Digital factory multi-mode disaster protection system based on dynamic weighted neural network and self-optimization method

Abstract

The invention discloses a digital factory multi-mode disaster protection system based on a dynamic weighted neural network and a self-optimization method. The self-optimization method comprises the following steps: S1, collecting multivariate data; s2, processing real-time data and constructing a feature tensor; s3, obtaining a hazard value through a hazard value calculation model, and matching the hazard value with a hazard level; s4, automatically starting countermeasures according to the hazard level; s5, inputting the accident data into an accident database, and updating the weight in the hazard value calculation model according to a weight self-optimization mechanism; the system comprises a multivariate data acquisition module, a feature fusion module, a dynamic hazard assessment module, an automatic response system module, a weight optimization module and a system integration and communication module, wherein the automatic response system module is used for automatically starting countermeasures according to hazard levels. According to the method, an algorithm-data-execution closed-loop architecture is formed, and the problems of response lag and assessment stiffness of a traditional factory safety system are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of digital factory safety protection, and in particular to a digital factory multi-modal disaster protection system based on a dynamic weighted neural network and a self-optimization method. BACKGROUND

[0002] With the deepening of industrial 4.0, the safety protection system of the digital factory is transforming from traditional manual supervision to intelligent active prevention and control.

[0003] The existing technology mainly realizes risk early warning through two types of schemes, namely, a visual monitoring system and a sensor alarm system. The visual monitoring system can only detect static risks and has a low recognition rate for dynamic behavior chains. The sensor alarm system can only make single-dimensional judgments, resulting in a relatively high false alarm rate and the inability to associate personnel location information. There are still many pain points in the entire industry: data island problem, visual, sensing, and equipment systems operating independently will lead to risk fragmentation and high corresponding delay. Static evaluation models are rigid, permissions are fixed, and cannot adapt to changes in the scene, ignore the impact of protective measures on risks, and require production to be stopped for retraining during model iteration. The traditional system only realizes a one-way process of "monitoring-alarm", leading to repeated occurrence of similar accidents, continuous existence of false alarm rules, and lack of a closed-loop optimization mechanism. Therefore, there is an urgent need for a digital factory multi-modal disaster protection system based on a dynamic weighted neural network and a self-optimization method to solve the above problems. SUMMARY

[0004] The purpose of the present application is to solve the problems existing in the prior art and to propose a digital factory multi-modal disaster protection system based on a dynamic weighted neural network and a self-optimization method. According to the data obtained by the camera and the sensor, a dynamic hazard evaluation algorithm system is used to construct a heterogeneous feature tensor using multi-modal spatio-temporal fusion technology, and a hazard value is calculated using a dynamic weighted neural network architecture and a feature weight dynamic updating mechanism, thereby quantifying the disaster value and quickly responding according to the value to reduce losses.

[0005] In order to achieve the above-mentioned purpose, the present application adopts the following technical solutions: A self-optimization method for digital factory multi-modal disaster protection based on a dynamic weighted neural network, comprising the following specific steps: S1: Multivariate data acquisition: real-time collection of digital factory worker, environment and equipment information; S2: Real-time feature fusion: processing real-time data and constructing a feature tensor; S3: Dynamic hazard evaluation engine: obtaining a hazard value and matching a hazard level through a hazard value calculation model; S4: Automatic response system: automatically starting response measures according to the hazard level; S5: Weight optimizer: accident data is entered into the accident database, and the weights in the hazard value calculation model are updated according to the weight self-optimization mechanism.

[0006] As a further technical solution of the present application, in S1, the collected multi-element data includes digital factory itself data and data generated by digital factory cameras, sensors and other devices, wherein the digital factory itself data includes but is not limited to personnel information and equipment information, and the data generated by digital factory cameras, sensors and other devices includes but is not limited to camera video stream, temperature, humidity, vibration, gas and other data.

[0007] As a further technical solution of the present application, S2 specifically comprises: obtaining a feature tensor by multi-source data fusion of the data obtained in step S1; S21: integrating the data collected in step S1 into three categories: visual features, sensor time sequence features and metadata, wherein the visual features are obtained by extracting spatial-motion features through a 3D convolution network according to the visual related data obtained in step S1, such as camera video stream; the sensor time sequence features are obtained by processing sensor time sequence dependency relationship according to the sensor data in step S1; and the metadata is digital factory itself data; S22: the formula for constructing the feature tensor is as follows: , Wherein: T is the feature tensor; V is the camera video stream (sampling frequency >= 30fps); S t is the sensor matrix (temperature / vibration / gas, t time data); M is the equipment or personnel metadata (work type, machine model, etc.); Concat is to connect the three types of data to construct the feature tensor; CNN is to extract spatial-motion features through a 3D convolution network; LSTM is to process sensor time sequence dependency relationship; and Embed is to directly embed the equipment or personnel metadata.

[0008] As a further technical solution of the present application, S21 specifically comprises: S211: visual feature extraction: taking the camera video stream as an example, it is a three-dimensional data structure, which can be represented as Wherein represents the video frame image data at the time, which is input into a 3D convolution network, and the size of the 3D convolution kernel is , wherein is the height and width of the convolution kernel in the spatial dimension, and the depth in the time dimension, and the convolution operation can be represented as: , Wherein: is the feature map after convolution; 、 、 respectively represent input video frame image data row index, column index in spatial dimension, and index in time dimension, specifically, corresponding to the height direction of the image, corresponding to the width direction of the image, corresponding to the time sequence direction of the video, used to locate the starting position of the convolution operation in the three-dimensional data structure; respectively are the dimensions of the convolution kernel in spatial and time dimensions; 、 、 respectively represent the row index, column index of the convolution kernel in the spatial dimension (height and width direction), and the index in the time dimension, which are used to traverse each element of the convolution kernel in the corresponding dimension, and perform element-wise multiplication operation with the elements in the corresponding position of the input data; is the bias term; After nonlinear transformation such as activation function (ReLU), visual features containing spatial-motion features can be obtained; S212: Sensor time sequence feature processing: assuming that the data collected by the sensor is a one-dimensional time sequence wherein represents the sensor measurement value at the time point, in order to process the time sequence dependency relationship, a long short-term memory network (LSTM) is adopted, and the memory cell state update formula of the LSTM is as follows: , , , , , , wherein: respectively are the outputs of the forget gate, the input gate, and the output gate; is the candidate memory cell state; is the updated memory cell state; is the hidden state; respectively are the weight matrix and the bias term; is a sigmoid activation function; represents element-level multiplication; is a hyperbolic tangent function; the final obtained hidden state sequence , that is, the sensor time sequence features containing time dependence after processing.

[0009] As a further technical solution of the present application, the S3 specifically comprises: S31: input the feature tensor and other parameters obtained by step S2 into a hazard value calculation model to obtain a hazard value, and the formula of the hazard value calculation model is as follows: , Wherein: H is the hazard value; is a Sigmoid normalization function; f i is the deep feature extractor (CNN+Attention) of the i-th feature channel; T is the feature tensor; is a dynamic feature weight; is a negative feedback adjustment coefficient; ReLU is an activation function; SD is a protective measure feature vector; W d is the weight of the feature vector SD; The hazard value calculation model increases the negative feedback adjustment coefficient, and for behaviors or factors that have positive effects on disasters, negative effects are given in the hazard value calculation to reduce the hazard value, so that the obtained hazard value is more accurate; S32: match the hazard level according to the hazard value obtained by the hazard value calculation model, and output the hazard level to an automatic response system.

[0010] As a further technical solution of the present application, the S32 specifically comprises: A series of hazard level threshold values are set , satisfying , comparing the calculated hazard value with these threshold values to match the hazard level: If , it is matched as hazard level 1, indicating a normal state; If , it is matched as hazard level 2, indicating a mild hazard; If , it is matched as hazard level , indicating an extremely serious hazard; The hazard level threshold values are used to divide the calculated hazard value into different hazard levels, and the setting of these threshold values needs to be determined according to a large amount of historical data and actual application scene hazard conditions through statistical analysis and expert experience judgment.

[0011] As a further technical solution of the present application, the S4 specifically comprises: The automatic response system triggers the automatic response measures corresponding to the hazard level according to the hazard level obtained in step S3, and the automatic response measures include but are not limited to: device linkage (such as automatically triggering factory fire extinguishing devices, automatically closing valves), personnel warning (such as reminding personnel in the disaster area to avoid through telephone, SMS, AR glasses real-time warning), escape guidance (such as intelligent emergency lights, AR glasses generating optimal evacuation path, guiding personnel to escape).

[0012] S31: Device linkage: Fire extinguishing device triggering: when the hazard level reaches the fire extinguishing threshold , the system automatically triggers the factory fire extinguishing device, and the triggering signal of the fire extinguishing device is represented as: , wherein 1 represents triggering the fire extinguishing device, 0 represents not triggering, and G is the current hazard level; Valve closing: when the hazard level reaches the valve closing threshold , the system automatically closes the valve, and the valve closing signal is represented as: ; S32: Personnel warning: Telephone / SMS reminder: when the hazard level reaches the warning threshold , the system sends a telephone or SMS reminder to personnel in the disaster area, and the warning information sending signal is represented as: ; at the same time, the system needs to obtain personnel information in the disaster area, such as a set of personnel position coordinates , so as to accurately send the warning information; AR glasses real-time warning: when (the AR glasses warning threshold), the AR glasses real-time warning signal is triggered, and its expression is similar to the above; S33: Escape guidance: Intelligent emergency light guidance: the system controls the intelligent emergency light to generate an escape guidance signal according to the personnel position and the disaster situation; assuming that the emergency light has multiple states, such as arrow indications pointing in different directions, the guidance state of the emergency light can be calculated according to the positional relationship between the personnel and the safe exit, and a simple vector pointing or a guidance direction determined based on a path planning algorithm is adopted; AR glasses generate optimal evacuation path: the optimal path from the current position of the personnel to the safe exit is calculated through a path planning algorithm (such as Dijkstra algorithm or A* algorithm); assuming that the current position of the personnel is , the position of the safe exit is , and the factory plan is represented as a graph structure , wherein is a node (such as each position point), Edges (representing the connectivity relationship between locations, and the edge weight can be distance or predicted travel time, etc.); use Dijkstra algorithm to calculate the shortest path The core idea is to maintain a distance array (used to store the shortest distance from the starting node to each node, which is maintained and updated in the Dijkstra algorithm) and a priority queue, and each time the node with the smallest distance is taken out for relaxation operation, and finally the optimal path information obtained is displayed to the personnel through the AR glasses.

[0013] As a further technical solution of the present application, S5 specifically includes: entering the data after the disaster into the accident database, automatically adjusting the dynamic characteristic weight according to the difference between the historical accident data and the model prediction result through the weight self-optimization mechanism, so as to improve the accuracy and adaptability of the model, and the formula of the weight dynamic updating rule is as follows: , Wherein: KL is the KL divergence penalty term (the true risk distribution vs the predicted risk distribution ); is the true risk distribution; is the predicted risk distribution; is the learning rate, which controls the step size of weight update, and the initial value is usually small (such as 0.01) to ensure the stability of the update process; is the experience distribution correction coefficient; is the loss function (predicted hazard value vs actual accident loss), which measures the difference between the model prediction result and the actual accident data, and the commonly used mean square error (MSE) is: , wherein: is the hazard value predicted by the model; is the hazard value when the accident occurs; is the number of training samples.

[0014] As a further technical solution of the present application, in S5, the data after the disaster includes various information at the time of the accident, including but not limited to: accident occurrence time (time stamp accurate to seconds, used to record the specific time of the accident), accident location (specific coordinate position in the factory, such as workshop number and equipment number, etc.), personnel involved (personnel information present at the time of the accident, including personnel ID and their role in the accident), equipment state (running state of the equipment at the time of the accident, such as temperature, pressure, etc. Key parameters), environmental conditions (environmental factors at the time of the accident, such as temperature, humidity, etc.), hazard value (hazard value at the time of the accident calculated by step S3), hazard level (corresponding to the hazard value, used to quickly evaluate the severity of the accident), these data constitute a complete portrait of the accident, providing a basis for subsequent analysis.

[0015] A digital factory multi-modal disaster protection system based on a dynamic weighted neural network is used to realize a self-optimization method for digital factory multi-modal disaster protection based on a dynamic weighted neural network, comprising the following modules: Multivariate data acquisition module: for real-time collection of digital factory worker, environment and equipment information; Feature fusion module: for processing real-time data and constructing feature tensor; Dynamic hazard assessment module: for obtaining hazard value and matching hazard level through hazard value calculation model; Automatic response system module: for automatically starting response measures according to hazard level; Weight optimization module: for entering accident data into accident database, and updating weights in hazard value calculation model according to weight self-optimization mechanism; System integration and communication module: for real-time monitoring and management of the running state of each module, coordinating task scheduling and resource allocation between modules; and for data transmission and communication between modules.

[0016] The beneficial effects of the present application are: 1. Dynamic weight mechanism: compared with the fixed weight of the traditional method, Automatic adjustment following scene changes, automatic adjustment following historical accident database data, making the calculation result more accurate.

[0017] 2. Dual-mode feature fusion: visual and sensor data are fused through timestamps, solving the data island problem and making the calculation result applicable to more scenarios.

[0018] 3. Negative feedback reinforcement: The coefficient realizes negative feedback adjustment of the hazard value, quantifies the positive effect in the disaster, and makes more accurate judgments for disaster protection in complex situations. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 A flowchart of a self-optimization method for digital factory multi-modal disaster protection based on a dynamic weighted neural network is proposed. DETAILED DESCRIPTION

[0020] In order to make the technical means, creative features, purposes and effects realized by the present application easy to understand, the present application is further described below in conjunction with specific embodiments.

[0021] Please refer to the accompanying Figure 1 A self-optimization method for digital factory multi-modal disaster protection based on a dynamic weighted neural network comprises the following specific steps: S1: Multivariate data collection: Collecting digital factory worker, environment and equipment information in real time; The collected multivariate data includes digital factory itself data and data generated by digital factory cameras, sensors and other equipment, wherein the digital factory itself data includes but is not limited to personnel information, equipment information, and the data generated by the digital factory cameras, sensors and other equipment includes but is not limited to camera video stream, temperature, humidity, vibration, gas and other data.

[0022] S2: Real-time feature fusion: processing real-time data and constructing a feature tensor; The data obtained in step S1 is subjected to multi-source data fusion to obtain a feature tensor; S21: The data collected in step S1 is integrated into three categories: visual features, sensor time series features, and metadata, wherein the visual features are obtained by extracting spatial-motion features through a 3D convolutional network according to the visual-related data obtained in step S1, such as camera video stream; the sensor time series features are obtained by processing sensor time series dependencies according to the sensor data in step S1; and the metadata is the digital factory itself data; S211: Visual feature extraction: Taking the camera video stream as an example, it is a three-dimensional data structure, which can be represented as , wherein represents the video frame image data at the i-th moment, which is input into a 3D convolutional network, and the size of the 3D convolution kernel is , wherein is the height and width of the convolution kernel in the spatial dimension, is the depth in the time dimension, and the convolution operation can be represented as: , wherein: is the feature map after convolution; , , respectively represent the row index, column index and index in the time dimension of the input video frame image data in the spatial dimension, specifically, corresponds to the height direction of the image, corresponds to the width direction of the image, and corresponds to the time sequence direction of the video, which is used to locate the starting position of the convolution operation in the three-dimensional data structure; and are respectively the sizes of the convolution kernel in the spatial and time dimensions; , , These represent the row index, column index, and time index of the convolution kernel in the spatial dimensions (height and width), respectively. They are used to traverse each element of the convolution kernel in the corresponding dimension and perform element-wise multiplication with the corresponding element of the input data. For bias terms; After nonlinear transformations such as activation functions (e.g., ReLU), visual features containing spatial-motion characteristics can be obtained. S212: Sensor Temporal Feature Processing: Assuming the data acquired by the sensor is a one-dimensional time series ,in Indicates the first To handle time-series dependencies, sensor measurements at each time point are used, and a Long Short-Term Memory (LSTM) network is employed. The LSTM memory cell state update formula is as follows: , , , , , , in: These are the outputs of the forget gate, input gate, and output gate, respectively. Candidate memory cell state; This is the updated memory cell state; It is in a hidden state; These are the weight matrix and the bias term, respectively. It is the sigmoid activation function; Represents element-wise multiplication; It is the hyperbolic tangent function; The final hidden state sequence That is, the processed sensor time-series features that contain time-dependent relationships; S22: The formula for constructing the feature tensor is as follows: , Where: T is the feature tensor; V is the camera video stream (sampling frequency ≥ 30fps); S t M represents the sensor matrix (temperature / vibration / gas, etc., data at time t); M represents the equipment or personnel metadata (job type, machine model, etc.); Concat is used to connect the three types of data to construct a feature tensor; CNN is a 3D convolutional network to extract spatial-motion features; LSTM is used to process the time-series dependencies of the sensors; and Embed is used to directly embed the equipment or personnel metadata.

[0023] S3: Dynamic hazard assessment engine: obtain hazard value and match hazard level through hazard value calculation model; S31: input the feature tensor and other parameters obtained in step S2 into the hazard value calculation model to obtain the hazard value, and the formula of the hazard value calculation model is as follows: , wherein H is the hazard value; is a Sigmoid normalization function; f i is the deep feature extractor (CNN+Attention) of the i-th feature channel; T is the feature tensor; is the dynamic feature weight; is the negative feedback adjustment coefficient; ReLU is the activation function; SD is the protective measure feature vector; W d is the weight of the feature vector SD; The hazard value calculation model adds a negative feedback adjustment coefficient, which gives a negative impact on behaviors or factors that have a positive effect on disasters in the hazard value calculation, reduces the hazard value, and makes the obtained hazard value more accurate; S32: match the corresponding hazard level according to the hazard value obtained by the hazard value calculation model, and output the hazard level to the automatic response system; A series of hazard level thresholds are set to match the hazard level by comparing the calculated hazard value with these thresholds: If , it is matched as hazard level 1, indicating normal state; If , it is matched as hazard level 2, indicating mild hazard; If , it is matched as hazard level , indicating extremely serious hazard; The hazard level threshold is used to divide the calculated hazard value into different hazard levels. In actual application, the setting of these thresholds needs to be analyzed in detail and determined by expert experience according to the specific production process of the factory, the characteristics of the equipment, the danger of the stored materials, and historical accident data, etc., to ensure that the risk degree can be accurately reflected and provide basis for effective emergency response; at the same time, with the change of the production conditions of the factory and the accumulation of new accident data, these thresholds also need to be evaluated and updated regularly to ensure their rationality and effectiveness.

[0024] The following is a specific value example of the hazard level threshold : Chemical plant fire risk scenario: a. Collect historical data: Collect data of fire accidents occurred in the chemical plant in the past 10 years, including key parameters such as temperature, smoke concentration, flammable gas concentration, etc. when the fire occurred; assume that there are 50 fire accidents, record the temperature of the reaction kettle, the smoke concentration of the workshop, the leakage amount of flammable gas and the corresponding fire hazard degree (such as property loss amount, personnel casualty, etc.) in each accident; b. Statistical analysis of data: statistical analysis is performed on these data to calculate the distribution of fire hazard degree under different temperatures, smoke concentrations and flammable gas concentrations; It is found that when the temperature of the reaction kettle is lower than , generally no fire will occur; when the temperature is between , there are a small number of fire accidents, and most of them are small fires that can be controlled at the initial stage; when the temperature is between , the number of fire accidents increases significantly, and the fire is difficult to control; when the temperature exceeds , major fires often occur, causing serious losses; Similarly, for smoke concentration, when it is lower than , there is basically no fire risk, , there is slight smoke accumulation but not necessarily fire, , the possibility of fire increases, and when it exceeds , it is a serious fire scene with thick smoke; When the flammable gas concentration is lower than (volumetric ratio), it is relatively safe, , there is an explosion risk but not yet at a very high risk level, and when it exceeds , it is in an extremely dangerous explosive fire environment; c. Set threshold: according to the above statistical analysis results, combined with the evaluation experience of experts on chemical fire risk, set the hazard level threshold: (low threshold): reaction kettle temperature , smoke concentration , flammable gas concentration ; when the hazard value (calculated according to temperature, smoke concentration, flammable gas concentration, etc.) is lower than , it is determined as low hazard level, and the plant is in a relatively safe state, but still needs normal on-duty monitoring; (medium threshold): reaction kettle temperature , smoke concentration , flammable gas concentration ; when the hazard value is between - , it is determined as medium hazard level, the inspection frequency needs to be increased, the on-site personnel need to be reminded to pay attention to safety and be prepared for emergencies; (high risk threshold): reactor temperature , smoke concentration , flammable gas concentration (assuming that the degree of harm rises sharply after ); when the hazard value exceeds , it is determined that the high hazard level is high, and the emergency plan is immediately activated, triggering a series of emergency response measures such as automatic fire extinguishing, ventilation and smoke exhaust, personnel evacuation, etc.

[0025] S4: automatic response system: automatically start response measures according to hazard level; The automatic response system triggers the automatic response measures corresponding to the hazard level according to the hazard level obtained in step S3, and the automatic response measures include but are not limited to: device linkage (such as automatically triggering factory fire extinguishing devices, automatically closing valves), personnel warning (such as reminding personnel in the disaster area to avoid through telephone, SMS, AR glasses real-time warning), escape guidance (such as intelligent emergency lights, AR glasses generating optimal evacuation path, guiding personnel to escape).

[0026] S41: device linkage: fire extinguishing device trigger: when the hazard level reaches the fire extinguishing threshold , the system automatically triggers the factory fire extinguishing device, and the trigger signal of the fire extinguishing device is represented as: , where 1 represents triggering the fire extinguishing device, 0 represents not triggering, is the current hazard level; The setting of the fire extinguishing threshold : a. Based on fire history data: collect data of past fire accidents occurred in the factory, including key parameters such as temperature, smoke concentration, flame intensity when the fire occurs, and corresponding hazard degree; for example, statistics show that when the reactor temperature exceeds and the smoke concentration exceeds , the fire often spreads rapidly and causes serious losses, so can be set as the hazard value corresponding to this combination of temperature and smoke concentration; b. Refer to industry standards and specifications: many industries have relevant fire safety standards and specifications, which may have clear requirements or recommendations for the fire extinguishing threshold in a particular environment; for example, in a chemical plant, for some specific chemical equipment and process areas, the relevant standards may stipulate that when the flammable gas concentration exceeds a certain value (such as lower explosive limit), the fire extinguishing device must be automatically started, so the corresponding hazard value can be used as a reference basis for ; c. Combining expert experience: Organize fire experts and experienced engineers to evaluate the fire risk of the factory, and determine the threshold value of valve closing according to their judgment of fire development law and fire extinguishing opportunity . Experts can comprehensively judge the hazard level under which immediate fire extinguishing measures need to be taken according to factors such as the layout of the actual factory, the characteristics of the equipment, the stored materials, etc.

[0027] Valve closing: when the hazard level reaches the valve closing threshold value , the system automatically closes the valve, and the valve closing signal is represented as: ; Setting of the valve closing threshold value : a. According to the safety requirements of equipment operation: for pipelines and equipment that store or transport flammable, explosive, toxic, hazardous and other dangerous media, when their operating parameters exceed the safety range, the valve needs to be closed in time to prevent leakage of dangerous media; for example, for pipelines transporting gasoline, when the pressure in the pipeline exceeds the design pressure or the flow rate is abnormal (such as more than 1.5 times the normal flow rate), it may cause pipeline rupture or leakage, and the hazard value corresponding to the combination of pressure and flow rate at this time can be used as ; b. Consider the severity of the accident consequences: analyze the influence of valve closing on the consequences of accidents under different working conditions to determine a reasonable . If the valve is not closed in time, it may cause a large amount of dangerous media to leak, and in turn cause serious accidents such as fire, explosion, poisoning, etc.; for example, in the liquefied petroleum gas storage tank area, when the pressure sensor of the storage tank detects a sharp rise in pressure, and the rise rate exceeds a certain value (such as 0.2 MPa per minute), it indicates that the storage tank may be in a critical state of overpressure, if the inlet valve is not closed in time, the storage tank may explode, and the hazard value corresponding to the pressure rise rate at this time should be used as ; c. Draw lessons from similar industry accident cases: study similar factory accident cases in the same industry, understand the key factors that led to the valve closing and the corresponding hazard level in these accidents; for example, in a pipeline leakage accident in a chemical enterprise, due to the failure to close the valve in time, a large amount of chemicals leaked and exploded, causing significant casualties and property losses, through the analysis of the accident, the set in similar situations to avoid similar accidents.

[0028] S42: Personnel alarm: Phone / SMS reminder: when the hazard level reaches the alarm threshold value , the system sends a phone or SMS reminder to the personnel in the disaster area, and the alarm information sending signal is represented as: Simultaneously, the system needs to obtain information on people within the disaster area, such as a set of personnel location coordinates. This is to ensure that alarm information is sent accurately; Alarm threshold Setting: a. Based on personnel safety risks: Prioritizing personnel safety, an alarm is triggered when the level of hazard poses a threat to personnel. For example, in a production workshop, if the noise generated by equipment exceeds 90 decibels and persists for a certain period (e.g., 1 minute), it may cause hearing damage to on-site personnel. In this case, the hazard value corresponding to the combination of noise level and duration can be used as a hazard assessment measure. For example, when the concentration of harmful gases in the workshop exceeds a certain value (such as carbon monoxide concentration exceeding...), This poses a significant threat to human health and should trigger an alarm. b. Consider the impact on production operations: For situations that may affect normal production operations but pose a relatively low risk to personnel safety, alarm thresholds can also be set; for example, when a minor malfunction occurs in a critical piece of equipment on the production line, leading to a decrease in production efficiency. While the situation may not immediately affect product quality or personnel safety, it could worsen if not addressed promptly. In this case, the hazard value corresponding to the combination of equipment failure severity and the percentage decrease in production efficiency can be used as a hazard assessment factor. This is so that operators can take timely measures for maintenance and adjustment; c. Refer to industry best practices: Learn from the alarm threshold settings of other companies in the same industry and make appropriate adjustments based on the actual situation of your own factory; for example, some advanced factories in the same industry have set reasonable alarm thresholds for parameters such as equipment vibration and temperature in equipment maintenance management. By learning from and referencing these experiences, you can initially determine the alarm thresholds for your own factory. And it is continuously optimized and improved in actual operation; AR glasses real-time alerts: When When the AR glasses alarm threshold is reached, the AR glasses will issue a real-time warning signal. The trigger expression is similar to that described above; AR glasses alarm threshold Setting: a. Application scenarios for AR glasses: AR glasses are typically used for real-time information prompts and warnings for on-site personnel. In certain high-risk work areas or complex work scenarios, when the level of hazard reaches a point where wearing AR glasses may be inappropriate, they may be used in more dangerous situations. When people wearing the glasses cause disturbance or danger, an AR glasses alarm is triggered; for example, in a high-altitude work site, when the wind speed exceeds [a certain threshold], [an alarm is triggered]. Furthermore, when workers are at a higher position, strong winds may affect their balance and safety. In this case, the hazard value corresponding to the combination of wind speed and working height can be used as... , remind the worker to take windproof measures or stop working in time through AR glasses; b. Consider the effectiveness and timeliness of information prompts: AR glasses warning information should be able to remind personnel in the early stage of hazards, while avoiding excessive invalid warnings that interfere with normal work. Therefore, the alarm threshold needs to be finely adjusted according to the actual working environment and personnel operation habits. For example, in the assembly workshop of an electronics factory, when the equipment on the production line fails and may affect the normal operation of adjacent workstations, considering that the AR glasses need to timely deliver failure information to personnel at relevant workstations, but cannot frequently misreport and interfere with production, the hazard value corresponding to the medium to high degree of equipment failure impact can be set. c. Coordinate with the overall emergency response system: AR glasses warnings should be coordinated with the overall emergency response system of the factory to ensure that, under different hazard levels, the glasses can provide warning information that matches the emergency response measures; for example, when a fire occurs in the factory and the hazard level reaches the point where full evacuation is needed, the threshold value should be coordinated with the fire alarm system to ensure that the AR glasses can timely issue the highest level of evacuation warning information to the wearer, guiding personnel.

[0029] S43: Escape guidance: Intelligent emergency light guidance: The system controls the intelligent emergency light to generate escape guidance signals based on personnel location and disaster situation; assuming that the emergency light has multiple states, such as arrow indicators pointing in different directions, its guidance state can be calculated based on the positional relationship between personnel and safety exits, using simple vector pointing or guidance directions determined based on path planning algorithms; AR glasses generate optimal evacuation path: calculate the optimal path from the current position of personnel to the safety exit using path planning algorithms (such as Dijkstra's algorithm or A* algorithm); let the current position of personnel be , the safety exit position be , and the factory floor plan be represented as a graph structure , where is a node (such as each position point), is an edge (representing the connectivity relationship between positions, and the edge weight can be distance or estimated travel time, etc.); use Dijkstra's algorithm to calculate the shortest path , the core idea of which is to maintain a distance array (storing the shortest distance from the starting node to other nodes, which is maintained and updated in Dijkstra's algorithm) and a priority queue, and each time the node with the smallest distance is taken out for relaxation operation. Finally, the optimal path information obtained is displayed to personnel through AR glasses.

[0030] ​S5: Weight optimizer: Enter accident data into the accident database and update the weights in the hazard value calculation model according to the weight self-optimization mechanism; The data after the disaster occurs is entered into the accident database, and the data after the disaster occurs includes various information at the time of the accident, including but not limited to: accident occurrence time (accurate to the second timestamp, used to record the specific time of the accident), accident location (specific coordinate location within the factory, such as workshop number and equipment number, etc.), personnel involved (personnel information present at the time of the accident, including personnel ID and their role in the accident), equipment state (running state of the equipment at the time of the accident, such as temperature, pressure, etc. Key parameters), environmental conditions (environmental factors at the time of the accident, such as temperature, humidity, etc.), hazard value (calculated by step S3 hazard value at the time of the accident), hazard level (corresponding to the hazard value, used to quickly assess the severity of the accident), these data constitute a complete portrait of the accident, providing a basis for subsequent analysis; Through the weight self-optimization mechanism, the dynamic feature weights are automatically adjusted according to the difference between historical accident data and model prediction results, to improve the accuracy and adaptability of the model, the formula of the weight dynamic updating rule is as follows: , Where: KL is the KL divergence penalty term (true risk distribution vs predicted risk distribution ); is the true risk distribution; is the predicted risk distribution; is the learning rate, which controls the step size of weight update (the specific value of is usually obtained by combining various methods: first, refer to the experience setting or use the framework default value to start training, and observe the training effect; then, refine the selection through grid search or learning rate range test; apply learning rate decay or adaptive method to dynamically adjust during the training process), the initial value is usually small (such as 0.01) to ensure the stability of the update process; is the experience distribution correction coefficient; is the loss function (predicted hazard value vs actual accident loss), which measures the difference between the model prediction result and the actual accident data, commonly used mean square error (MSE): , where: is the hazard value predicted by the model; is the hazard value at the time of the actual accident; is the number of training samples; A digital factory multi-modal disaster protection system based on a dynamic weighted neural network is used to implement a self-optimization method for a digital factory multi-modal disaster protection based on a dynamic weighted neural network, including the following modules: Multivariate data acquisition module: used for real-time collection of digital factory worker, environment and equipment information, including the following sub-modules: Visual information acquisition sub-module: equipped with high-definition cameras, infrared thermal imagers and other visual sensors, deployed in key areas of the factory, around equipment and in densely populated places, real-time acquisition of video stream and image data; for example, install cameras in warehouse areas to monitor cargo stacking and personnel activities; in the production workshop, use infrared thermal imagers to monitor temperature changes when equipment is running; preliminary preprocessing of collected visual data, such as image size adjustment, format conversion, etc., to meet subsequent processing needs and ensure data consistency and processability.

[0031] Device operation information acquisition sub-module: connected to the control systems of various devices in the factory (such as PLC, DCS, etc.), real-time reading of device operating parameters, including temperature, pressure, speed, current, voltage, etc.; for example, in chemical production equipment, accurately collect the temperature and pressure changes in the reaction kettle, as well as the motor speed and current data, and timely grasp the running state of the equipment; monitor the fault alarm signals of the equipment, once the equipment appears abnormal conditions such as overheating, overload, etc., the alarm information can be immediately captured and input as important data to the subsequent evaluation link.

[0032] Environmental information acquisition sub-module: deploy environmental sensors such as temperature and humidity sensors, smoke sensors, flammable gas sensors, toxic gas sensors, etc., to monitor the factory environment comprehensively, install corresponding sensors in factory workshops, warehouses, pipelines, etc., to obtain real-time environmental data; real-time monitoring and analysis of environmental data, when the environmental parameters exceed the safety threshold, timely warning signals are sent; for example, when the smoke sensor detects that the smoke concentration exceeds the set safety value, or the flammable gas sensor detects gas leakage, these key information is quickly transmitted to the subsequent processing module Feature fusion module: used for processing real-time data and constructing feature tensor, including the following sub-modules: Visual feature extraction sub-module: receives visual data from the multivariate data acquisition module, uses 3D convolution network to process video stream, extracts spatial-motion features, specifically, a series of continuous video frames collected by the camera are input into the 3D convolution neural network, through convolution layers, pooling layers, etc. operations, capture the position changes, motion trajectories and shape features of objects in space; for example, in the personnel behavior monitoring of the factory, by extracting the motion features of personnel in the video, it is judged whether the personnel have abnormal behaviors such as running quickly in the prohibited area, suddenly falling down, etc., to provide basis for subsequent safety evaluation.

[0033] Sensor time series feature processing submodule: For the time series data transmitted by the equipment operation information acquisition unit and the environment information acquisition unit, the long short-term memory network (LSTM) and other time series processing algorithms are used to mine the time series dependency relationship in the data; for the temperature, pressure and other parameter change sequences of the equipment over time, LSTM can learn the correlation between these parameters at different times, and extract features reflecting the running trend and state change of the equipment; for example, through time series analysis of the temperature data in the chemical production equipment reactor, it is determined whether the temperature change conforms to the normal production process curve, if there is an abnormal temperature fluctuation or rising trend, this key feature is extracted in time for subsequent hazard assessment.

[0034] Metadata processing submodule: The metadata of the digital factory is sorted and processed, including the layout information of the factory, the model and location of the equipment, the process flow parameters, the post and responsibility of the personnel, etc.; these structured metadata are normalized and coded to make them compatible with visual features and sensor time series features; for example, the installation position coordinates of the equipment are combined with the equipment running state features to accurately determine the specific location of the potential hazard source in the factory during disaster assessment, providing precise spatial information support for subsequent emergency response.

[0035] Feature tensor construction submodule: The processed and extracted visual features, sensor time series features and metadata features are spliced and combined according to certain rules to form a feature tensor; this feature tensor is a multi-dimensional data structure that can comprehensively and comprehensively represent the state information of personnel, equipment and environment of the digital factory at the current time; the construction process of the feature tensor needs to consider the correlation and complementarity between different features, and through reasonable design of fusion strategy, it is ensured that the feature tensor can fully reflect the actual situation of the factory, providing high-quality input data for subsequent dynamic hazard assessment.

[0036] Dynamic hazard assessment module: used to obtain the hazard value and match the hazard level through the hazard value calculation model, including the following submodules: Hazard Value Calculation Submodule: Receives the feature tensor constructed by the feature fusion module and inputs it into the pre-trained dynamic weighting neural network model. According to the weights of various features in the feature tensor, the model calculates the hazard value under the current factory state through nonlinear combination operation; for example, for a situation where equipment is running at high temperature, environmental smoke concentration is rising, and personnel are active in dangerous areas, the model will consider the different influences of these factors on the degree of harm, assign appropriate weights, and calculate a value reflecting the overall degree of harm; the weights of the dynamic weighting neural network model will be adjusted according to system feedback and actual situation to ensure that the calculation of the hazard value can accurately reflect the actual risk situation of the factory. The adjustment of weights is based on historical data and accident cases accumulated in the system, which are automatically optimized through machine learning algorithms, so that the model can adapt to changes in factory environment and production process.

[0037] Hazard Level Matching Submodule: According to the calculated hazard value, compare it with the preset hazard level threshold to divide the degree of harm into different levels, such as low, medium, high, and extremely dangerous; these thresholds are set according to a large amount of historical accident data, expert experience, and the specific safety standards of the factory, which can accurately reflect the risk level corresponding to different degrees of harm; for example, when the hazard value is below a certain low-risk threshold, the system determines that the current factory is in a safe state; when the hazard value exceeds the high-risk threshold, the corresponding high-risk alarm is triggered immediately, and the hazard level information is transmitted to the automatic response system module to take effective emergency measures in time.

[0038] Automatic Response System Module: Used to automatically start response measures according to the hazard level, including the following submodules: Device Linkage Control Submodule: Connected with various safety devices and automation control systems in the factory, automatically triggers corresponding device linkage operations according to the hazard level signal output by the dynamic hazard assessment module; for example, when the hazard level reaches high, it can automatically start the factory's fire extinguishing device to spray water or release fire extinguishing gas in the fire area; in the case of gas leakage, automatically close the related valves to cut off the leakage source and prevent the accident from further expanding; the execution of device linkage operations is monitored and fed back in real time to ensure that the devices can act correctly according to the predetermined instructions, and if device linkage fails or malfunctions, an alarm is sent to the system in time and attempts are made to take alternative solutions or notify relevant personnel for manual intervention to ensure the reliability of emergency response.

[0039] Personnel alarm notification submodule: Use various communication means, such as factory internal broadcast system, SMS platform, telephone system and mobile application based push notification, to send alarm information to personnel in disaster area, alarm information content includes disaster type, occurrence location, severity and emergency evacuation instruction, etc.; for example, in the event of fire, broadcast fire alarm to each area of the factory through the broadcast system, inform the specific fire location and evacuation direction; at the same time, send SMS or push mobile application notification to relevant personnel, ensure that even in noisy environment or personnel in mobile state, alarm information can be received in time, improve the alertness and emergency response speed of personnel.

[0040] Escape guidance indication submodule: Based on the layout information of the factory and the location of the current disaster, combined with the real-time location data of the personnel (which can be obtained through personnel positioning system, such as Wi-Fi, Bluetooth positioning or indoor positioning sensor, etc.), use path planning algorithm (such as Dijkstra algorithm, A* algorithm, etc.) to calculate the optimal escape path; through intelligent emergency light, AR glasses and other devices to provide intuitive escape guidance for personnel, intelligent emergency light can flash different light signals according to the indication direction of escape path, guide personnel to evacuate to safety exit; AR glasses can display real-time escape route, safety exit location and refuge area information in the field of view of personnel, help personnel quickly and accurately find the escape direction in complex and chaotic disaster scene, improve the success rate of escape.

[0041] Weight optimization module: used for entering accident data into accident database, and updating weight in hazard value calculation model according to weight self-optimization mechanism, including the following submodules: Accident data management submodule: responsible for collecting, sorting and storing various accident data occurred in digital factory, accident data includes time, location, personnel casualty, equipment damage degree, environmental influence range and corresponding hazard value and hazard level, etc. Detailed information; classify and label accident data, according to different disaster types, severity and other factors for archiving management, convenient for subsequent data query, analysis and mining, through in-depth analysis of accident data, provide rich sample data and actual case support for weight optimization, help the system better understand the characteristics and rules of different types of disasters.

[0042] Weight adjustment algorithm submodule: Use optimization algorithms in machine learning, such as gradient descent, genetic algorithm, particle swarm optimization algorithm, etc., to automatically adjust the weights in the dynamic weighting neural network model combined with accident data and system historical operation data; taking gradient descent as an example, calculate the gradient of the loss function (such as mean square error) on the weight, update the weight value along the gradient descent direction, make the model output more close to the actual accident data hazard value and hazard level, so as to improve the accuracy and reliability of the model; regularly evaluate and verify the weight optimization results, through cross-validation, test set evaluation and other methods, ensure that the optimized weights have good generalization performance on new data samples, avoid overfitting or underfitting; if the optimization effect is not ideal, adjust the parameters of the optimization algorithm or select other more appropriate algorithms for re-optimization, ensure the efficiency and effectiveness of the weight optimization module.

[0043] System integration and communication module: used for real-time monitoring and management of the running state of each module, coordinating task scheduling and resource allocation among modules; and used for data transmission and communication among modules, including the following submodules: Data transmission and interface submodule: responsible for data transmission and communication among modules, ensuring that multi-element data acquisition module, feature fusion module, dynamic hazard assessment module, automatic response system module and weight optimization module can interact and share data in real time and accurately, using high-speed and stable data transmission protocols (such as TCP / IP, MQTT, etc.), establishing communication interfaces between modules to meet the real-time requirements of the system for data; encrypt and protect data transmission process to prevent data from being stolen, tampered with or subjected to malicious attacks, ensuring the data security and reliability of the system, while optimizing and managing the bandwidth, delay and other performance indicators of data transmission to ensure the stable operation of the system in complex factory environment.

[0044] Central control and coordination submodule: as the control center of the whole system, it monitors and manages the running state of each module, coordinates the task scheduling and resource allocation among modules; according to the overall operation situation and priority setting of the system, reasonably arrange the work tasks of each module, ensure that the system can respond quickly and efficiently in case of disaster, and fully play the function of each module; provide a man-machine interface for the system, which is convenient for operators to configure, monitor and operate the system, through an intuitive graphical interface, operators can real-time view the running state of the factory, disaster warning information, equipment linkage situation and personnel evacuation progress, etc., timely intervene and adjust the system, improve the operability and controllability of the system.

[0045] System running process: 1. The multi-element data acquisition module collects visual information, equipment operation information and environmental information of the digital factory in real time, and transmits the raw data to the feature fusion module.

[0046] 2. The feature fusion module processes and extracts features from the collected multi-element data, including visual features, sensor time sequence features and metadata extraction and fusion, constructs a feature tensor, and then transmits the feature tensor to the dynamic hazard assessment module.

[0047] 3. The dynamic hazard assessment module uses a dynamic weighted neural network model to calculate the feature tensor, obtains a hazard value, and compares it with a preset threshold to determine a hazard level, and sends the hazard level information to the automatic response system module and the weight optimization module.

[0048] 4. The automatic response system module triggers corresponding device linkage control, personnel alarm notification and escape guidance indication operation according to the received hazard level signal, realizes timely response and processing of disasters.

[0049] 5. The weight optimization module continuously collects accident data during system operation, and adjusts and optimizes the weights in the dynamic weighted neural network model according to the accident data and optimization algorithm, and feeds back the optimized weights to the dynamic hazard assessment module to improve the evaluation accuracy of the model.

[0050] 6. The system integration and communication module is responsible for data transmission, communication coordination and central control management between modules during the whole system operation, ensures the stable and efficient operation of the system, and realizes the effective protection of the multi-modal disaster of the digital factory.

[0051] From the above description, it can be seen that the above-mentioned embodiments of the present application realize the following technical effects: the dynamic hazard assessment algorithm system, the feature tensor constructed by the multi-source data fusion layer, the hazard value calculation model to obtain the hazard value, and the weight self-optimization mechanism to automatically optimize the weights form a closed-loop architecture of algorithm-data-execution, solving the problems of response lag and evaluation rigidity of traditional factory safety systems.

[0052] Those skilled in the art should understand that the discussion of any of the above embodiments is only exemplary and is not intended to imply that the scope of the present application is limited to these examples; under the idea of the present application, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other changes of different aspects of the present application as described above. In order to be brief, they are not provided in detail.

[0053] The present application is intended to embrace all such alternatives, modifications, and variations that fall within the broad scope of the appended claims. Accordingly, any and all such alternatives, modifications and variations should be included within the scope of the present application.

Claims

1. A self-optimization method for digital factory multi-modal disaster prevention based on a dynamically weighted neural network, characterized in that, Comprise the following specific steps: S1: Multivariate data acquisition: Real-time collection of digital factory worker, environment and equipment information; S2: Real-time feature fusion: Process real-time data and build feature tensor; S3: Dynamic hazard assessment engine: Obtain hazard value and match hazard level through hazard value calculation model; S4: Automatic response system: Automatically start response measures according to hazard level; S5: Weight optimizer: Enter accident data into accident database, and update weights in hazard value calculation model according to weight self-optimization mechanism.

2. The self-optimization method for digital factory multi-modal disaster protection based on dynamic weighted neural network according to claim 1, characterized in that, In the S1, the collected multivariate data includes digital factory itself data and data generated by digital factory cameras and sensors, wherein the digital factory itself data includes but is not limited to personnel information and equipment information, and the data generated by the digital factory cameras and sensors includes but is not limited to camera video stream, temperature, humidity, vibration and gas data.

3. The self-optimization method for digital factory multi-modal disaster protection based on dynamic weighted neural network according to claim 2, characterized in that, The S2 specifically comprises: obtaining a feature tensor by fusing the data obtained in step S1; S21: The data collected in step S1 is integrated into three categories: visual features, sensor time series features and metadata, wherein the visual features are obtained by extracting spatial-motion features through a 3D convolution network according to the visual related data obtained in step S1; the sensor time series features are obtained by processing sensor time series dependency according to the sensor data in step S1; and the metadata is digital factory itself data; The formula for constructing the feature tensor in S22 is as follows: ， Where: T is the feature tensor; V is the camera video stream; S t is the sensor matrix; M is the device or person metadata; Concat is to concatenate the three types of data to construct the feature tensor; CNN is the 3D convolutional network to extract spatial-motion features; LSTM is to handle the temporal dependence of the sensor time series; Embed is to directly embed the device or person metadata.

4. The self-optimization method for digital factory multi-modal disaster protection based on dynamic weighted neural network according to claim 3, characterized in that, The S21 specifically comprises: S211: visual feature extraction: taking the camera video stream as an example, it is a three-dimensional data structure, which can be represented as , wherein represents the video frame image data at the moment, which is input into the 3D convolution network, and the convolution operation can be represented as: ， wherein: is the feature map after convolution; , , respectively represent input video frame image data row index, column index in spatial dimension and index in time dimension; respectively are the size of the convolution kernel in spatial and time dimensions; , , respectively represent row index, column index of the convolution kernel in spatial dimension and index in time dimension; is the bias term; After nonlinear transformation by an activation function, visual features containing spatial-motion features can be obtained; S212: Sensor time series feature processing: assuming that the data collected by the sensor is a one-dimensional time series wherein denotes the sensor measurement value at the time point, for processing the time series dependency relationship, a long short-term memory network, LSTM, is adopted, and the memory cell state update formula of the LSTM is as follows: ， ， ， ， ， ， wherein: are outputs of a forget gate, an input gate, and an output gate, respectively; is a candidate memory cell state; is an updated memory cell state; is a hidden state; are a weight matrix and a bias term, respectively; is a sigmoid activation function; denotes element-wise multiplication; is a hyperbolic tangent function; The final hidden state sequence is the processed sensor time series feature that contains temporal dependencies.

5. The self-optimization method for digital factory multi-modal disaster protection based on dynamic weighted neural network according to claim 1, characterized in that, The S3 specifically comprises: S31: Input the feature tensor calculated in step S2 and other parameters into a hazard value calculation model to obtain a hazard value, and the formula of the hazard value calculation model is as follows: ， Wherein: H is the hazard value; is a Sigmoid normalization function; f i is the deep feature extractor of the i th feature channel; T is the feature tensor; is a dynamic feature weight; is a negative feedback adjustment coefficient; ReLU is an activation function; SD is a protective measure feature vector; W d is the weight of the feature vector SD; S32: Match the hazard level corresponding to the hazard value obtained by the hazard value calculation model, and output the hazard level to the automatic response system.

6. The self-optimization method for digital factory multi-modal disaster protection based on dynamic weighted neural network according to claim 5, characterized in that, The S32 specifically comprises: Setting a series of hazard level thresholds , satisfying , comparing the calculated hazard values to these thresholds to match a hazard level: If then the match is of hazard level 1, indicating a normal state; If then the match is a hazard level 2, indicating a light hazard; If then the match is of a hazard level which indicates an extremely serious hazard.

7. The self-optimization method for digital factory multi-modal disaster protection based on dynamic weighted neural network according to claim 1, characterized in that, The S4 specifically comprises: The automatic response system triggers the automatic response measures corresponding to the hazard level obtained in step S3, and the automatic response measures include but are not limited to: device linkage, personnel alarm and escape guidance.

8. The self-optimization method for digital factory multi-modal disaster protection based on dynamic weighted neural network according to claim 1, characterized in that, The S5 specifically comprises: Enter the data after the disaster into the accident database, automatically adjust the dynamic feature weights according to the difference between the historical accident data and the model prediction results through the weight self-optimization mechanism, and the formula of the weight dynamic update rule is as follows: ， wherein: KL is a KL divergence penalty term; is a true risk distribution; is a predicted risk distribution; is a learning rate; is an experience distribution correction coefficient; is a loss function.

9. The self-optimization method for digital factory multi-modal disaster protection based on dynamic weighted neural network according to claim 8, characterized in that, In the S5, the data after the disaster includes various information at the time of the accident, including but not limited to: accident time, accident location, personnel involved, equipment state, environmental conditions, hazard value and hazard level.

10. A digital factory multi-modal disaster protection system based on dynamic weighted neural network, characterized in that, A self-optimization method for realizing the digital factory multi-modal disaster protection based on dynamic weighted neural network according to any one of claims 1-9, comprising the following modules: Multivariate data acquisition module: for real-time collection of digital factory worker, environment and equipment information; Feature fusion module: for processing real-time data and building feature tensor; Dynamic hazard assessment module: for obtaining hazard value and matching hazard level through hazard value calculation model; Automatic response system module: used for automatically starting response measures according to the hazard level; Weight optimization module: used for entering accident data into an accident database and updating weights in a hazard value calculation model according to a weight self-optimization mechanism; System integration and communication module: used for real-time monitoring and management of the running states of various modules, coordination of task scheduling and resource allocation among the modules, and data transmission and communication among the modules.

Images