios and automatically controls the production line, fire protection, and alarm system. Time point: 10:17:35 am; Location: welding workshop path segment P12. Determination result: the risk index R(P12) returned by the graph neural network is 0.81, which has exceeded the second threshold T2 set by the park, which is 0.75; the recommended tour route is adjusted to "P11→P13". The route optimization submodule receives the "P11→P13" route and the risk index of each path segment. The threshold determination submodule detects that R(P12)≥T2, generates a high-risk event EV-ZX412, and records the event timestamp. The AR rendering submodule calls the real-time coordinates in the Unity digital twin scene to generate a yellow flashing arrow pointing to the P11 exit; a red warning area covers the P12 ground. This rendering result is pushed to the visitor's mobile APP and AR glasses in glTF stream format. The APP side pops up a prompt within 200ms: "High-temperature welding operation in front, please detour according to the arrow guidance". The industrial control interface submodule calls the OPC UA method StopLine(id="WeldLine_3"), and issues it to the welding line PLC; after receiving it, the PLC successively disconnects the main motor and the direct current welding source within 500ms, and the total stop line time is controlled within 800ms. At the same time, the fire gateway is called to start the top spraying device; the valve is fully open with a delay of 1.2s. The audible and visual alarm controller is triggered by the Modbus TCP instruction, and the workshop warning light is accompanied by an 80dB buzzer sounding for 15s. The safety operation interface submodule packages the risk index curve of the event EV-ZX412, the control instruction receipt, and the visitor evacuation record to generate a JSON event package. The event package is uploaded to the enterprise SOC via HTTPS, and is simultaneously written into the municipal emergency management platform MQ channel, realizing cross-department linkage. The event hash is also chained to ensure that the post-audit is tamper-proof. The state monitoring and recovery platform layer refreshes the risk index of P12 at a frequency of 1Hz. When the index is below T1=0.45 for 30s continuously, the threshold determination submodule generates a release instruction; the industrial control interface calls the stepwise recovery production; the fire gateway turns off the sprinkler; the audible and visual alarm is automatically extinguished, and the APP prompts "danger removed, please continue the tour according to the green arrow". Effect verification evacuation efficiency: the visitors detour through P11-P13 within 35s, maintaining a distance of at least 12m from the high-risk area; production loss: stop time 2min, welding line loss ≤80 pieces of production capacity, which is lower than the threshold set by the safety department; accident avoidance real-time temperature measurement shows that the instantaneous spark temperature peak in the P12 area is 750℃, and if the stop and spraying are not taken, the surrounding plastic wire slot is at risk of melting.The embodiment shows that the platform layer of the application can complete risk judgment, AR navigation, industrial emergency stop and cross-department reporting in a high-risk scene in seconds, and realizes closed-loop safety management from prediction to disposal.

[0073] Preferably, the hidden layer of the graph neural network model adopts an improved activation function f(x i , D i ):

[0074] f(x, D i ) = σ((1 + tanh(mu * D i )) * x)

[0075] wherein x i is a feature vector of an i-th path segment output by a previous network layer, D i is a risk factor of the i-th path segment, sigma() represents a ReLU nonlinear operation, tanh() represents a hyperbolic tangent function, and mu is a learnable positive parameter.

[0076] Preferably, the crowd density is obtained by dividing the number of tourists entering the path segment per unit time by the effective area of the path segment; and generating the tourist feature vector includes reading the center point coordinates of the current path segment in a three-dimensional Cartesian coordinate system of the park, and concatenating the center point coordinates with the number of tourists, the crowd density, the average travel speed and the age distribution information corresponding to the path segment in sequence to generate a tourist feature vector for subsequent graph neural network processing.

[0077] Preferably, the age distribution includes detecting human faces in real-time video images first, and then calling a pre-trained deep learning age estimation model to give an age prediction value for each human face; all prediction values are classified according to preset age ranges, at least including a juvenile range of fourteen years old and below, an adult range of fifteen to sixty years old, and a senior range of sixty years old and above, the number of tourists in each age range is counted, and the proportion of each age range is calculated based on the total number of tourists to form the age distribution of the path segment.

[0078] Preferably, the combustible gas concentration includes the concentrations of hydrogen, methane, propane, butane, acetylene, toluene and xylene; and the generation of the equipment hazard source feature vector includes normalizing the real-time collected temperature, pressure, speed, dust concentration, carbon dioxide concentration and combustible gas concentration data of each industrial equipment, and concatenating the equipment spatial coordinates, equipment type identifier and preset hazard level coefficient to form an equipment hazard source feature vector for subsequent graph neural network processing.

[0079] Preferably, the step S4 of performing feature fusion on the tourist feature vector, the equipment hazard source feature vector and the risk factors of each path segment to construct a park state space graph specifically includes:

[0080] S41: taking each path segment as a first type of node, splicing the corresponding tourist feature vector and the path segment risk factor according to a fixed field order to form a complete path segment node attribute;

[0081] S42: taking each industrial equipment as a second type of node, directly using the equipment risk source feature vector as the equipment node attribute; spatial topology connection, between the path segment nodes, according to the geometric adjacency relationship in the park digital twin model, establishing adjacent walking channel edges to describe the continuity of the walkable tourists;

[0082] S43: between the path segment node and the equipment node, when the spatial distance is lower than a preset threshold, establishing a radiation influence edge, and recording the distance, equipment risk level and other attributes on the edge to represent the potential risk effect of the equipment on the path segment;

[0083] S44: the above two types of nodes and two types of edges jointly form a heterogeneous state space graph; by introducing a time stamp or a sliding window method, the continuous frame graph structure is stacked to form a space-time graph sequence, providing dynamic evolution information for the graph neural network.

[0084] Preferably, the graph neural network model adopts the following hierarchical structure:

[0085] The input embedding layer linearly maps the path segment node attribute and the industrial equipment node attribute respectively, so that the feature dimensions of different node types are consistent;

[0086] The space-time graph attention layer performs multi-head attention aggregation on the heterogeneous state space graph at each time step, respectively learns weights according to node-edge types, and captures the walkable relationship between the path segments and the risk radiation relationship of the equipment to the path segments;

[0087] The distance gate activation layer uses a distance gate activation function containing a path segment risk factor after each hidden layer to adaptively amplify or suppress the features of each node;

[0088] The time sequence convolution layer performs one-dimensional convolution on the historical feature sequence of the same node along the time dimension to model the dynamic change trend of the risk;

[0089] The global readout layer performs average pooling and maximum pooling splicing on the node features output by the multiple space-time graph layers to generate a comprehensive representation for each path segment;

[0090] The multi-task output head sets the risk index regression branch and the optimal route classification branch in parallel, and simultaneously optimizes them through a joint loss function in the training stage.

[0091] The application also provides an automobile industrial park industrial tourism safety management system, which comprises the above-mentioned graph neural network model. Figure 2As shown, the perception layer includes: 4K panoramic camera (hood + PoE): installed above each path segment for passenger flow detection and face recognition. Millimeter wave 77GHz radar: deployed at corners and elevated positions to detect personnel and AGV movement under obstructions. All-in-one environmental sensor: integrates temperature, pressure, PM2.5, VOC, CO2, H2, CH4, etc. gas probes, and is magnetically fixed to the device guardrail. Industrial robot state collector: embeds an EtherCAT module in the welding / spraying robot control cabinet to read speed, current, and temperature rise in real time. UWB positioning tag (optional): worn on the visitor's safety helmet for centimeter-level positioning and evacuation tracking.

[0092] The edge computing layer includes: Jetson AGX Orin edge box: one for every 3-5 path segments, responsible for video encoding, YOLOv9-Pose, and image-radar fusion inference. Industrial IoT gateway: supports OPCUA, Modbus-TCP, EtherNet / IP protocols, and converts device PLC data into MQTT streams.

[0093] Network and security isolation includes gigabit PoE switches: connect cameras and edge boxes for power and data on one line. Wi-Fi 6 & 5G small base stations: provide wireless backhaul for visitor AR glasses, mobile phones, and UWB readers. Industrial firewall: divides the production control network into an OT area, with edge boxes and IoT gateways in the DMZ; only allows white-listed protocols to access horizontally.

[0094] The platform layer includes private cloud GPU servers (2x A100): deploy K8s clusters to run spatio-temporal graph neural networks and federated PPO path optimization strategies. High-availability Kafka-Flink stream processing cluster: implements a millisecond-level event bus to push risk indexes to the SOC and AR rendering service. AR rendering server: generates glTF navigation instructions in real time and multicasts them to mobile devices.

[0095] Control and linkage includes redundant PLC cabinet (Safety I / O): receives platform emergency stop commands to control the welding line, conveyor belt, and overhead valve. Acoustic alarm (MODBUS TCP / IP): installed on the workshop passageway and above the equipment, starts the buzzer and red flash after receiving the gateway command. Fire pump frequency conversion cabinet: connected to the PLC safety relay through hard-wired contacts, ensuring manual / automatic switching in case of communication failure.

[0096] Interconnection overview, perception layer -> edge box: Gigabit PoE (camera), CAN / RS-485 (environmental probe), Radar CANFD; edge box -> IoT gateway: local Ethernet, using MQTT over TLS; IoT gateway -> private cloud: 10 GbE fiber uplink, data into Kafka topic; private cloud -> AR terminal: rendering server pushes low-latency stream via WebSocket + WebRTC, distributed via Wi-Fi6 / 5G; private cloud -> PLC: write emergency stop position to PLC via OPC UA PubSub in DMZ, PLC triggers safety relay via hardwiring; Hard-wired redundancy, ensuring electrical cascading action; Bidirectional WebSocket, uploading UWB positioning and visitor confirmation information.

[0097] Through the above hardware and connection architecture, the system forms a closed loop of "end-side multi-modal perception - edge real-time inference - cloud intelligent decision - industrial cascading control", realizing high-precision risk prediction, second-level emergency disposal and visitor immersive navigation.

[0098] The visitor multi-modal feature acquisition module acquires the number of visitors, the crowd density, the average speed of travel and the age distribution of the path segment through target detection and face recognition on real-time video images by deploying cameras on each path segment, and generates a visitor feature vector based on the path segment center point coordinates;

[0099] The industrial equipment hazard source acquisition module acquires the environmental parameters of each industrial equipment Ej currently running in the automobile industry park, including temperature, pressure, speed, dust concentration, flammable gas concentration and carbon dioxide concentration, and records the device spatial coordinates and hazard level coefficient H j , and generates a device hazard source feature vector;

[0100] The distance-enhanced feature calculation module calculates the Euclidean distance d i，j between each path segment and each industrial equipment currently running, and calculates the hazard factor D i of the i-th path segment according to the Euclidean distance:

[0101] D i = min(H j / (d i，j + ε))

[0102] Where min() represents the minimum value for each device, i is the path segment index, and j is the index of the industrial equipment currently running; ε represents the adjustment parameter value, which is a positive number greater than 0 and less than 1;

[0103] The graph neural network model prediction module fuses the tourist feature vector, the equipment hazard source feature vector and the hazard factors of each path segment, constructs a park state space graph, inputs the state space graph into the trained graph neural network model, and the graph neural network model outputs the recommended visiting route and risk index of each path segment.

[0104] The response module pushes the AR path guide to the tourist APP according to the visiting route and risk index, and automatically triggers the industrial equipment production line emergency stop, starts the fire sprinkler and reports in real time when the risk index exceeds the threshold.

[0105] Preferably, the hidden layer of the graph neural network model adopts an improved activation function f(x i , D i ):

[0106] f(x, D i ) = sigma((1 + tanh(mu * D i )) * x)

[0107] Wherein, x i is the feature vector of the previous network layer output of the i-th path segment, D i is the hazard factor of the i-th path segment, sigma() represents ReLU nonlinear operation, tanh() represents hyperbolic tangent function, and mu is a learnable positive parameter.

[0108] Preferably, the crowd density is obtained by dividing the number of tourists entering the path segment per unit time by the effective area of the path segment; generating a tourist feature vector includes reading the center point coordinates of the current path segment in the three-dimensional Cartesian coordinate system of the park, and sequentially splicing the center point coordinates with the corresponding number of tourists, crowd density, average speed and age distribution information of the path segment to generate a tourist for subsequent graph neural network processing.

[0109] The present application provides an automobile industry park industrial tourism safety management system and method, which can realize the following beneficial technical effects:

[0110] 1. The present application uniformly splices the tourist feature vector, the equipment hazard source feature vector and the distance-weighted hazard factor of each path segment according to the heterogeneous node-edge norm, constructs a park state space graph, realizes the deep fusion of the three-dimensional elements of "people-road-machine", and compared with the existing safety management scheme which only depends on video or single sensor, the present application can simultaneously capture the changes of passenger flow, equipment working condition and space coupling relationship, so that the risk distribution presents a centimeter-level spatial resolution and a second-level time resolution, and the predictability is significantly improved.

[0111] 2. The improved activation function of the path segment danger factor is adopted in the hidden layer of the graph neural network, the path segment danger factor is added to the calculation process of the activation function as a consideration factor, the distance between the equipment and the crowd path segment and the danger coefficient factor are considered, the judgment accuracy and precision are greatly improved, and the path recommendation and danger degree judgment accuracy are improved.

[0112] 3. The system can trigger sound and light alarm at the edge side in seconds, and through the industrial control interface, the production line emergency stop and fire sprinkler are linked; the AR guide pushes the evacuation path to the tourist terminal. Field verification shows that the potential accident response time is shortened from 90 seconds of the traditional scheme to 35 seconds, avoiding the risk of stopping production of equipment worth millions of yuan.

[0113] The above describes in detail a kind of automobile industrial park industrial tourism safety management system and method, the principle and implementation mode of the present application are described in this paper by applying specific examples; the above example is only used to help understand the core idea of the present application; meanwhile, for those skilled in the art, according to the idea and method of the present application, the specific implementation mode and application range will be changed; in view of the above, the content of the specification should not be understood as a limitation of the present application.

Claims

1. A method for safety management of industrial tourism in an automotive industrial park, characterized in that, Including the following steps: S 1: By using cameras deployed on each path segment to perform target detection and face recognition on real-time video images, the number of tourists, crowd density, average walking speed and age distribution of the path segment are obtained, and tourist feature vectors are generated based on the coordinates of the center point of the path segment. S2: Obtain E data for every industrial piece of equipment currently in operation within the automotive industrial park. j The environmental parameters include temperature, pressure, rotational speed, dust concentration, combustible gas concentration, and carbon dioxide concentration, and the equipment's spatial coordinates and hazard level coefficient H are also recorded. j Generate the feature vector of equipment hazard sources; S3: Calculate the Euclidean distance d between each path segment and each operating industrial device. i，j Calculate the hazard factor D of the i-th path segment based on Euclidean distance. i : D i =min(H j / (d i，j +ε)) Where min() represents the minimum value calculated for each device, i is the path segment index, j is the index of the industrial device currently in operation; ε represents the adjustment parameter value, which is a positive number greater than 0 and less than 1; S4: The feature vectors of tourists, the feature vectors of equipment hazards, and the hazard factors of each path segment are fused to construct a state space diagram of the park. The state space diagram is then input into a trained graph neural network model, which outputs suggested tour routes and risk indices for each path segment. S5: The platform layer pushes AR path guidance to the visitor's APP based on the visitor route and risk index, and automatically triggers emergency stop of industrial equipment production line, starts fire sprinkler system and alarms in real time when the risk index exceeds the threshold.

2. The method for safety management of industrial tourism in an automotive industrial park as described in claim 1, characterized in that, The hidden layer of the graph neural network model uses an improved activation function f(x). i D i ): f(x,D i )=σ((1+tanh(μ*D i ))*x) Where, x i D outputs the feature vector of the i-th path segment from the previous network layer. i Let σ be the risk factor for the i-th path segment, σ() represent the ReLU nonlinear operation, tanh() represent the hyperbolic tangent function, and μ be a learnable positive parameter.

3. The method for safety management of industrial tourism in an automotive industrial park as described in claim 1, characterized in that, The pedestrian density is obtained by dividing the number of tourists entering the path segment per unit time by the effective area of ​​the path segment; generating the tourist feature vector includes reading the coordinates of the center point of the current path segment in the park's three-dimensional Cartesian coordinate system, and sequentially concatenating the center point coordinates with the number of tourists, pedestrian density, average walking speed, and age distribution information corresponding to the path segment to generate a tourist feature vector for subsequent graph neural network processing.

4. The method for safety management of industrial tourism in an automotive industrial park as described in claim 1, characterized in that, The age distribution involves first detecting faces in real-time video images, then calling a pre-trained deep learning age estimation model to provide an age prediction value for each face; classifying all prediction values ​​into preset age groups, including at least a children's group of 14 years and under, an adult group of 15 to 60 years, and an elderly group of 60 years and above; counting the number of tourists in each age group; and calculating the proportion of each age group based on the total number of tourists to form the age distribution of the path segment.

5. The method for safety management of industrial tourism in an automotive industrial park as described in claim 1, characterized in that, The combustible gas concentration includes the concentrations of hydrogen, methane, propane, butane, acetylene, toluene, and xylene; the generation of equipment hazard source feature vectors includes: normalizing the temperature, pressure, rotation speed, dust concentration, carbon dioxide concentration, and combustible gas concentration data collected in real time for each industrial equipment, and concatenating the equipment spatial coordinates, equipment type identifier, and preset hazard level coefficients to form an equipment hazard source feature vector for subsequent graph neural network processing.

6. The method for safety management of industrial tourism in an automotive industrial park as described in claim 1, characterized in that, Step S4 involves fusing the visitor feature vector, equipment hazard source feature vector, and hazard factors of each path segment to construct a park state space map. Specifically, this includes: S41: Take each path segment as the first type of node, and concatenate the corresponding tourist feature vector with the risk factor of the path segment according to a fixed field order to form the complete path segment node attributes; S42: Using each piece of industrial equipment as a second type of node, the feature vector of the equipment hazard source is directly used as the attribute of the equipment node; spatial topology connection, between path segment nodes, adjacent walking passage edges are established according to the geometric adjacency relationship in the digital twin model of the park to describe the continuity of walkable paths for tourists. S43: When the spatial distance between a path segment node and a device node is less than a preset threshold, a radiation influence edge is established, and attributes such as distance and device hazard level are recorded on the edge to characterize the potential risk effect of the device on the path segment. S44: Combine the two types of nodes and the two types of edges mentioned above to form a heterogeneous state space graph; stack the continuous frame graph structure by introducing timestamps or sliding windows to form a spatiotemporal graph sequence, providing dynamic evolution information for graph neural networks.

7. The method for safety management of industrial tourism in an automotive industrial park as described in claim 1, characterized in that, The graph neural network model adopts the following hierarchical structure: The input embedding layer performs linear mapping on the attributes of path segment nodes and industrial equipment nodes respectively, so that the feature dimensions of different node types are consistent. The space-time graph attention layer performs multi-head attention aggregation on the heterogeneous state space graph at each time step, and learns weights according to node-edge type to capture the walkable relationship between path segments and the risk radiation relationship of the device to the path segment; After each hidden layer, the distance-gated activation layer uses a distance-gated activation function that includes path segment hazard factors to adaptively amplify or suppress the features of each node. Temporal convolutional layers perform one-dimensional convolution on the historical feature sequences of the same node along the time dimension to model the dynamic changing trend of risk; The global readout layer performs average pooling and max pooling on the node features output from multiple spatiotemporal layers to concatenate them and generate a comprehensive representation for each path segment. The multi-task output head sets up the risk index regression branch and the optimal route classification branch in parallel, and optimizes them simultaneously through a joint loss function during the training phase.

8. A safety management system for industrial tourism in an automotive industrial park, characterized in that, include: The multimodal feature acquisition module for tourists uses cameras deployed on each path segment to perform target detection and face recognition on real-time video images, obtains the number of tourists, crowd density, average walking speed and age distribution of the path segment, and generates tourist feature vectors based on the coordinates of the center point of the path segment. The industrial equipment hazard source acquisition module acquires E data from every piece of industrial equipment currently operating in the automotive industrial park. j The environmental parameters include temperature, pressure, rotational speed, dust concentration, combustible gas concentration, and carbon dioxide concentration, and the equipment's spatial coordinates and hazard level coefficient H are also recorded. j Generate the feature vector of equipment hazard sources; The distance enhancement feature calculation module calculates the Euclidean distance d between each path segment and each operating industrial device. i，j Calculate the hazard factor D of the i-th path segment based on Euclidean distance. i : D i =min(H j / (d i，j +ε)) Where min() represents the minimum value calculated for each device, i is the path segment index, j is the index of the running industrial device; ε represents the adjustment parameter value, which is a positive number greater than 0 and less than 1; The graph neural network model prediction module fuses the features of visitor feature vectors, equipment hazard source feature vectors, and hazard factors of each path segment to construct a park state space graph. The state space graph is then input into the trained graph neural network model, which outputs suggested tour routes and risk indices for each path segment. The response module, at the platform layer, pushes AR path guidance to the visitor's APP based on the visitor route and risk index. When the risk index exceeds the threshold, it automatically triggers an emergency stop of the industrial equipment production line, starts the fire sprinkler system, and issues a real-time alarm.

9. The industrial tourism safety management system for an automotive industrial park as described in claim 8, characterized in that, The hidden layer of the graph neural network model uses an improved activation function f(x). i D i ): f(x,D i )=σ((1+tanh(μ*D i ))*x) Where, x i D outputs the feature vector of the i-th path segment from the previous network layer. i Let σ be the risk factor for the i-th path segment, σ() represent the ReLU nonlinear operation, tanh() represent the hyperbolic tangent function, and μ be a learnable positive parameter.

10. The industrial tourism safety management system for an automotive industrial park as described in claim 8, characterized in that, The pedestrian density is obtained by dividing the number of tourists entering the path segment per unit time by the effective area of ​​the path segment; generating the tourist feature vector includes reading the coordinates of the center point of the current path segment in the park's three-dimensional Cartesian coordinate system, and sequentially concatenating the center point coordinates with the number of tourists, pedestrian density, average walking speed, and age distribution information corresponding to the path segment to generate a tourist feature vector for subsequent graph neural network processing.

Images