Station building safety risk online evaluation system based on edge AI large model

Abstract

The invention discloses a station building safety risk online assessment system based on an edge AI large model, which relates to the technical field of station building safety systems and comprises a data acquisition module, a large model reasoning module, an edge calculation module, a cloud collaboration module, a risk assessment module and a user interaction module. The multi-modal fusion module can perform fusion analysis on multi-modal data such as videos, sensors, texts and the like to form comprehensive risk assessment, the multi-modal fusion module adopts dynamic weighting, utilizes an attention mechanism and dynamically distributes weights according to the importance of modals, and through the design, the working personnel can perform multi-modal risk assessment according to the current time period, the working state or external factors. In addition, the multi-modal fusion module supports adaptive sampling and dynamic updating in the use process, and the data sampling frequency of the data acquisition module is dynamically adjusted according to the risk level, so that the flexibility of the system is further improved.

Description

Technical Field

[0001] This invention relates to the field of railway station safety system technology, specifically to an online assessment system for railway station safety risks based on an edge AI large model. Background Technology

[0002] The online safety risk assessment system for railway stations using edge AI big data models is an innovative safety management solution. It aims to improve the safety and efficiency of railway station operations through advanced technology. The system combines edge computing and artificial intelligence big data model technology, enabling it to accurately assess and manage the safety risks of railway stations in a real-time environment. This provides an intelligent and efficient solution for railway station safety management, helping to prevent safety accidents and ensure the safety and stability of railway station operations.

[0003] Currently available online station safety risk assessment systems typically require processing large amounts of data. However, when dealing with large volumes of data, data collection, transmission, and processing may become slow or even crash due to excessive processing volume. In addition, data from different modalities may have different collection frequencies and inconsistent data quality, leading to delays or information loss. Furthermore, the standards of existing online station safety risk assessment systems are relatively uniform and cannot be flexibly adjusted according to the specific conditions of the monitored station.

[0004] Therefore, in view of this, we study and improve the existing structure and its shortcomings, and propose an online assessment system for station safety risks based on edge AI large model. Summary of the Invention

[0005] The purpose of this invention is to provide an online assessment system for station safety risks based on edge AI large models, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] An online assessment system for station safety risks based on edge AI large model includes a data acquisition module, a large model inference module, an edge computing module, a cloud collaboration module, a risk assessment module, and a user interaction module;

[0008] The output of the data acquisition module is connected to the large model inference module, and the output of the large model inference module is connected to the edge computing module, the cloud collaboration module, and the risk assessment module. The output of the risk assessment module is connected to the user interaction module.

[0009] The data acquisition module is used to collect data on temperature, humidity, air pressure, and smoke concentration in the station building, as well as temperature, current, voltage, and vibration data of electrical equipment. It also collects video streams inside and outside the station building, personnel entry and exit records, movement trajectories, and past risk events in the station building, and processes the collected data.

[0010] The large model inference module is used to analyze sensor data and video streams in real time, detect abnormal behavior or abnormal state, and predict potential risks based on historical data of equipment failure, fire, and illegal intrusion and real-time collected data. At the same time, it integrates and analyzes multimodal data such as video, sensor, and text to form a comprehensive risk assessment.

[0011] The edge computing module is used to run large AI models on edge devices, reduce data transmission latency, achieve real-time risk assessment, optimize performance and power consumption, and still perform basic risk assessment when the network is interrupted. It also utilizes an edge caching mechanism to temporarily store data in the event of network instability and upload it in batches after the network is restored.

[0012] The cloud collaboration module is used to periodically train and optimize large models and distribute the updated models to edge devices.

[0013] The risk assessment module is used to predict potential risks based on historical and real-time data of equipment failures, fires, and illegal intrusions.

[0014] The user interaction module is used by users to view risk assessment results and query station risk status, and periodically generate station safety risk reports;

[0015] The data acquisition module includes an environmental data acquisition module, an equipment operation status acquisition module, a video surveillance data acquisition module, a personnel activity data acquisition module, a historical data acquisition module, and a data preprocessing module. The data preprocessing module is used to fill or remove outliers in the data acquired by the data acquisition module, reduce noise in the data, align the time, and unify multimodal data to the same time dimension. In addition, the data preprocessing module uses keyframe extraction and ROI detection to transmit only risk areas or abnormal frames. The data preprocessing module evaluates the importance of the data in real time through a lightweight large model on the edge device and uploads important data to the cloud.

[0016] The large model inference module includes an anomaly detection module, a risk prediction module, a semantic understanding module, and a multimodal fusion module. The multimodal fusion module adopts dynamic weighting and uses an attention mechanism to dynamically allocate weights according to the importance of the modality. During use, the multimodal fusion module supports adaptive sampling and dynamic updates, and dynamically adjusts the data sampling frequency of the data acquisition module according to the risk level. The sampling frequency is reduced when the environmental parameters fluctuate less and increased when the environmental parameters fluctuate more.

[0017] Furthermore, the environmental data acquisition module is used to collect data on temperature, humidity, air pressure, and smoke concentration in the station building, while the equipment operation status acquisition module is used to collect data on temperature, current, voltage, and vibration of the electrical equipment in the station building.

[0018] Furthermore, the video surveillance data acquisition module is used to collect video streams inside and outside the station building, the personnel activity data acquisition module is used to collect personnel entry and exit records and movement trajectories, and the historical data acquisition module is used to collect past risk events and equipment maintenance records of the station building.

[0019] Furthermore, the anomaly detection module is used to analyze sensor data and video streams in real time to detect abnormal behavior or abnormal states, and the risk prediction module predicts potential risks based on historical data of equipment failure, fire, and illegal intrusion, as well as real-time collected data.

[0020] Furthermore, the semantic understanding module parses operator instructions and logs through natural language processing to identify potential problems, and the multimodal fusion module fuses and analyzes multimodal data, including video, sensor, and text, to form a comprehensive risk assessment.

[0021] Furthermore, the edge computing module includes a real-time processing module, a low-power, high-efficiency computing module, a network outage fault-tolerant module, and an edge caching module. The real-time processing module runs large AI models on edge devices, and the low-power, high-efficiency computing module uses lightweight large models.

[0022] Furthermore, the network outage fault tolerance module supports offline processing, and can still perform basic risk assessment when the network is interrupted. The edge caching module uses the edge caching mechanism to temporarily store data in the case of network instability, and upload it in batches after the network is restored to avoid data loss.

[0023] Furthermore, the cloud collaboration module includes a model update module, a data synchronization and storage module, and a global optimization module. The model update module performs regular training and optimization on the large model and distributes the updated model to edge devices.

[0024] Furthermore, the risk assessment module includes a risk rating module, a visualization module, an alarm mechanism module, and an emergency plan recommendation module. The risk rating module is divided into three levels: low, medium, and high, based on the severity of the risk. The visualization module displays the station's risk status in an intuitive way, such as charts and heat maps.

[0025] Furthermore, the user interaction module includes a visual interface module, a voice interaction module, and a report generation module. The visual interface module provides multi-platform support, the voice interaction module allows users to query the station's risk status via voice, and the report generation module periodically generates station safety risk reports, supporting export to PDF and Excel.

[0026] This invention provides an online assessment system for station safety risks based on a large edge AI model, which has the following advantages:

[0027] 1. The data preprocessing module of this invention can fill in or remove constant sensor disconnections and device unresponsiveness in the collected data during the data processing process. At the same time, it can reduce noise in noisy data such as video flickering and sensor jitter, which can reduce the difficulty of subsequent data processing. In addition, the data preprocessing module can unify the video frames and sensor acquisition frequencies of multimodal data to the same time dimension. This can effectively avoid the situation where different acquisition frequencies and inconsistent data quality of different modal data lead to data transmission delays or information loss. This can greatly improve the system's working efficiency and stability. In addition, the data preprocessing module can also use keyframe extraction and ROI region of interest detection to transmit only risky areas or abnormal frames, thereby reducing the data processing burden of the system. At the same time, the data preprocessing module can evaluate the importance of data in real time through a lightweight large model on the edge device and upload only important data to the cloud, thereby reducing the difficulty of log reading.

[0028] 2. The multimodal fusion module of this invention can fuse and analyze multimodal data such as video, sensor, and text to form a comprehensive risk assessment. The multimodal fusion module adopts dynamic weighting and uses an attention mechanism to dynamically allocate weights according to the importance of the modality. Through this design, staff can adjust the risk assessment standards according to the current time period, work status, or external factors, making the risk assessment of the equipment more flexible. In addition, the multimodal fusion module supports adaptive sampling and dynamic updates during use. It dynamically adjusts the data sampling frequency of the data acquisition module according to the risk level. The sampling frequency can be reduced when the environmental parameters fluctuate less and increased when the fluctuation is greater, which further improves the flexibility of the system.

[0029] 3. The dynamic weighting of the multimodal fusion module of this invention can also automatically adjust weights based on multiple levels and data types. When the system adjusts weights based on the equipment layer, the weight of key equipment (such as the main transformer) can be higher than that of secondary equipment (such as auxiliary equipment). Under extreme weather conditions such as high temperature, high humidity, and thunderstorms, the weight of environmental factors will be increased. The weight of environmental data can be obtained through edge device networking or adjusted by obtaining historical environmental data from logs. The dynamic weighting of the multimodal fusion module can also utilize regional layer weights for weight adjustment. When adjusting regional layer weights, the system can combine regional risk handling. When multiple stations are located in the same area, regional risks can be considered (…). (For example, during typhoons and floods) the weight of each station building is adjusted. When a station building is located in an area prone to blizzards, high temperatures, floods, thunderstorms, etc., its weight will automatically increase during the above-mentioned risk periods. At the same time, the multimodal fusion module can also consider the dynamic changes of the time dimension with time-layer weights. The weight distribution of the time layer can be different according to the weight distribution of different time periods. During the night, the weight of personnel activity monitoring will be reduced, while the weight of equipment automated operation status will be increased. During the period when equipment is operating under high load (such as the summer peak electricity consumption), the weight of equipment status will be increased. In addition, the system can also dynamically learn the optimal weight distribution based on the real-time risk assessment effect and user feedback (when a user reports that an assessment result is misjudged as low risk, the system will automatically adjust the weight of relevant factors).

[0030] 4. When the edge computing module of this invention is working, the real-time processing module can run large AI models on edge devices, reducing data transmission latency and enabling real-time risk assessment. The low-power, high-efficiency computing module uses a lightweight large model to optimize performance and power consumption. By working with the data preprocessing module, the system's workload and assessment accuracy can be further improved. In addition, the network outage fault tolerance module supports offline processing and can still perform basic risk assessments when the network is interrupted. The edge caching module utilizes the edge caching mechanism to temporarily store data in the event of network instability and upload it in batches after the network is restored, avoiding data loss. Through the above design, the system's working stability can be greatly improved.

[0031] 6. This invention completes sensitive data processing locally through an edge computing module, reducing the frequency of uploading critical data (such as video surveillance and personnel trajectories) to the cloud, lowering the risk of data leakage, and meeting the needs of scenarios with strict privacy protection requirements. At the same time, the edge device adopts a lightweight large model, reducing the dependence on hardware computing power, and can be deployed on old equipment or low-cost edge nodes, reducing the initial investment in the intelligent transformation of the station. In addition, the decoupled design between the data acquisition module, preprocessing module, and risk assessment module makes it easy to flexibly add or remove functions according to the needs of different stations, such as adding new sensor types or expanding risk assessment dimensions.

[0032] 7. This invention integrates and analyzes multi-source data such as video, sensor, and text. Even if a single data source (such as a camera malfunction) fails, the system can still maintain its risk assessment function through other data, avoiding a complete system crash. Furthermore, the network outage fault tolerance module works in conjunction with the cloud collaboration module, allowing the edge to continue operating independently and caching data during network fluctuations or cloud service interruptions, ensuring continuous system availability. Additionally, the system can automatically optimize weight allocation strategies and risk assessment logic through user feedback and global data analysis, achieving a virtuous cycle. The use of a dynamic weighting mechanism can automatically adjust assessment strategies for special environments (such as low-oxygen monitoring in high-altitude stations or salt spray corrosion monitoring in coastal stations), expanding the system's applicable geographical range. Moreover, by dynamically adjusting the data sampling frequency and low-power computing modules, energy consumption is reduced during off-peak hours, aligning with the trend of low-carbon operation. Attached Figure Description

[0033] Figure 1 This is a flowchart illustrating the overall operation of an online station safety risk assessment system based on an edge AI large model, as described in this invention.

[0034] Figure 2 This is a schematic diagram of the data acquisition module of an online station safety risk assessment system based on an edge AI large model according to the present invention;

[0035] Figure 3 This is a schematic diagram of the large model inference module of an online assessment system for station safety risks based on edge AI large model according to the present invention;

[0036] Figure 4 This is a schematic diagram of the edge computing module of an online assessment system for station safety risks based on a large edge AI model, according to the present invention.

[0037] Figure 5 This is a schematic diagram of the cloud collaboration module of an online assessment system for station safety risks based on an edge AI large model, according to the present invention.

[0038] Figure 6 This is a schematic diagram of the risk assessment module of an online station safety risk assessment system based on an edge AI large model according to the present invention;

[0039] Figure 7 This is a schematic diagram of the user interaction module of an online station safety risk assessment system based on an edge AI large model according to the present invention.

[0040] Figure 8 This is a schematic diagram of the online assessment process of an online station safety risk assessment system based on an edge AI large model, according to the present invention. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] Please see Figures 1 to 8 The present invention provides a technical solution: an online assessment system for station safety risks based on edge AI large model, including a data acquisition module, a large model inference module, an edge computing module, a cloud collaboration module, a risk assessment module and a user interaction module. The data acquisition module includes an environmental data acquisition module, an equipment operation status acquisition module, a video surveillance data acquisition module, a personnel activity data acquisition module, a historical data acquisition module and a data preprocessing module.

[0043] The environmental data acquisition module is used to collect data on temperature, humidity, air pressure, and smoke concentration in the station building; the equipment operation status acquisition module is used to collect data on temperature, current, voltage, and vibration of the electrical equipment in the station building; the video surveillance data acquisition module is used to collect video streams inside and outside the station building; the personnel activity data acquisition module is used to collect personnel entry and exit records and movement trajectories; the historical data acquisition module is used to collect past risk events in the station building and equipment maintenance records; the anomaly detection module is used to analyze sensor data and video streams in real time to detect abnormal behavior or abnormal states; and the risk prediction module predicts potential risks based on historical data and real-time collected data on equipment failures, fires, and illegal intrusions.

[0044] The data preprocessing module fills in or removes outliers such as sensor disconnections and device unresponsiveness in the acquired data. It also reduces noise in the video footage, addressing flickering and sensor jitter. Furthermore, it aligns the time frame, unifying the video frames and sensor acquisition frequencies of the multimodal data to the same time dimension. Additionally, the module can extract keyframes and detect regions of interest (ROIs), transmitting only high-risk or outlier frames. The data preprocessing module uses a lightweight, large-scale model on the edge device to assess the importance of the data in real time, uploading only the most important data to the cloud.

[0045] The importance analysis formula for real-time evaluation data in the data preprocessing module is as follows:

[0046] Sensor data Calculate the score for each data point:

[0047] ;

[0048] in, It is the average value of the sensor data. It is the fluctuation characteristic of the data. For weight, only transmit The video surveillance data acquisition module uses the following formula to extract keyframes from video data:

[0049] For video stream Perform keyframe selection and retain frames. The following conditions must be met: ;

[0050] in, The difference between the current frame and the previous frame is measured by pixel changes or structural similarity. The preset threshold and large model inference module include an anomaly detection module, risk prediction module, semantic understanding module, and multimodal fusion module. The anomaly detection module can analyze sensor data and video streams in real time to detect abnormal behavior or abnormal states. The risk prediction module can predict potential risks based on historical data and real-time collected data of equipment failure, fire, and illegal intrusion. The semantic understanding module can parse operator instructions and logs through natural language processing to identify potential problems. The multimodal fusion module can fuse and analyze multimodal data such as video, sensor, and text to form a comprehensive risk assessment.

[0051] The multimodal fusion module employs dynamic weighting, utilizing an attention mechanism to dynamically allocate weights based on the importance of each modality. The formula is as follows:

[0052] ;

[0053] In addition, the multimodal fusion module supports adaptive sampling and dynamic updates during use. It dynamically adjusts the data sampling frequency of the data acquisition module according to the risk level. It can reduce the sampling frequency when the environmental parameters fluctuate less and increase the sampling frequency when the fluctuations are greater.

[0054] The edge computing module includes a real-time processing module, a low-power, high-efficiency computing module, a network outage fault-tolerant module, and an edge caching module. The real-time processing module runs large AI models on edge devices, reducing data transmission latency and enabling real-time risk assessment. The low-power, high-efficiency computing module uses lightweight large models to optimize performance and power consumption. The network outage fault-tolerant module supports offline processing and can still perform basic risk assessments when the network is interrupted. The edge caching module utilizes an edge caching mechanism to temporarily store data in unstable network conditions and upload it in batches after the network is restored, avoiding data loss.

[0055] The cloud-based collaboration module includes a model update module, a data synchronization and storage module, and a global optimization module. The model update module performs regular training and optimization of large models in the cloud and distributes the updated models to edge devices. The data synchronization and storage module uploads key data from the edge devices to the cloud for long-term storage and in-depth analysis. The global optimization module aggregates data from multiple stations in the cloud to identify global risks.

[0056] The risk assessment module includes a risk rating module, a visualization module, an alarm mechanism module, and an emergency plan recommendation module. The risk rating module classifies risks into three levels: low, medium, and high, based on their severity. The visualization module displays the station's risk status in an intuitive way, using charts, heat maps, and other methods. The alarm mechanism module supports multiple alarm formats, including sound, SMS, and APP push notifications, and provides detailed risk analysis reports. The emergency plan recommendation module recommends corresponding emergency response plans based on the type of risk.

[0057] The user interaction module includes a visual interface module, a voice interaction module, and a report generation module. The visual interface module provides multi-platform support for PCs, mobile devices, and tablets, making it convenient for users to view risk assessment results. The voice interaction module allows users to query the risk status of the station building via voice. The report generation module periodically generates station building safety risk reports and supports exporting to PDF and Excel.

[0058] In summary, this online safety risk assessment system for railway stations based on edge AI large-scale models, when in use, firstly, the data acquisition module collects data on temperature, humidity, air pressure, and smoke concentration within the station; the equipment operation status acquisition module collects data on temperature, current, voltage, and vibration of electrical equipment within the station; in addition, the video surveillance data acquisition module collects video streams inside and outside the station; the personnel activity data acquisition module collects personnel entry and exit records, movement trajectories, and historical data; and the historical data acquisition module collects past risk events and equipment maintenance records. After the above modules collect the data, it can be sent to the data preprocessing module for data processing. During the data processing, the data preprocessing module can identify and handle issues such as constant sensor disconnections and equipment malfunctions in the collected data. The system should perform filling or culling, and can also reduce noise in video footage and sensor jitter, which can reduce the difficulty of subsequent data processing. In addition, the data preprocessing module can unify the video frames and sensor acquisition frequencies of multimodal data to the same time dimension. This can effectively avoid the situation where different acquisition frequencies and inconsistent data quality of different modal data lead to data transmission delays or information loss. This can greatly improve the system's work efficiency and stability. In addition, the data preprocessing module can also use keyframe extraction and ROI region of interest detection to transmit only risky areas or abnormal frames, thereby reducing the data processing burden of the system. At the same time, the data preprocessing module can evaluate the importance of data in real time through lightweight large models on edge devices, and only upload important data to the cloud to reduce the difficulty of log reading.

[0059] After the data preprocessing module processes the data, it transmits the data to the large-scale model inference module via the network. The anomaly detection module within the large-scale model inference module analyzes sensor data and video streams in real time to detect abnormal behaviors or states inside and outside the monitoring station. The risk prediction module compares historical data on equipment failures, fires, and unauthorized intrusions with real-time collected data to predict potential risks. The semantic understanding module uses natural language processing to parse operator instructions and logs to identify potential problems, reducing the operational difficulty for technicians. The multimodal fusion module integrates and analyzes multimodal data such as video, sensor data, and text to form a comprehensive risk assessment. The multimodal fusion module employs dynamic weighting and utilizes an attention mechanism. The system dynamically assigns weights based on the importance of each modality. This design allows staff to adjust risk assessment standards according to the current time period, work status, or external factors, making equipment risk assessment more flexible. Furthermore, the dynamic weighting of the multimodal fusion module can automatically adjust weights based on multiple levels and data types. When the system adjusts weights at the equipment level, the weight of critical equipment (such as the main transformer) can be higher than that of secondary equipment (such as auxiliary equipment). Under extreme weather conditions such as high temperature, high humidity, and thunderstorms, the weight of environmental factors will be increased. The weight of environmental data can be obtained through edge device networking or adjusted by acquiring historical environmental data from logs. The dynamic weighting of the multimodal fusion module... The system can also utilize regional layer weights for weight adjustment. During regional layer weight adjustment, the system can incorporate regional risk management. When multiple stations are located in the same area, the weight of each station can be adjusted based on regional risks (such as typhoons and floods). When a station is located in a high-risk area such as blizzards, high temperatures, floods, or thunderstorms, its weight will automatically increase during these risk periods. Simultaneously, the multimodal fusion module can also consider dynamic changes in the time dimension using time-layer weights. Time-layer weights can vary according to different time periods. During nighttime, the weight of personnel activity monitoring will decrease, while the weight of automated equipment operation status will increase. During periods of high equipment load operation (such as peak summer electricity consumption), the weight of equipment status will increase. Furthermore, the system can also adjust weights based on real-time risk assessments. The system assesses performance and user feedback, dynamically learning the optimal weight distribution (automatically adjusting the weights of relevant factors when a user reports a misjudgment of an assessment as low risk). Furthermore, the multimodal fusion module supports adaptive sampling and dynamic updates, dynamically adjusting the data sampling frequency of the data acquisition module based on the risk level. The sampling frequency can be reduced when environmental parameters fluctuate little and increased when they fluctuate significantly, further enhancing the system's flexibility. During edge computing, the real-time processing module can run large AI models on edge devices, reducing data transmission latency and enabling real-time risk assessment. Meanwhile, the low-power, high-efficiency computing module uses lightweight large models, optimizing performance and power consumption, in conjunction with the data preprocessing module.This design significantly improves system workload and assessment accuracy. Furthermore, the network outage tolerance module supports offline processing, enabling basic risk assessment even during network interruptions. The edge caching module utilizes an edge caching mechanism to temporarily store data in unstable network conditions, allowing for batch uploading once the network is restored, thus preventing data loss. These design features greatly enhance system stability.

[0060] In addition, during the use of the cloud collaboration module, the model update module can regularly train and optimize large models in the cloud and distribute the updated models to edge devices, while the data synchronization and storage module uploads key data from the edge to the cloud for long-term storage and in-depth analysis. The global optimization module aggregates data from multiple stations through the cloud to identify global risks. The use of the cloud collaboration module can greatly improve the system's scalability and upgrades.

[0061] When the risk assessment module is working, the risk rating module can classify the risk into three levels—low, medium, and high—based on the calculation results of the large model reasoning module, according to the severity of the risk. After obtaining the risk level, the visualization module can display the risk status of the station through intuitive methods such as charts and heat maps. After the risk assessment and calculation are completed, the alarm mechanism module can issue an alarm. The alarm mechanism module supports multiple alarm forms such as sound, SMS, and APP push, and provides a detailed risk analysis report. The emergency plan recommendation module recommends corresponding emergency handling plans based on the risk type, which greatly facilitates engineers in subsequent risk handling.

[0062] The user interaction module's visual interface supports multiple platforms, including PCs, mobile devices, and tablets, making it convenient for users and engineers to view risk assessment results. The voice interaction module allows users to query the station's risk status via voice, which greatly facilitates engineers' risk assessment during risk management. The report generation module periodically generates station safety risk reports and supports exporting to PDF and Excel, which greatly facilitates subsequent log generation, processing, and summary processes.

[0063] The embodiments of the present invention are given for illustrative and descriptive purposes only, and are not intended to be exhaustive or to limit the invention to the forms disclosed. Many modifications and variations will be apparent to those skilled in the art. The embodiments were chosen and described in order to better illustrate the principles and practical application of the invention, and to enable those skilled in the art to understand the invention and to design various embodiments with various modifications suitable for a particular purpose.

Claims

1. An online assessment system for station safety risks based on edge AI large-scale models, characterized in that, It includes a data acquisition module, a large model inference module, an edge computing module, a cloud collaboration module, a risk assessment module, and a user interaction module; The output of the data acquisition module is connected to the large model inference module, and the output of the large model inference module is connected to the edge computing module, the cloud collaboration module, and the risk assessment module. The output of the risk assessment module is connected to the user interaction module. The data acquisition module is used to collect data on temperature, humidity, air pressure, and smoke concentration in the station building, as well as temperature, current, voltage, and vibration data of electrical equipment. It also collects video streams inside and outside the station building, personnel entry and exit records, movement trajectories, and past risk events in the station building, and processes the collected data. The large model inference module is used to analyze sensor data and video streams in real time, detect abnormal behavior or abnormal state, and predict potential risks based on historical data of equipment failure, fire, and illegal intrusion and real-time collected data. At the same time, it integrates and analyzes multimodal data such as video, sensor, and text to form a comprehensive risk assessment. The edge computing module is used to run large AI models on edge devices, reduce data transmission latency, achieve real-time risk assessment, optimize performance and power consumption, and still perform basic risk assessment when the network is interrupted. It also utilizes an edge caching mechanism to temporarily store data in the event of network instability and upload it in batches after the network is restored. The cloud collaboration module is used to periodically train and optimize large models and distribute the updated models to edge devices. The risk assessment module is used to predict potential risks based on historical and real-time data of equipment failures, fires, and illegal intrusions. The user interaction module is used by users to view risk assessment results and query station risk status, and periodically generate station safety risk reports; The data acquisition module includes an environmental data acquisition module, an equipment operation status acquisition module, a video surveillance data acquisition module, a personnel activity data acquisition module, a historical data acquisition module, and a data preprocessing module. The data preprocessing module is used to fill or remove outliers in the data acquired by the data acquisition module, reduce noise in the data, align the time, and unify multimodal data to the same time dimension. In addition, the data preprocessing module uses keyframe extraction and ROI detection to transmit only risk areas or abnormal frames. The data preprocessing module evaluates the importance of the data in real time through a lightweight large model on the edge device and uploads important data to the cloud. The large model inference module includes an anomaly detection module, a risk prediction module, a semantic understanding module, and a multimodal fusion module. The multimodal fusion module adopts dynamic weighting and uses an attention mechanism to dynamically allocate weights according to the importance of the modality. During use, the multimodal fusion module supports adaptive sampling and dynamic updates, and dynamically adjusts the data sampling frequency of the data acquisition module according to the risk level. The sampling frequency is reduced when the environmental parameters fluctuate less and increased when the environmental parameters fluctuate more.

2. The online station safety risk assessment system based on edge AI large model according to claim 1, characterized in that, The environmental data acquisition module is used to collect temperature, humidity, air pressure, and smoke concentration data of the station building, while the equipment operation status acquisition module is used to collect temperature, current, voltage, and vibration data of the electrical equipment in the station building.

3. The online station safety risk assessment system based on an edge AI large model according to claim 1, characterized in that, The video surveillance data acquisition module is used to collect video streams inside and outside the station building, the personnel activity data acquisition module is used to collect personnel entry and exit records and movement trajectories, and the historical data acquisition module is used to collect past risk events and equipment maintenance records of the station building.

4. The online station safety risk assessment system based on edge AI large model according to claim 1, characterized in that, The anomaly detection module is used to analyze sensor data and video streams in real time to detect abnormal behavior or abnormal states. The risk prediction module predicts potential risks based on historical data of equipment failure, fire, and illegal intrusion, as well as real-time collected data.

5. The online station safety risk assessment system based on edge AI large model according to claim 1, characterized in that, The semantic understanding module uses natural language processing to parse operator instructions and logs to identify potential problems, while the multimodal fusion module integrates and analyzes multimodal data, including video, sensor, and text, to form a comprehensive risk assessment.

6. The online station safety risk assessment system based on an edge AI large model according to claim 1, characterized in that, The edge computing module includes a real-time processing module, a low-power, high-efficiency computing module, a network outage fault-tolerant module, and an edge caching module. The real-time processing module runs large AI models on edge devices, and the low-power, high-efficiency computing module uses lightweight large models.

7. The online station safety risk assessment system based on an edge AI large model according to claim 6, characterized in that, The network outage fault tolerance module supports offline processing and still performs basic risk assessment when the network is interrupted. The edge caching module uses the edge caching mechanism to temporarily store data in the event of network instability, and then uploads it in batches after the network is restored to avoid data loss.

8. The online assessment system for station safety risks based on an edge AI large model according to claim 1, characterized in that, The cloud collaboration module includes a model update module, a data synchronization and storage module, and a global optimization module. The model update module performs regular training and optimization on the large model and distributes the updated model to edge devices.

9. The online station safety risk assessment system based on edge AI large model according to claim 1, characterized in that, The risk assessment module includes a risk rating module, a visualization module, an alarm mechanism module, and an emergency plan recommendation module. The risk rating module divides risks into three levels: low, medium, and high, based on their severity. The visualization module displays the station's risk status in an intuitive way, using charts, heat maps, and other visual methods.

10. The online assessment system for station safety risks based on an edge AI large model according to claim 1, characterized in that, The user interaction module includes a visual interface module, a voice interaction module, and a report generation module. The visual interface module provides multi-platform support, the voice interaction module allows users to query the station's risk status via voice, and the report generation module periodically generates station safety risk reports, supporting export to PDF and Excel.

Images