Multi-mode security risk intelligent identification and local early warning method

Abstract

The invention provides a multi-mode safety risk intelligent identification and local early warning method, which relates to the technical field of building safety and comprises the steps of data acquisition, edge processing, risk assessment, local early warning, event uploading and feedback and the like. According to the invention, through multi-source data fusion, edge calculation and a low-power-consumption field alarm mechanism, active identification and local instant early warning of risk events can be realized.

Description

Technical Field

[0001] The present invention relates to the field of building safety technology, and in particular to a multimodal safety risk intelligent identification and local early warning method. Background Art

[0002] Construction sites like high-rise buildings and municipal infrastructure are complex and densely populated, often accompanied by safety hazards such as falls from heights, electric shocks, poisoning, and fires. Currently, widely used security monitoring methods include video surveillance, facial recognition, and environmental monitoring. However, most rely on centralized processing, resulting in long response delays, weak on-site linkage, and high energy consumption. These methods struggle to meet the demand for fast, low-power, and highly reliable localized early warning systems.

[0003] At the same time, existing systems generally rely on a single data source for information fusion and risk identification, resulting in limited accuracy and a lack of flexible deployment and model adaptability, making them difficult to operate stably under complex working conditions. Therefore, a new construction site-specific methodology with multimodal fusion capabilities, edge intelligent processing, and low-power early warning capabilities is urgently needed to enable rapid identification and on-site response to typical construction risks. Summary of the Invention

[0004] In response to the above-mentioned existing technologies, the present invention proposes a multimodal safety risk intelligent identification and local early warning method for building and municipal construction scenarios. Through multi-source data fusion, edge computing and low-power on-site alarm mechanism, active identification of risk events and local immediate early warning are achieved.

[0005] The present invention provides a multimodal security risk intelligent identification and local early warning method, comprising the following steps: S1. Data Collection: Multimodal sensors deployed on-site collect data on personnel behavior, working environment, equipment status, etc., including but not limited to images, temperature, smoke, gas, current, inclination, and other information; S2, Edge Processing: On-site edge computing nodes pre-process the raw data, extract features, and perform preliminary event recognition to form a structured feature stream; S3. Risk assessment: Utilize multi-source fusion algorithms to jointly analyze collected data to determine whether there are any abnormal conditions. If the abnormal conditions exceed the set threshold, it is determined to be a potential safety risk. S4. Local warning: Once a risk event is identified, the low-power terminal warning device will immediately trigger an on-site sound and light alarm or visual signal to remind relevant personnel to avoid danger; S5. Event upload and feedback: Upload risk event-related data and judgment results to the cloud platform wirelessly, and continuously optimize local model parameters based on remote feedback.

[0006] Preferably, the multimodal sensor includes an infrared thermal imager, a CO / CH4 gas sensor, a current transformer, a micro-vibration sensor, a camera and an environmental noise sensor.

[0007] Preferably, the edge computing node integrates an AI processing chip and has TinyML capabilities, can locally complete the operation of the dangerous behavior recognition model, and supports offline reasoning.

[0008] Preferably, the risk assessment adopts a multi-dimensional feature weighted fusion mechanism, including strategies such as time series analysis, spatial thermal distribution and personnel trajectory reconstruction.

[0009] Preferably, the early warning device is an independently powered low-power unit with sound and light alarm functions, and the alarm response time is less than 200ms.

[0010] Preferably, in the risk assessment, the risk level is divided into three levels according to the severity of the event: minor anomaly, dangerous behavior, and emergency accident, corresponding to different alarm priorities and prompt methods.

[0011] Preferably, the system supports model self-learning and dynamic parameter update mechanisms during event upload and feedback. Cloud-based feedback results are used to modify edge recognition models to adapt to new operating conditions. Furthermore, the system can adaptively adjust recognition model strategies based on risk type. For example, for "height work," it prioritizes identifying edge displacement and failure to wear a safety belt; for "electrical fire," it prioritizes analyzing overload characteristics and temperature rise gradients.

[0012] Preferably, a time window accumulation mechanism is introduced into the local early warning to reduce frequent alarms caused by false alarms and interference events.

[0013] Preferably, the early warning device has an ultra-low power consumption wake-up mechanism, which is in a deep sleep mode when not in operation, and automatically wakes up only after a high-risk determination or a timed heartbeat interruption, thereby increasing the device's battery life to more than 6 months.

[0014] Preferably, in the local warning, if multiple warning devices in the same area are triggered at the same time, the system automatically enters the "cooperative alarm" mode, expands the alarm range and generates a compound risk event identifier on the back end, which is uploaded first and remotely scheduled for intervention.

[0015] Compared with the prior art, the present invention has the following advantages: 1. Multimodal fusion recognition: By integrating multi-source information such as images, gas, electricity, and vibration, the recognition accuracy of various types of safety events (such as fire, collapse, fall, electric shock, etc.) is improved.

[0016] 2. Edge intelligent local processing: Use lightweight AI models to run locally on edge nodes to avoid data upload delays and bandwidth consumption, and meet real-time requirements.

[0017] 3. Ultra-low power warning mechanism: The warning device adopts a wake-up working mechanism and has the ability to link sound and light on-site, realizing "local recognition and local response" and has a long battery life.

[0018] 4. Model self-learning mechanism: Back-end recognition feedback results are used to update local model parameters, improve adaptability, and reduce the need for human intervention.

[0019] 5. Collaborative alarm linkage mechanism: The collaboration of multiple devices can integrate and identify local group risk events, improving system stability and overall response efficiency.

[0020] 6. The present invention is particularly suitable for high-risk operating environments and has good practical value and promotion prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 This is a flow chart of data processing and early warning triggering in an embodiment of the present invention.

[0022] Figure 2 Schematic diagram of the structure of the system in an embodiment of the present invention. DETAILED DESCRIPTION

[0023] In order to make the technical means, creative features, objectives and effects achieved by the present invention easier to understand, the present invention is further described below in conjunction with specific embodiments.

[0024] Example: Figure 1 、 2 The illustrated multimodal intelligent identification and local early warning method for safety risks combines artificial intelligence and IoT technologies. Through the fusion of multiple on-site sensors and local AI inference, it achieves real-time identification of multiple types of hidden dangers at construction sites and provides low-power terminal early warnings. This system uses a combination of sensors, including cameras, vibration / pressure sensors, tilt sensors, and gas / liquid leakage sensors, to collect information on construction worker behavior, machine status, and environmental parameters. The system then runs behavior recognition and status analysis algorithms on low-power terminals on-site.

[0025] The specific steps include: S1. Multimodal data acquisition and preprocessing: Cameras deployed in key areas continuously collect on-site video images for identifying personnel, machinery, and environmental characteristics. At the same time, various IoT sensors (such as infrared thermal imagers, CO / CH4 gas sensors, current transformers, micro-vibration sensors, environmental noise sensors, load sensors, displacement sensors, etc.) collect corresponding status data. After image enhancement and resolution adjustment, video data is input into the local visual AI model. Sensor data is preprocessed by filtering, normalization, and other methods before being input into the local analysis module.

[0026] S2. Behavior recognition and status detection algorithms: Deploy lightweight neural network models (such as optimized YOLOv8 / CNN models) on-site terminals to perform personnel behavior and equipment detection, such as detecting whether workers are wearing safety belts, helmets, reflective clothing, and whether they have entered dangerous areas (virtual safety fences). Analyze equipment information captured by cameras and sensors, such as using visual recognition and sensor measurements to determine whether tower cranes, excavators and other lifting equipment are overloaded and whether the operating radius is legal. Identify whether guardrails are missing in high-altitude working areas and detect whether cables have oil stains or abnormal humidity (in combination with gas / liquid leakage sensors). Each recognition module outputs a signal indicating whether the corresponding risk has been detected.

[0027] S3. Risk fusion assessment and classification: This embodiment proposes mapping multi-source detection results to a unified risk indicator system, assigning weights to various types of behavioral and state hazards; the system fuses the image recognition results with the sensor signals at the terminal through a rule engine or machine learning model to generate a comprehensive risk value; based on preset thresholds and scenario models (such as the risk factor for working at height, the electrical risk factor, etc.), the severity of the identified hazards is assessed, and the risks are classified into three levels: general (minor anomalies), high (dangerous behavior), and severe (emergency accidents).

[0028] S4. Early warning triggering logic: When the risk value of a certain hidden danger exceeds the threshold, the local terminal on site will immediately trigger an early warning: the terminal uses built-in sound and light alarms, vibrators, etc. to provide intuitive prompts, without relying on external network push. For example, if a worker is detected not wearing a safety belt and is in a high-altitude working area, a high-level sound and light alarm can be immediately issued; if the machinery is found to be overloaded, the local warning light and vibration alarm in the machinery area will be triggered; the early warning information is also recorded in the local log, and the status summary can be pushed to the on-site management screen through short-range communication (such as local Wi-Fi, Bluetooth or near-field communication); the entire early warning process does not need to be uploaded to the cloud, and the response is fast; The early warning device is an independently powered, low-power unit with audible and visual alarm functions, and an alarm response time of less than 200ms. A time window accumulation mechanism is introduced into the local early warning to reduce frequent alarms caused by false alarms and interference events. The early warning device has an ultra-low power wake-up mechanism. It is in deep sleep mode when not in operation and automatically wakes up only after a high-risk determination or a timed heartbeat interruption, thereby extending the device's battery life to more than 6 months. If multiple early warning devices in the same area are triggered simultaneously, the system automatically enters "cooperative alarm" mode, expanding the alarm range and generating a composite risk event identifier on the back end, which is uploaded first and remotely dispatched for intervention. S5. Event Upload and Feedback: Risk event-related data and judgment results are wirelessly uploaded to the cloud platform, while local model parameters are continuously optimized based on remote feedback. The system supports model self-learning and dynamic parameter update mechanisms, and cloud-based feedback is used to modify the edge recognition model to adapt to new working conditions. Furthermore, the system can adaptively adjust the recognition model strategy based on risk type. For example, for "height work," it prioritizes edge displacement and failure to wear a safety belt; for "electrical fire," it prioritizes overload characteristics and temperature rise gradients.

[0029] In this embodiment, the system is divided into a sensing and data collection layer, an edge computing and analysis layer, and an end-point alarm layer. The sensing layer consists of cameras and various sensors. The edge computing layer includes an embedded computing module with an integrated low-power AI chip, along with locally deployed behavior recognition and status detection models. The end-point alarm layer integrates an acoustic, optical, and vibration output device. This architecture can adapt to different sensor combinations and supports multi-node collaboration, with each node capable of independent operation. Deployment costs are low, and flexible deployment can be tailored to actual construction scenarios (for example, video surveillance can be deployed at high-altitude work sites and hazardous areas, gas sensors can be deployed in cable wells, and load sensors can be installed on heavy machinery). To adapt to power-scarce and off-grid environments at construction sites, edge devices utilize embedded AI chips or microcontroller platforms supporting the TinyML framework. For example, these platforms utilize an ARM Cortex-M series MCU with an ultra-lightweight neural network model, enabling object detection and rule recognition in just tens of KB. The local inference process is performed with extremely low power consumption, with an average single-node power consumption of less than 50mW. This allows for long-term independent operation with micro-power solutions such as batteries or solar power, adapting to unmanned and decentralized site deployment requirements.

[0030] The above are merely embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent solutions made using the contents of the present invention specification, directly or indirectly applied in other related technical fields, are also within the patent protection scope of the present invention.

Claims

1. A multimodal security risk intelligent identification and local early warning method, characterized in that: The following steps are involved: S1. Data Collection: Multimodal sensors deployed on-site collect data on personnel behavior, working environment, equipment status, etc., including but not limited to images, temperature, smoke, gas, current, inclination, and other information; S2, Edge Processing: On-site edge computing nodes pre-process the raw data, extract features, and perform preliminary event recognition to form a structured feature stream; S3. Risk assessment: Utilize multi-source fusion algorithms to jointly analyze collected data to determine whether there are any abnormal conditions. If the abnormal conditions exceed the set threshold, it is determined to be a potential safety risk. S4. Local warning: Once a risk event is identified, the low-power terminal warning device will immediately trigger an on-site sound and light alarm or visual signal to remind relevant personnel to avoid danger; S5. Event upload and feedback: Upload risk event-related data and judgment results to the cloud platform wirelessly, and continuously optimize local model parameters based on remote feedback.

2. The multimodal security risk intelligent identification and local early warning method according to claim 1, characterized in that: The multimodal sensor includes an infrared thermal imager, a CO / CH4 gas sensor, a current transformer, a micro-vibration sensor, a camera and an environmental noise sensor.

3. The multimodal security risk intelligent identification and local early warning method according to claim 1 or 2, characterized in that: The edge computing node integrates an AI processing chip and has TinyML capabilities. It can locally complete the operation of the dangerous behavior recognition model and support offline reasoning.

4. The multimodal security risk intelligent identification and local early warning method according to claim 1 or 2, characterized in that: The risk assessment adopts a multi-dimensional feature weighted fusion mechanism, including strategies such as time series analysis, spatial thermal distribution and personnel trajectory reconstruction.

5. The multimodal security risk intelligent identification and local early warning method according to claim 1 or 2, characterized in that: The early warning device is an independently powered low-power unit with sound and light alarm functions, and the alarm response time is less than 200ms.

6. The multimodal security risk intelligent identification and local early warning method according to claim 1 or 2, characterized in that: In the risk assessment, the risk level is divided into three levels according to the severity of the event: minor anomalies, dangerous behaviors, and emergencies, corresponding to different alarm priorities and prompt methods.

7. The multimodal security risk intelligent identification and local early warning method according to claim 1 or 2, characterized in that: In the event uploading and feedback, the system supports model self-learning and dynamic parameter update mechanism, and the cloud feedback results are used to correct the edge recognition model to adapt to new working conditions.

8. The multimodal security risk intelligent identification and local early warning method according to claim 1 or 2, characterized in that: A time window accumulation mechanism is introduced into the local early warning to reduce frequent alarms caused by false alarms and interference events.

9. The multimodal security risk intelligent identification and local early warning method according to claim 1 or 2, characterized in that: The early warning device has an ultra-low power consumption wake-up mechanism, which is in a deep sleep mode when not in operation, and automatically wakes up only after a high-risk determination or a timed heartbeat interruption.

10. The multimodal security risk intelligent identification and local early warning method according to claim 1 or 2, characterized in that: In the local warning, if multiple warning devices in the same area are triggered at the same time, the system automatically enters the "cooperative alarm" mode, expands the alarm range and generates a compound risk event identifier on the back end, which is uploaded first and remotely scheduled for intervention.

Images