Security element intelligent early warning method based on hybrid model

Abstract

The invention relates to the technical field of production safety management and artificial intelligence, and discloses a safety factor intelligent early warning method based on a hybrid model. The method specifically comprises the following steps: S1, constructing a multi-modal data acquisition layer, an edge reasoning and local early warning layer, a cloud large model training and depth reasoning layer, a visualization and decision-making auxiliary layer and edge equipment; S2, acquiring multi-modal data through the multi-modal data acquisition layer in the S1, and performing time synchronization, space calibration, denoising and formatting processing. According to the invention, a multi-modal data acquisition layer is constructed, an NTP or Beidou clock module is adopted to realize millisecond-level time synchronization, a space definition label is utilized to perform unified compilation on sensing data in the same space, space calibration is completed, and meanwhile, targeted denoising and formatting processing is performed on data such as images, sensors, audios, texts and the like. The problem that the multi-modal data protocol and format are not uniform is solved.

Description

Technical Field

[0001] The present invention relates to the fields of production safety management and artificial intelligence technology, and in particular to an intelligent early warning method for safety factors based on a hybrid model. Background Art

[0002] In high-risk industries such as mining, electricity, and petrochemicals, production safety monitoring requires the simultaneous acquisition of multiple types of information. These multimodal data come from sensors and systems from different manufacturers, and the protocols and formats are not unified. There is often a "data island" phenomenon, making it difficult to achieve deep cross-modal and cross-system integration. At the same time, there is no unified modeling and analysis mechanism for the causal relationship between multiple factors such as "personnel-equipment-environment-process-regulations".

[0003] In existing technologies, multimodal production safety factor data (such as visible light, infrared, gas concentration, vibration / sound, environmental parameters, and MES / PLC logs) are often collected by sensors and systems from different manufacturers. The protocols and formats are not unified, which makes it difficult to align the time series and calibrate the space of the data. It is impossible to form a joint correlation analysis of the multiple factors of "personnel-equipment-environment-process-regulation", which is prone to omissions or false alarms. In addition, although the lightweight small models deployed at the edge can achieve low-latency local inference, due to the limited coverage of training samples and small model capacity, they are not suitable for new environments or rare risks ( Such as chemical leaks and rare failure modes) have poor adaptability and lack the ability to judge the complex risks after the interaction of multiple factors; although the large cloud model has the advantages of few-sample / zero-sample learning capabilities and deep causal chain reasoning, it is limited by network bandwidth and latency, and cannot guarantee real-time warning at critical moments; sinking the complete large model to the edge will lead to computing power bottlenecks, high storage costs, inconvenient updates and maintenance, etc.; there is also a lack of a collaborative switching mechanism that can call the cloud large model when the network is unobstructed, fall back to the edge small model when the network is restricted, and dynamically update the edge model with the cloud iteration results through OTA. Summary of the Invention

[0004] In order to make up for the above deficiencies, the present invention provides a safety factor intelligent early warning method based on a hybrid model, which aims to improve the problem of " " mentioned in the prior art.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a hybrid model-based intelligent early warning method for safety factors, specifically comprising the following steps:

[0006] S1: Build a multimodal data acquisition layer, an edge reasoning and local warning layer, a cloud-based large model training and deep reasoning layer, a visualization and decision support layer, and edge devices;

[0007] S2: Collect multimodal data through the multimodal data acquisition layer in S1, and perform time synchronization, spatial calibration, denoising and formatting;

[0008] S3: Uses lightweight models through the edge inference and local warning layer in S1 to perform local inference on the data processed in S2, outputs a real-time risk score, issues local alerts based on the score and local threshold library, and uploads data fragments and inference results to the cloud when a high risk level is reached.

[0009] S4: Receives data uploaded from S3 through the cloud-based large model training and deep inference layer in S1, performs deep inference using the multimodal large model, outputs risk measurement vectors and treatment recommendations, and pushes the inference results to edge devices in the form of incremental small model patches.

[0010] S5: The edge devices in S4 switch between different working modes based on network quality and computing load, and automatically retransmit locally cached data after the network is restored;

[0011] S6: The large cloud model output is visualized through the visualization and decision-making support layer in S1, and reports and suggestions are generated and pushed to relevant personnel.

[0012] As a further description of the above technical solution:

[0013] The multimodal data in S2 includes visible light data, infrared data, gas concentration data, temperature and humidity data, dust data, vibration data, sound data, RFID / Beidou positioning data and MES / PLC text log data. The time synchronization adopts NTP or Beidou clock module to achieve millisecond-level synchronization. The spatial calibration unifies the sensor data in the same space through spatial definition tags to achieve pixel alignment between synchronous frames. The denoising and formatting processing includes denoising convolution or filtering and resolution unification of the image, sliding window filtering and linear / high-order interpolation of the sensor data, short-time Fourier transform of the audio to generate a Mel spectrum diagram, and sentence segmentation, word segmentation and keyword extraction of the text.

[0014] As a further description of the above technical solution:

[0015] In the edge inference and local warning layer, the lightweight small model includes a visual small model, a sensor timing small model, an audio small model, a text small model and a fusion small model. The visual small model is based on MobileNetV3 or pruned ResNet-18, and is used for preliminary anomaly detection of visible light and infrared fusion frames. The sensor timing small model adopts 1D-CNN+unidirectional LSTM for fast anomaly recognition of time series data. The audio small model converts the on-site sound into a Mel spectrum graph and inputs it into a lightweight CNN for abnormal sound detection. The text small model is based on Disti lBERT and performs keyword filtering and preliminary event classification on MES / PLC logs. The fusion small model splices the feature vectors output by each branch and inputs it into a layer of lightweight TransformerEncoder to output a real-time risk score.

[0016] As a further description of the above technical solution:

[0017] In the edge inference and local warning layer, edge model dynamic compression technology is introduced to adjust the structure and parameters of the small model in real time according to the computing power resources of the edge device. Edge small model adaptive threshold adjustment technology is used to dynamically adjust the alarm threshold according to historical data and current environmental changes.

[0018] As a further description of the above technical solution:

[0019] In the cloud-based large model training and deep reasoning layer, the multimodal large model includes a visual branch, a sensor branch, an audio branch, a text branch, a cross-modal fusion layer and a causal association reasoning layer. The visual branch uses a dual-stream ResNet-50 or ViT for deep feature extraction of visible light and infrared images. The sensor branch uses 1D-CNN+bidirectional LSTM to capture temporal patterns. The audio branch is based on a CNN+Transformer network for complex sound recognition. The text branch is based on a BERT-base or larger pre-trained model to parse potential hidden danger information in MES / PLC logs. The cross-modal fusion layer uses multi-head attention to fuse the feature vectors of each branch into a unified representation. The causal association reasoning layer is based on a knowledge graph and uses GAT for deep causal chain reasoning to output a high-precision risk measurement vector.

[0020] As a further description of the above technical solution:

[0021] In the cloud-based large model training and deep inference layer, cloud-based model active learning technology is added to actively select key samples for training based on the uncertainty of edge feedback data. During the large model training process, federated learning technology is used to combine data from multiple companies for training.

[0022] As a further description of the above technical solution:

[0023] The edge device switches between different working modes according to network quality and computing power load: when the network is stable and computing power is idle, it enters the "joint mode" and dynamically adjusts the fused small model based on the feedback from the large model; when the network is unstable or computing power is tight, it automatically returns to the "all-small model mode"; if the network is unavailable, it switches to the "small model full-stack mode", transmits a brief alarm summary through emergency communication, and automatically retransmits the complete multi-modal original data cached locally to the cloud after the network is restored. Intelligent predictive switching technology is introduced to predict network changes based on historical network data and current status, and perform model mode switching in advance.

[0024] As a further description of the above technical solution:

[0025] In the visualization and decision-making support layer, based on the output of the cloud-based large model, combined with GIS / 3D maps, risk heat maps are rendered in real time, and the evolution of the causal chain is displayed in animated form. On-site accident reports, risk assessment reports and disposal suggestions are automatically generated and pushed to relevant personnel through channels including but not limited to mobile apps, DingTalk / SMS. In addition, augmented reality technology is introduced to display the real-time risk status and potential hazards of on-site equipment in a visual interface.

[0026] As a further description of the above technical solution:

[0027] In S2, a data repair technology based on deep learning is used to repair severely missing data. In the multimodal data acquisition layer, an adaptive sampling frequency adjustment technology is newly added to dynamically adjust the sampling frequency of the sensor according to environmental changes.

[0028] As a further description of the above technical solution:

[0029] A hybrid model-based intelligent early warning system for safety factors, comprising a multimodal data acquisition layer, an edge reasoning and local early warning layer, a cloud-based large model training and deep reasoning layer, an edge device, and a visualization and decision-making assistance layer, wherein the multimodal data acquisition layer is used to execute step S1 of claim 1, the edge reasoning and local early warning layer is used to execute step S2 of claim 1, the cloud-based large model training and deep reasoning layer is used to execute step S3 of claim 1, the edge device is used to execute step S4 of claim 1, and the visualization and decision-making assistance layer is used to execute step S5 of claim 1.

[0030] The present invention has the following beneficial effects:

[0031] 1. In the present invention, by constructing a multimodal data acquisition layer, using NTP or Beidou clock modules to achieve millisecond-level time synchronization, using space definition tags to uniformly compile sensor data in the same space, completing spatial calibration, and simultaneously performing targeted denoising and formatting on image, sensor, audio, text and other data, the problem of inconsistent multimodal data protocols and formats is solved, the "data island" phenomenon is eliminated, and deep cross-modal and cross-system integration is achieved. It can effectively form a joint correlation analysis between multiple factors of "personnel-equipment-environment-process-regulations", avoiding omissions or false alarms caused by the inability to integrate data.

[0032] 2. This invention utilizes a hybrid model combining edge reasoning with a local warning layer and cloud-based large-model training with a deep reasoning layer. The edge layer deploys lightweight small models for vision, sensor timing, audio, text, and fusion to achieve low-latency local reasoning. The cloud-based large model performs deep reasoning using a multimodal large model, outputs risk metric vectors and treatment recommendations, and pushes optimized incremental small model patches to edge devices. Simultaneously, the cloud-based model optimizes using active learning and federated learning techniques, addressing the issues of lightweight small models' poor adaptability to new environments or rare risks and their lack of ability to judge complex risks. This also avoids the cloud-based large model's inability to provide real-time warnings due to network limitations.

[0033] 2. The present invention supports multiple communication modes and can adapt to various extreme environments such as underground, outdoor, and chemical plants. The edge device can intelligently switch between "joint mode", "full small model mode" and "small model full stack mode" according to the network quality and computing power load, and introduces intelligent prediction switching technology to predict network changes based on historical network data and current status, and switch model modes in advance. After the network is restored, the local cached data is automatically re-transmitted, and the edge model can be dynamically updated with the cloud iteration results through OTA. It solves the problem of lack of network smoothness and model collaborative switching under restricted conditions in the existing technology, avoids the computing power, storage and maintenance problems brought about by sinking the complete large model to the edge, and ensures that the system can run stably and provide effective early warning under different network and computing power conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 Schematic diagram of the overall logical architecture of the present invention. DETAILED DESCRIPTION

[0035] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0036] Reference Figure 1 The present invention provides an embodiment of a hybrid model-based intelligent early warning method for safety factors, which specifically includes the following steps:

[0037] S1: Build a multimodal data acquisition layer, an edge reasoning and local warning layer, a cloud-based large model training and deep reasoning layer, a visualization and decision support layer, and edge devices;

[0038] S2: Multimodal data is collected through the multimodal data acquisition layer in S1, and time synchronization, spatial calibration, denoising and formatting are performed. Deep learning-based data repair technology is used to repair severely missing data. Specifically, multimodal data includes visible light data, infrared data, gas concentration data, temperature and humidity data, dust data, vibration data, sound data, RFID / Beidou positioning data, and MES / PLC text log data. Time synchronization uses NTP or Beidou clock module to achieve millisecond-level synchronization. Spatial calibration uses spatial definition tags to unify the sensor data in the same space to achieve pixel alignment between synchronous frames. Denoising and formatting processing includes denoising convolution or filtering and resolution unification of images, sliding window filtering and linear / high-order interpolation of sensor data, short-time Fourier transform of audio to generate Mel spectrum, and sentence segmentation, word segmentation and keyword extraction of text. Furthermore, in the multimodal data acquisition layer, an adaptive sampling frequency adjustment technology is added to dynamically adjust the sampling frequency of the sensor according to environmental changes;

[0039] S3: Through the edge inference and local warning layer in S1, a lightweight small model is used to perform local inference on the data processed in S2, output a real-time risk score, issue a local alarm based on the score and the local threshold library, and upload data fragments and inference results to the cloud when a higher risk level is reached. In the edge inference and local warning layer, the lightweight small models include a visual small model, a sensor time series small model, an audio small model, a text small model, and a fusion small model. The visual small model is based on MobileNetV3 or pruned ResNet-18 and is used for preliminary anomaly detection of visible light and infrared fusion frames. The sensor time series small model uses 1D-CNN+unidirectional LSTM for fast anomaly recognition of time series data. The audio small model converts the on-site sound into a Mel spectrum graph and inputs it into a lightweight CNN for abnormal sound detection. The text small model is based on DistilBERT and performs keyword filtering and preliminary event classification on MES / PLC logs. The fusion small model splices the feature vectors output by each branch and inputs them into a layer of lightweight TransformerEncoder to output a real-time risk score.

[0040] Furthermore, an example is given in the local threshold library for local alarms:

[0041] Green safety: risk score <0.3;

[0042] Yellow warning: 0.3≤score<0.6;

[0043] Orange alert: 0.6≤score<0.8;

[0044] Red danger: risk score ≥ 0.8;

[0045] When the score reaches yellow warning or above, the edge device will issue a prompt through a local sound and light alarm (buzzer + sound and light), and at the same time:

[0046] Yellow warning: The warning will only flash on the local interface and vibrate on the wearable device.

[0047] Orange alert: Immediately switch the sound and light to high-speed flashing, and upload the visible light frames, infrared frames, sensor raw data, audio clips, and log clips for several seconds / minutes before and after the current moment to the cloud via 5G or LoRa in the form of data packets;

[0048] Red danger: Immediately activate the local highest intensity sound and light alarm, and simultaneously send an alarm SMS / App push to the on-site responsible person, while also fully reporting the multi-modal raw data.

[0049] Further description, the edge model dynamic compression technology is introduced in the edge inference and local warning layer, and the structure and parameters of the small model are adjusted in real time according to the computing power resources of the edge device. The edge small model adaptive threshold adjustment technology is used to dynamically adjust the alarm threshold according to historical data and current environmental changes.

[0050] S4: Receive the data uploaded in S3 through the cloud-based large model training and deep reasoning layer in S1, perform deep reasoning through the multimodal large model, output risk measurement vectors and disposal suggestions, and push the reasoning results to the edge device in the form of incremental small model patches. Specifically, in the cloud-based large model training and deep reasoning layer, the multimodal large model includes a visual branch, a sensor branch, an audio branch, a text branch, a cross-modal fusion layer, and a causal association reasoning layer. The visual branch uses a dual-stream ResNet-50 or ViT for deep feature extraction of visible light and infrared images. The sensor branch uses 1D-CNN (three layers) + bidirectional LSTM to capture temporal patterns. The audio branch is based on CNN+ The Transformer network is used for complex sound recognition. The text branch is based on BERT-base or a larger pre-trained model to parse potential hidden danger information in MES / PLC logs. The cross-modal fusion layer uses multi-head attention to fuse the feature vectors of each branch into a unified representation. The causal association reasoning layer is based on the knowledge graph and uses GAT for deep causal chain reasoning to output a high-precision risk measurement vector. Specifically, after receiving edge uploaded data, the visual branch extracts deep features such as cracks on the tank surface and temperature distribution; the sensor branch analyzes the temporal correlation of gas concentration, pressure, and temperature; the audio branch identifies the frequency characteristics of gas leakage sound; and the text branch parses the conflicts between equipment maintenance records and operating instructions in the log. The cross-modal fusion layer uses a multi-head attention mechanism to fuse the 512-dimensional feature vectors of each branch into a unified representation. The causal inference layer, based on a constructed petrochemical industry knowledge graph (including equipment maintenance procedures, safety regulations, and historical accident cases), uses GAT to infer the causal chain of "valve seal aging → gas leakage → explosion upon encountering a heat source." The calculated risk metric vector shows an explosion probability of 0.85 and outputs the following response: "Immediately activate the emergency shut-off valve → evacuate personnel within a 500-meter radius → activate the fire sprinkler system." Simultaneously, the cloud creates an incremental patch for the optimized convolutional layer parameters of the small visual model and pushes it to the edge device via a secure channel.

[0051] Further description: In the cloud-based large model training and deep inference layer, cloud-based model active learning technology is added to actively select key samples for training based on the uncertainty of edge feedback data. During the large model training process, federated learning technology is used to combine data from multiple companies for training.

[0052] S5: The edge devices in S4 switch between different working modes according to the network quality and computing power load, and automatically retransmit the locally cached data after the network is restored. Specifically, the edge devices switch between different working modes according to the network quality and computing power load: when the network is stable and the computing power is idle, they enter the "joint mode" and dynamically adjust the fused small model based on the feedback of the large model. When the network is unstable or the computing power is tight, they automatically return to the "full small model mode". If the network is unavailable, they switch to the "small model full stack mode", transmit a brief alarm summary through emergency communication, and automatically retransmit the complete multimodal original data of the local cache to the cloud after the network is restored. The edge device uses an industrial computer pre-installed with a visual model, a sensor time series model, an audio model, a text model, and a fusion model. When the infrared thermal imager detects a sudden increase in the local temperature of the tank and the visible light camera simultaneously captures an unusual reflection on the tank surface, the visual model outputs a heat source detection probability of 0.7. At the same time, the sensor time series model detects that the nearby combustible gas concentration has increased from 1% LEL to 15% LEL over a five-minute period, outputting an abnormality probability of 0.8. The audio model detects a slight hissing sound of gas leakage with a probability of 0.6. The text model analyzes the "valve adjustment abnormality" record in the MES log and determines an abnormality probability of 0.7. After reducing the dimensionality of each branch feature vector to 64, the fusion model outputs a risk score of 0.75, reaching the red alert threshold. The edge device immediately triggers an audible and visual alarm and uploads the 30 seconds of imagery, sensor data, audio clips, and log clips preceding the incident to the cloud via the 5G network. Simultaneously, a vibration alarm is sent to the on-site inspector's smart helmet.

[0053] In addition, intelligent predictive switching technology is introduced to predict network changes based on historical network data and current status, and switch model modes in advance;

[0054] S6: The visualization and decision-making support layer in S1 visualizes the cloud-based large model output, generates reports, and sends recommendations to relevant personnel. Specifically, the visualization and decision-making support layer uses the cloud-based large model output, combined with GIS / 3D maps, to render real-time risk heat maps and animated causal chain evolution. It automatically generates on-site incident reports, risk assessment reports, and response recommendations, which are then sent to relevant personnel via mobile apps, DingTalk, and SMS. Augmented reality technology is also introduced to visualize the real-time risk status and potential hazards of on-site equipment within the interface. Specifically, risky locations in the oil tank area are highlighted in red, and 3D animations illustrate gas diffusion paths and potential explosion ranges. The system automatically generates incident reports, including an event timeline, a multimodal data chain of evidence, and a causal analysis diagram. Response recommendations are then sent to relevant personnel, such as the safety director and fire chief, via WeChat. When managers view the scene wearing AR glasses, they can overlay real-time equipment risk scores, potential hazard locations, and emergency response instructions.

[0055] The present invention also provides a hybrid model-based intelligent early warning system for security factors, including a multimodal data acquisition layer, an edge reasoning and local early warning layer, a cloud-based large model training and deep reasoning layer, an edge device, and a visualization and decision-making assistance layer. The multimodal data acquisition layer is used to execute step S1, the edge reasoning and local early warning layer is used to execute step S2, the cloud-based large model training and deep reasoning layer is used to execute step S3, the edge device is used to execute step S4, and the visualization and decision-making assistance layer is used to execute step S5.

[0056] Finally, it should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A hybrid model-based intelligent early warning method for safety factors, characterized by: The specific steps include: S1: Build a multimodal data acquisition layer, an edge reasoning and local warning layer, a cloud-based large model training and deep reasoning layer, a visualization and decision support layer, and edge devices; S2: Collect multimodal data through the multimodal data acquisition layer in S1, and perform time synchronization, spatial calibration, denoising and formatting; S3: Uses lightweight models through the edge inference and local warning layer in S1 to perform local inference on the data processed in S2, outputs a real-time risk score, issues local alerts based on the score and local threshold library, and uploads data fragments and inference results to the cloud when a high risk level is reached. S4: The cloud-based large model training and deep inference layer in S1 receives the data uploaded from S3, performs deep inference using the multimodal large model, outputs risk measurement vectors and treatment recommendations, and pushes the inference results to edge devices in the form of incremental small model patches. S5: The edge devices in S4 switch between different working modes based on network quality and computing load, and automatically retransmit locally cached data after the network is restored; S6: The large cloud model output is visualized through the visualization and decision-making support layer in S1, and reports and suggestions are generated and pushed to relevant personnel.

2. The hybrid model-based intelligent early warning method for safety factors according to claim 1 is characterized by: The multimodal data in S2 includes visible light data, infrared data, gas concentration data, temperature and humidity data, dust data, vibration data, sound data, RFID / Beidou positioning data and MES / PLC text log data. The time synchronization adopts NTP or Beidou clock module to achieve millisecond-level synchronization. The spatial calibration unifies the sensor data in the same space through spatial definition tags to achieve pixel alignment between synchronous frames. The denoising and formatting processing includes denoising convolution or filtering and resolution unification of the image, sliding window filtering and linear / high-order interpolation of the sensor data, short-time Fourier transform of the audio to generate a Mel spectrum diagram, and sentence segmentation, word segmentation and keyword extraction of the text.

3. The hybrid model-based intelligent early warning method for safety factors according to claim 1 is characterized by: In the edge inference and local warning layer, the lightweight small model includes a visual small model, a sensor timing small model, an audio small model, a text small model and a fusion small model. The visual small model is based on MobileNetV3 or pruned ResNet-18, and is used for preliminary anomaly detection of visible light and infrared fusion frames. The sensor timing small model adopts 1D-CNN+unidirectional LSTM for fast anomaly recognition of time series data. The audio small model converts the on-site sound into a Mel spectrum graph and inputs it into a lightweight CNN for abnormal sound detection. The text small model is based on Disti lBERT and performs keyword filtering and preliminary event classification on MES / PLC logs. The fusion small model splices the feature vectors output by each branch and inputs it into a layer of lightweight TransformerEncoder to output a real-time risk score.

4. The hybrid model-based intelligent early warning method for safety factors according to claim 1 is characterized by: In the edge inference and local warning layer, edge model dynamic compression technology is introduced to adjust the structure and parameters of the small model in real time according to the computing power resources of the edge device. Edge small model adaptive threshold adjustment technology is used to dynamically adjust the alarm threshold according to historical data and current environmental changes.

5. The hybrid model-based intelligent early warning method for safety factors according to claim 1 is characterized by: In the cloud-based large model training and deep reasoning layer, the multimodal large model includes a visual branch, a sensor branch, an audio branch, a text branch, a cross-modal fusion layer and a causal association reasoning layer. The visual branch uses a dual-stream ResNet-50 or ViT for deep feature extraction of visible light and infrared images. The sensor branch uses 1D-CNN+bidirectional LSTM to capture temporal patterns. The audio branch is based on a CNN+Transformer network for complex sound recognition. The text branch is based on a BERT-base or larger pre-trained model to parse potential hidden danger information in MES / PLC logs. The cross-modal fusion layer uses multi-head attention to fuse the feature vectors of each branch into a unified representation. The causal association reasoning layer is based on a knowledge graph and uses GAT for deep causal chain reasoning to output a high-precision risk measurement vector.

6. The hybrid model-based intelligent early warning method for safety factors according to claim 5 is characterized by: In the cloud-based large model training and deep inference layer, cloud-based model active learning technology is added to actively select key samples for training based on the uncertainty of edge feedback data. During the large model training process, federated learning technology is used to combine data from multiple companies for training.

7. The hybrid model-based intelligent early warning method for safety factors according to claim 1 is characterized by: The edge device switches between different working modes according to network quality and computing power load: when the network is stable and computing power is idle, it enters the "joint mode" and dynamically adjusts the fused small model based on the feedback from the large model. When the network is unstable or the computing power is tight, it automatically returns to the "all-small model mode". If the network is unavailable, it switches to the "small model full-stack mode", transmits a brief alarm summary through emergency communication, and automatically retransmits the complete multi-modal original data cached locally to the cloud after the network is restored. Intelligent predictive switching technology is introduced to predict network changes based on historical network data and current status, and switch model modes in advance.

8. The hybrid model-based intelligent early warning method for safety factors according to claim 1 is characterized by: In the visualization and decision-making support layer, based on the output of the cloud-based large model, combined with GIS / 3D maps, risk heat maps are rendered in real time, and the evolution of the causal chain is displayed in animated form. On-site accident reports, risk assessment reports and disposal suggestions are automatically generated and pushed to relevant personnel through channels including but not limited to mobile apps, DingTalk / SMS. In addition, augmented reality technology is introduced to display the real-time risk status and potential hazards of on-site equipment in a visual interface.

9. The hybrid model-based intelligent early warning method for safety factors according to claim 1 is characterized by: In S2, a data repair technology based on deep learning is used to repair severely missing data. In the multimodal data acquisition layer, an adaptive sampling frequency adjustment technology is newly added to dynamically adjust the sampling frequency of the sensor according to environmental changes.

10. A hybrid model-based intelligent early warning system for safety factors, comprising the hybrid model-based intelligent early warning method for safety factors according to claims 1-9, characterized in that: It includes a multimodal data acquisition layer, an edge reasoning and local warning layer, a cloud-based large model training and deep reasoning layer, an edge device, and a visualization and decision-making assistance layer. The multimodal data acquisition layer is used to execute the S1 step described in claim 1, the edge reasoning and local warning layer is used to execute the S2 step described in claim 1, the cloud-based large model training and deep reasoning layer is used to execute the S3 step described in claim 1, the edge device is used to execute the S4 step described in claim 1, and the visualization and decision-making assistance layer is used to execute the S5 step described in claim 1.

Images