Safety production hidden danger intelligent identification system based on end-cloud collaboration and small sample learning

Abstract

The invention discloses a safety production hidden danger intelligent identification system based on end-cloud collaboration and small sample learning, and the system comprises an expert knowledge base construction module which is used for constructing an expert knowledge base through expert documents, accident cases, troubleshooting records and equipment specification unstructured data of the chemical industry; the image acquisition module is used for acquiring industrial scene images; the data management module is used for carrying out cold and hot data separation, scene classification and data preprocessing operation on the collected industrial scene image to obtain a small-scale hidden danger sample; the visual model building module is used for carrying out fine tuning on small-scale hidden danger samples through a pre-training visual feature extractor by adopting transfer learning and integrated learning technologies, providing a joint compression strategy and building a visual recognition model; and the data mining module is used for acquiring network picture data through a Web crawling algorithm, an association rule algorithm and a clustering algorithm. According to the invention, high-precision identification and intelligent auxiliary decision making of multi-scene industrial potential safety hazards are realized.

Description

Technical Field

[0001] This invention belongs to the field of interdisciplinary technology of artificial intelligence and safe production, specifically a safe production hazard intelligent identification system based on edge-cloud collaboration and few-sample learning. Background Technology

[0002] Work safety is a crucial guarantee for economic and social development and is related to the safety of people's lives and property. Currently, there are numerous and widespread safety hazards in high-risk industries in my country. Traditional hazard investigation relies on on-site verification by industry experts, which suffers from problems such as complex procedures, low efficiency, high labor costs, and incomplete coverage. Moreover, there is a shortage of qualified and capable grassroots supervisors and a lack of expert resources, making it difficult to meet the demand for comprehensive and high-precision hazard investigation.

[0003] With the development of artificial intelligence technology, the application of technologies such as deep learning and computer vision in industrial scenarios has gradually become widespread. However, the identification of safety hazards still faces many technical bottlenecks: First, data resources are scarce and of low quality. Safety hazard-related data are mostly in unstructured form, with few positive samples and incomplete scenario coverage, resulting in insufficient data support for model training. Second, the models have poor adaptability to complex industrial scenarios. The characteristics of hazards vary greatly across different industries, making it difficult for existing models to achieve high-precision identification across multiple scenarios. Third, deployment costs are high. Complex neural network models have high hardware resource requirements, and fixed terminals are difficult to cover complex industrial sites, with insufficient collaboration between edge devices and the cloud. Fourth, the challenge of small-sample learning is prominent. High-risk industry hazard scenarios are scarce, and traditional deep learning models have low accuracy and weak generalization ability in small-sample situations.

[0004] Currently, there is no system that can effectively solve the problem of high-precision identification of industrial safety hazards under conditions of multiple scenarios, small samples, and low-quality data. This makes it difficult to meet the requirements of the "Industrial Internet + Safe Production" action plan for intelligent supervision and pre-emptive prevention. There is an urgent need to develop an intelligent identification system that is adaptable to complex industrial environments, has low data dependence, and is flexible in deployment. Summary of the Invention

[0005] To address the aforementioned problems in existing technologies, this invention provides an intelligent safety production hazard identification system based on edge-cloud collaboration and few-sample learning, which achieves high-precision identification and intelligent auxiliary decision-making for industrial safety hazards in multiple scenarios.

[0006] The technical solution to achieve the above objectives is: A safety production hazard intelligent identification system based on edge-cloud collaboration and few-shot learning includes: Security expert data platform: The expert knowledge base construction module is used to build an expert knowledge base using unstructured data such as expert documents, accident cases, investigation records, and equipment manuals from the chemical industry. Image acquisition module, used to acquire images of industrial scenes; The data governance module is used to perform hot and cold data separation, scene classification, and data preprocessing on the collected industrial scene images to obtain small-scale potential hazard samples. The visual model building module is used to fine-tune a visual recognition model on a small number of potential hazard samples by using transfer learning and ensemble learning techniques through a pre-trained visual feature extractor, and proposes a joint compression strategy. Security expert application platform: The data mining module is used to obtain online image data through web crawling algorithms, association rule algorithms, and clustering algorithms; The application software module is used to drive the visual recognition model to classify scenes and identify potential hazards in images, and output image features and recognition labels. The interaction module is used to link with the expert knowledge base to realize the closed loop of expert verification of pseudo-labels and real labels, and to complete expert review and knowledge supplementation. The decision output module combines knowledge and visual recognition results to output expert recommendations and accident predictions.

[0007] Preferably, in the expert knowledge base construction module, the expert knowledge base construction process is as follows: Natural language processing technology is used to process expert documents, accident cases, troubleshooting records, and equipment manuals, including: Named entity recognition: Identifies entities such as equipment names, components, risk points, personnel, and operating environment; Relationship extraction: Obtain the hierarchical relationship between "equipment – ​​component – ​​cause of failure – type of potential hazard"; Event extraction: Identify key events in the incident chain, including but not limited to leaks, short circuits, and overheating; The above information is written into the knowledge graph in the form of triples, forming a multi-scenario industrial safety knowledge system covering the elements of "human-machine-environment"; To ensure the continuous evolution of the expert knowledge base, a knowledge supplementation mechanism of "model preprocessing + expert review" is provided: Perform automated entity recognition and preliminary structuring on the newly added data; Experts review, supplement, and correct the content through the interactive interface; After passing the knowledge consistency verification, the knowledge is written into the expert knowledge base.

[0008] Preferably, the data governance module employs a hot and cold data separation mechanism, multi-scenario classification and noise cleaning technology, sample removal and consistency verification, and implements a label quality management mechanism. The data is expanded through multi-dimensional enhancement strategies, including geometric enhancement, lighting and background simulation, random occlusion, and noise perturbation. Geometric enhancements include rotation, clipping, and scaling.

[0009] Preferably, in the visual model building module, the joint compression strategy is: Low-rank matrix technology is used to reduce parameter storage overhead, and 8-bit integer quantization and binary quantization techniques are combined to compress the model size; pruning techniques are introduced to remove redundant parameters, and the network structure and loss function of the YOLO9 object detection model are optimized.

[0010] Preferably, the security expert data platform further includes: The database is used to provide storage for image libraries, algorithm libraries, and tag libraries, supported by a large K8S cluster and cloud database, and to provide access interfaces for the image libraries, algorithm libraries, and tag libraries to the security expert application platform.

[0011] Preferably, the database consists of an image library of 10 million unsupervised images and over 500,000 positive samples, an algorithm library of a constructed visual recognition model, model parameters, and inference logic, and a tag library of labeled information.

[0012] Compared with the prior art, the beneficial effects of the present invention are: 1) The knowledge base construction mechanism of this invention automatically transforms unstructured information from expert documents, accident cases, and inspection images into structured knowledge, and realizes the dynamic evolution of knowledge through automated extraction, expert verification, and continuous completion. This mechanism significantly improves data integration efficiency, solves the problems of low efficiency and insufficient coverage of traditional manual sorting, and provides a high-quality data foundation for subsequent small sample learning and model training. 2) This invention achieves a success rate of no less than 90% in identifying potential hazards in multiple scenarios through a combination of data augmentation, few-shot learning and transfer learning strategies. Its performance reaches the level of domain experts and is suitable for the needs of multiple industries such as oil and gas, coal, and mining. It provides strong technical support for risk identification in complex industrial environments. 3) The edge-cloud collaborative architecture (security expert application platform - security expert data platform) constructed by this invention can offload most computing tasks to the cloud for execution, while the edge only runs lightweight models, thereby avoiding excessive requirements on the configuration of on-site equipment. With the help of model compression technologies such as pruning and quantization, real-time hidden danger identification can be achieved on low-cost hardware, which can significantly reduce the system construction cost and subsequent maintenance cost. 4) This invention integrates industry expert knowledge with artificial intelligence technology to form a closed loop of "machine recognition + expert verification", which not only improves the efficiency of investigation, but also realizes the accumulation and reuse of expert knowledge, forming a continuously learning and evolving intelligent hidden danger investigation system. 5) By integrating capabilities such as real-time identification, fault reasoning, and risk warning, this invention can achieve continuous monitoring and intelligent assessment of the safety status of industrial sites, providing enterprises and regulatory authorities with the system capabilities of pre-discovery, real-time response, and closed-loop handling, which is of great significance for improving the modernization of the safety production governance system and governance capabilities. Attached Figure Description

[0013] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a block diagram of the intelligent identification system for safety production hazards based on edge-cloud collaboration and few-sample learning, as described in this invention. Figure 2 This is the overall architecture diagram of the intelligent identification system for safety production hazards based on edge-cloud collaboration and few-sample learning in this invention. Detailed Implementation

[0014] 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, and 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.

[0015] like Figure 1 , 2 As shown, the intelligent identification system for safety production hazards based on edge-cloud collaboration and few-sample learning includes: a safety expert data platform and a safety expert application platform; The security expert data platform includes: expert knowledge base construction module 1, image acquisition module 2, data governance module 3, visual model building module 4, and database 5; Module 1, the Expert Knowledge Base Construction Module, is used to build an expert knowledge base using unstructured data such as expert documents, accident cases, investigation records, and equipment manuals from the chemical industry.

[0016] In this embodiment, the expert knowledge base construction process is as follows: Natural language processing technology is used to process expert documents, accident cases, troubleshooting records, and equipment manuals, including: Named entity recognition: Identifies entities such as equipment names, components, risk points, personnel, and operating environment; Relationship extraction: Obtain the hierarchical relationship between "equipment – ​​component – ​​cause of failure – type of potential hazard"; Event extraction: Identify key events in the incident chain, including but not limited to leaks, short circuits, and overheating; The above information is written into the knowledge graph in the form of triples, forming a multi-scenario industrial safety knowledge system covering the elements of "human-machine-environment"; To ensure the continuous evolution of the expert knowledge base, a knowledge supplementation mechanism of "model preprocessing + expert review" is provided: Perform automated entity recognition and preliminary structuring on the newly added data; Experts review, supplement, and correct the content through the interactive interface; After passing knowledge consistency verification, the knowledge is written into the expert knowledge base; To develop a sustainable, evolving, and explainable industrial safety knowledge system.

[0017] This invention combines knowledge reasoning with machine learning. By integrating traditional machine learning recommendation algorithms with relational reasoning, entity completion, and semantic matching strategies, it performs fault chain correlation analysis, generates hazard type prediction results and corresponding handling suggestions, and provides expert-level decision support for hazard identification models.

[0018] Image acquisition module 2 is used to acquire images of industrial scenes.

[0019] Data governance module 3 is used to perform hot and cold data separation, scene classification, and data preprocessing on the collected industrial scene images to obtain small-scale hazard samples.

[0020] In this embodiment, for more than 10 million original data entries, the present invention employs a hot and cold data separation mechanism, multi-scenario classification and noise cleaning technology, sample elimination and consistency verification, and implements a label quality management mechanism (label ratio ≥10%, positive samples ≥5%). Furthermore, the data is expanded through multi-dimensional enhancement strategies such as geometric enhancement, lighting and background simulation, random occlusion, and noise perturbation. Combined with generative adversarial algorithms (GAN) to generate positive and negative sample pairs, the data space is effectively expanded and the problem of low-quality and scarce data is alleviated. Geometric enhancements include rotation, clipping, and scaling.

[0021] The visual model building module 4 is used to fine-tune a visual recognition model on a small number of potential hazard samples by using transfer learning and ensemble learning techniques through a pre-trained visual feature extractor, and proposes a joint compression strategy.

[0022] In this embodiment, to address the problems of scarce positive samples and complex scenarios, the present invention constructs a visual recognition model system that supports small sample sizes and high generalization ability, including: Based on pre-trained visual feature extractors (such as ResNet and EfficientNet), transfer learning and ensemble learning techniques are used to fine-tune the model on a small number of potential hazard samples to improve the model's generalization ability. To enable the model to be deployed at the edge, this invention proposes a joint compression strategy: using low-rank matrix (LORA) technology to reduce parameter storage overhead, and combining quantization techniques such as 8-bit integer quantization and binary quantization to compress the model size; introducing pruning techniques to remove redundant parameters, optimizing the network structure and loss function of the YOLO9 object detection model, and developing a lightweight model adapted to edge devices.

[0023] Database 5 is used to provide storage for image libraries, algorithm libraries, and tag libraries, supported by a large K8S cluster and cloud database, and to provide access interfaces for image libraries, algorithm libraries, and tag libraries to the security expert application platform.

[0024] In this embodiment, the image library consists of 10 million unsupervised images and over 500,000 positive samples, the algorithm library consists of a constructed visual recognition model, model parameters, and inference logic, and the tag library consists of annotation information.

[0025] The security expert application platform includes: data mining module 6, application software module 7, interaction module 8, and decision output module 9.

[0026] Data mining module 6 is used to obtain network image data through web crawling algorithms, association rule algorithms, and clustering algorithms.

[0027] Application software module 7 is used to drive the visual recognition model to classify scenes and identify potential hazards in images, and output image features and recognition labels.

[0028] Interaction module 8 is used to link with the expert knowledge base to realize the expert verification closed loop between pseudo-labels and real labels, and to complete expert review and knowledge supplementation.

[0029] Decision output module 9 is used to combine knowledge and visual recognition results to output expert suggestions and accident prediction results.

[0030] The intelligent safety hazard identification system of this invention is an edge-cloud collaborative architecture suitable for industrial environments, and includes the following: Optimized architecture design: The edge device (security expert application platform) is responsible for real-time data collection, preliminary inference and rapid response, while the cloud (security expert data platform) is responsible for large-scale data storage, model training and optimization and complex decision support. Edge computing reduces data transmission latency and bandwidth consumption.

[0031] Real-time data processing and decision-making: Edge devices enable immediate perception and preliminary identification of potential risks, while the cloud performs in-depth analysis and model iteration on complex data uploaded from edge devices, forming a collaborative mechanism of "rapid response at the edge + in-depth optimization in the cloud".

[0032] Enhanced security and reliability: Encrypted communication channels are built to ensure data transmission and storage security, and an adaptive computing strategy is adopted to dynamically adjust resource allocation based on edge device resources, thereby improving system stability.

[0033] Example 1: Implementation Method of Intelligent Identification System for Pipeline Leakage Hazards in the Chemical Industry 1. Scenario Description Oil, gas, and chemical medium pipelines in chemical enterprises often suffer from hidden dangers such as dripping, seepage, and jetting. Traditional inspections rely on manual observation, which has problems such as low inspection frequency, difficulty in covering the entire pipeline area, and high rate of missed inspections.

[0034] 2. Implementation Steps (1) Knowledge base construction Collect unstructured data such as safety regulations, equipment maintenance manuals, accident cases, and pipeline operating parameters from the chemical industry.

[0035] BERT-wwm-chinese was used for entity extraction to identify core entities such as "pipeline type", "leakage characteristics" and "response measures", and a Neo4j knowledge graph (approximately 100,000 entities and 200,000 relationships) was constructed.

[0036] Newly added hidden danger cases are reviewed and supplemented through an expert interaction platform, forming a continuously updated knowledge base for hidden dangers in chemical pipelines.

[0037] (2) Image acquisition and model training Approximately 20,000 images of normal and leak scenarios were collected from chemical plants, including about 1,000 positive leak samples.

[0038] The images were annotated at the pixel level, and positive samples were amplified to approximately 15,000 using geometric augmentation and GAN generation methods.

[0039] The YOLOv8 lightweight model was selected, and LoRA fine-tuning and 8-bit quantization optimization were used to complete the model training and compression.

[0040] After training, the model achieved a leak detection accuracy of ≥90%, a recall of ≥85%, and a single-frame inference latency of ≤0.2 seconds on the test set.

[0041] (3) Hazard identification and knowledge reasoning The edge-end model performs rapid initial leakage screening on real-time acquired images; Suspected leak images are uploaded to the cloud, where a large model combined with a knowledge graph performs a second assessment: identifying the leak type (dripping / jetting leak), inferring possible causes (abnormal pressure, loose flanges, etc.), and matching appropriate measures (closing valves, diverting water, etc.). The system then outputs the leak level and treatment recommendations, which are automatically recorded in the hazard management system.

[0042] 3. Technical Details Deployed at the edge, running an INT8 quantization model; The inference and knowledge base services are run on an Alibaba Cloud ECS cluster in the cloud. End-to-cloud transmission uses TLS encryption to ensure the security of monitoring data; The model is automatically updated weekly based on new samples.

[0043] 4. Effects This embodiment verifies the effectiveness of the invention in the scenario of pipeline leakage in chemical enterprises: The efficiency of hazard investigation is improved by more than 60%, and the accuracy of leak identification is improved by about 25% compared with manual inspection, effectively reducing downtime losses and manual inspection costs.

[0044] Example 2: Fault Identification and Knowledge Reasoning of Coal Mine Conveyor Belt Equipment 1. Scenario Description Coal mine conveyor belt systems are prone to problems such as belt misalignment, tearing, coal piling, and belt aging. Traditional inspections rely on manual observation, which carries high risks and the risk of missed inspections.

[0045] 2. Implementation Steps (1) The acquisition system is a fixed camera installed next to the conveyor belt, which acquires one frame every 5 seconds; (2) The image enters the lightweight model at the edge and performs deviation angle detection, material accumulation detection, and surface texture anomaly recognition. (3) The recognition results are transmitted to the cloud for knowledge reasoning: If the error is identified as "deviation", the knowledge base will search for possible causes such as: roller wear, idler deformation, and insufficient tension. If identified as "tear", match it with historical accident event chains to predict the risk level; (4) Integrate the test results with knowledge reasoning to generate system suggestions, such as: check the tensioning device, clean up the fallen coal, and replace the damaged idler roller; (5) The results are displayed on the monitoring platform and simultaneously entered into the hidden danger management system.

[0046] 3. Technical Details Using a few-sample learning strategy, common hidden dangers can be identified with only a few dozen positive samples. A portion of scarce potential sample data was generated using GAN adversarial generative techniques. The knowledge base is implemented using a graph database, which supports reasoning about complex fault chains.

[0047] 4. Effects It reduces manual inspection costs by more than 60%, achieves a 92% accuracy rate in identifying deviation-related hazards, and has a response time of less than 1 second.

[0048] Finally, it should be noted that the above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A safety production hazard intelligent identification system based on edge-cloud collaboration and few-sample learning, characterized in that, include: Security expert data platform: The expert knowledge base construction module is used to build an expert knowledge base using unstructured data such as expert documents, accident cases, investigation records, and equipment manuals from the chemical industry. Image acquisition module, used to acquire images of industrial scenes; The data governance module is used to perform hot and cold data separation, scene classification, and data preprocessing on the collected industrial scene images to obtain small-scale potential hazard samples. The visual model building module is used to fine-tune a visual recognition model on a small number of potential hazard samples by using transfer learning and ensemble learning techniques through a pre-trained visual feature extractor, and proposes a joint compression strategy. Security expert application platform: The data mining module is used to obtain online image data through web crawling algorithms, association rule algorithms, and clustering algorithms; The application software module is used to drive the visual recognition model to classify scenes and identify potential hazards in images, and output image features and recognition labels. The interaction module is used to link with the expert knowledge base to realize the closed loop of expert verification of pseudo-labels and real labels, and to complete expert review and knowledge supplementation. The decision output module combines knowledge and visual recognition results to output expert recommendations and accident predictions.

2. The intelligent safety production hazard identification system based on edge-cloud collaboration and few-sample learning according to claim 1, characterized in that, The expert knowledge base construction module describes the expert knowledge base construction process as follows: Natural language processing technology is used to process expert documents, accident cases, troubleshooting records, and equipment manuals, including: Named entity recognition: Identifies entities such as equipment names, components, risk points, personnel, and operating environment; Relationship extraction: Obtain the hierarchical relationship between "equipment – ​​component – ​​cause of failure – type of hazard"; Event extraction: Identify key events in the incident chain, including but not limited to leaks, short circuits, and overheating; The above information is written into the knowledge graph in the form of triples, forming a multi-scenario industrial safety knowledge system covering the elements of "human-machine-environment"; To ensure the continuous evolution of the expert knowledge base, a knowledge supplementation mechanism of "model preprocessing + expert review" is provided: Perform automated entity recognition and preliminary structuring on the newly added data; Experts review, supplement, and correct the content through the interactive interface; After passing the knowledge consistency verification, the knowledge is written into the expert knowledge base.

3. The intelligent safety hazard identification system based on edge-cloud collaboration and few-sample learning according to claim 1, characterized in that, The data governance module employs a hot and cold data separation mechanism, multi-scenario classification and noise cleaning technology, sample removal and consistency verification, and implements a label quality management mechanism. The data is expanded through multi-dimensional enhancement strategies, including geometric enhancement, lighting and background simulation, random occlusion, and noise perturbation. Geometric enhancements include rotation, clipping, and scaling.

4. The intelligent safety production hazard identification system based on edge-cloud collaboration and few-sample learning according to claim 1, characterized in that, In the visual model building module, the joint compression strategy is as follows: Low-rank matrix technology is used to reduce parameter storage overhead, and 8-bit integer quantization and binary quantization techniques are combined to compress the model size; pruning techniques are introduced to remove redundant parameters, and the network structure and loss function of the YOLO9 object detection model are optimized.

5. The intelligent safety production hazard identification system based on edge-cloud collaboration and few-sample learning according to claim 1, characterized in that, The security expert data platform also includes: The database is used to provide storage for image libraries, algorithm libraries, and tag libraries, supported by a large K8S cluster and cloud database, and to provide access interfaces for the image libraries, algorithm libraries, and tag libraries to the security expert application platform.

6. The intelligent safety production hazard identification system based on edge-cloud collaboration and few-sample learning according to claim 5, characterized in that, The database consists of an image library containing 10 million unsupervised images and over 500,000 positive samples, an algorithm library containing a constructed visual recognition model, model parameters, and inference logic, and a tag library containing annotation information.

Images