CN-121998410-A - Production safety digital management method and system based on space-time risk field evolution

CN121998410ACN 121998410 ACN121998410 ACN 121998410ACN-121998410-A

Abstract

The application discloses a production safety digital management method and system based on space-time risk field evolution, and belongs to the technical field of computer application and artificial intelligence. The method comprises the steps of constructing a multi-mode safety data set and a specific data set for training a production safety model, carrying out multi-mode pre-training on a preset safety base large model based on the multi-mode safety data to obtain a pre-training model with universal safety situation sensing capability, carrying out embedded fine adjustment on the safety mechanism of the pre-training model based on the specific data set to obtain a production safety model capable of understanding the safety mechanism and carrying out risk deduction, obtaining multi-mode real-time data of a real-time production site, carrying out industrial reasoning on the multi-mode real-time data by utilizing the production safety model to obtain reasoning results comprising risk grades, evolution trends and causal paths, generating early warning information, an interpretable risk tracing path and a device control instruction according to the reasoning results, and pushing the early warning information, the interpretable risk tracing path and the device control instruction to target equipment.

Inventors

GUO JIAN
YANG LEI
ZHU FEIFEI
ZHAO LEILEI
WANG FENG
LI HOURONG
HU JIWEI
LIU YING

Assignees

中信重工机械股份有限公司
中国中信有限公司

Dates

Publication Date: 20260508
Application Date: 20251226

Claims (10)

1. A production security digital management method based on space-time risk field evolution, which is characterized by being used for training a production security model and utilizing the production security model to conduct security management, the method comprising: The method comprises the steps of constructing a multi-mode safety data set and a specific data set, wherein the multi-mode safety data set is used for training a production safety model and comprises correlated historical sensor data, historical video monitoring data and unstructured safety text data, and the specific data set comprises correlated accident chain knowledge maps and space-time risk field evolution data, wherein the accident chain knowledge maps are constructed based on historical accident and risk and operability analysis; based on the multi-mode security data, carrying out multi-mode pre-training on a preset security base large model to obtain a pre-training model with general security situation sensing capability; based on the specific data set, carrying out embedded fine adjustment on the safety mechanism of the pre-training model so as to obtain a production safety model capable of understanding the safety mechanism and carrying out risk deduction; acquiring multi-mode real-time data of a real-time production site, and carrying out industrial reasoning on the multi-mode real-time data by utilizing the production safety model to obtain a reasoning result comprising a risk level, an evolution trend and a causal path; and generating early warning information, an interpretable risk tracing path and a device control instruction according to the reasoning result, and pushing the early warning information, the interpretable risk tracing path and the device control instruction to target devices.
2. The method of claim 1, wherein the operation of constructing a multimodal safety dataset for training the production safety model comprises: acquiring historical sensor data, historical video monitoring data and unstructured safe text data of a target production environment, wherein the unstructured safe text data comprises national safety standards, industry operation rules and historical accident reports; carrying out semantic analysis on the unstructured safe text data, and extracting key safe concepts and constraint conditions; aligning and associating the key safety concept with personnel behaviors and equipment states in the historical video monitoring data and the historical sensor data in a semantic level to form a cross-modal semantic association data pair; and cleaning, denoising and standardizing each semantic association data pair to generate the multi-mode safety data set.
3. The method of claim 1, wherein the operation of constructing a particular data set for training the production safety model comprises: acquiring historical accident reports, risk and operability analysis records, equipment state historical time sequence data, personnel behavior historical tracks and environmental parameters; Carrying out structured information extraction on the historical accident report and the dangerous and operability analysis record, and identifying and defining key event nodes and causal relations in the accident evolution process; Constructing an accident chain knowledge graph with the event as a node and the causal logic as an edge based on the key event node and the causal relation; modeling a production site as a dynamically evolving multidimensional risk field according to the equipment state historical time sequence data, the personnel behavior historical track and the environmental parameter, and generating space-time risk field evolution data for describing risk space-time diffusion and evolution; mapping and correlating key event nodes in the accident chain knowledge graph with physical parameters in the space-time risk field evolution data to generate the specific data set.
4. The method of claim 2, wherein performing a multi-modal pre-training operation on a pre-set security base large model based on the multi-modal security data comprises: the method comprises the steps of extracting characteristics of historical sensor data, historical video monitoring data and parsed unstructured safe text data in the multi-mode safe data set through a sensor encoder, a visual encoder and a text encoder respectively, and mapping the characteristics to a unified semantic embedded space to obtain standardized characteristic representation of each mode of data; Fusing the standardized feature representations of all the modal data to generate a fused feature vector capable of representing the overall security situation; the security base large model is pre-trained through a cross-modal comparison learning loss function based on the fusion feature vector, wherein the cross-modal comparison learning loss function comprises an information noise comparison estimation loss item and a cross-modal alignment loss item, the information noise comparison estimation loss item is used for shortening the distance of semantic association data pairs in the semantic embedding space and pushing the distance of non-association data pairs, and the cross-modal alignment loss item is used for restraining the consistency of feature distribution output by different modal encoders in the semantic embedding space.
5. A method according to claim 3, wherein performing security mechanism embedded fine-tuning of the pre-trained model based on the specific data set comprises: The accident chain knowledge graph and the space-time risk field evolution data which are related to each other in the specific data set are used as fine tuning samples together; The nodes and edges in the accident chain knowledge graph are coded to obtain corresponding graph vector representations, and the graph vector representations are input into the pre-training model for learning, so that the pre-training model can obtain causal deduction capability; Inputting the space-time risk field evolution data into the pre-training model, constructing a physical information loss item by calculating residual errors between a risk field state predicted by the pre-training model and a physical conservation law followed in the space-time risk field evolution data, and restricting training of the pre-training model by using the physical information loss item so that the predicted output of the pre-training model accords with the physical law of the risk field evolution; In a safe digital twin environment synchronous with a physical production site, taking the pre-training model as an intelligent agent, taking the minimized system risk as a reward function, and optimizing a decision strategy of the pre-training model through reinforcement learning.
6. The method of claim 1, wherein the act of industrially reasoning the multi-modal real-time data using the production safety model to derive reasoning results including risk levels, evolution trends and causal paths comprises: Inputting the acquired multi-modal real-time data into the production safety model, wherein the multi-modal real-time data comprises real-time sensor data, real-time video stream data and production context data; Carrying out fusion analysis on the multi-mode real-time data through the production safety model, identifying risk events, simulating dynamic changes of a risk field under the influence of the risk events, and predicting evolution trends of risk intensity, influence range and diffusion direction in a specific future period; Carrying out causal deduction on the identified risk event through the production safety model to generate a corresponding accident development path and event occurrence probability; And generating an inference result comprising a comprehensive risk level, a space-time map for representing the risk evolution trend and a key causal path according to the evolution trend, the accident development path and the event occurrence probability.
7. The method of claim 6, wherein generating early warning information, an interpretable risk trace path, and device control instructions based on the inference results comprises: generating early warning information corresponding to the risk level according to the comprehensive risk level; based on the key causal path, invoking the accident chain knowledge graph to generate a visual causal deducing path from a risk source to a potential result as an interpretable risk tracing path; And generating a device control instruction sequence aiming at the target device through a preset safety strategy rule according to the risk evolution trend space-time spectrum and the accident development path.
8. A storage medium comprising a stored program, wherein the method of any one of claims 1 to 7 is performed by a processor when the program is run.
9. Production safety digital management system based on space-time risk field evolution, which is characterized by comprising: A security data infrastructure layer configured to build a multimodal security data set and a specific data set for training a production security model; the multi-mode security data set comprises correlated historical sensor data, historical video monitoring data and unstructured security text data, wherein the specific data set comprises correlated accident chain knowledge maps and space-time risk field evolution data, and the accident chain knowledge maps are constructed based on historical accident and risk and operability analysis; the multi-mode safety data fusion layer is configured to perform multi-mode pre-training on a preset safety base large model based on the multi-mode safety data so as to obtain a pre-training model with universal safety situation sensing capability; The safety mechanism embedding and risk deduction layer is configured to conduct embedded fine adjustment on the safety mechanism of the pre-training model based on the specific data set so as to obtain a production safety model capable of understanding the safety mechanism and conducting risk deduction; the intelligent early warning and decision interaction layer is configured to acquire multi-mode real-time data of a real-time production site, and perform industrial reasoning on the multi-mode real-time data by utilizing the production safety model to acquire reasoning results comprising risk grades, evolution trends and causal paths; And the intelligent security application layer is configured to generate early warning information, an interpretable risk tracing path and a device control instruction according to the reasoning result and push the early warning information, the interpretable risk tracing path and the device control instruction to the target device.
10. The system of claim 9, wherein the security data infrastructure layer constructs the multimodal security data set by obtaining historical sensor data, historical video monitoring data, and unstructured security text data of a target production environment, wherein the unstructured security text data comprises national security standards, industry operational procedures, and historical incident reports, performing semantic parsing on the unstructured security text data, extracting key security concepts and constraint conditions, performing semantic level alignment and association on the key security concepts and personnel behaviors and equipment states in the historical video monitoring data and the historical sensor data to form cross-modal semantic association data pairs, and performing cleaning, denoising, and standardization processing on each semantic association data pair to generate the multimodal security data set.

Description

Production safety digital management method and system based on space-time risk field evolution Technical Field The application relates to the technical field of computer application and artificial intelligence, in particular to a production safety digital management method and system based on space-time risk field evolution. Background With the development of industrial internet of things (IoT) and big data technologies, modern production security management systems have turned from traditional manual inspection and post-hoc response to digital and informative management. However, the existing production safety management system still has the problems of poor generalization capability, lack of active prediction capability, and incapacitation of safety mechanism. The method specifically comprises the following steps: (1) The existing video monitoring and sensor (such as gas and temperature) data and personnel inspection records are split, so that cross-mode fusion analysis cannot be performed; (2) Passive response but not active prediction, wherein the traditional system seriously depends on post report and fixed threshold alarm, and can not predict and deduce the chain reaction of risks according to multidimensional factors; (3) The existing system does not understand the causal logic behind the 'safe SOP (standard operation program)', 'national safety Specification' or 'accident chain', but only presents simple data, so that the result output by the model is not accurate and reliable enough. Therefore, the existing production safety management system has poor generalization capability, weak prediction capability and lack of understanding of a safety mechanism, and is a problem to be solved urgently. Disclosure of Invention The embodiment of the disclosure provides a production safety digital management method and system based on space-time risk field evolution, which at least solve the technical problems of poor generalization capability, weak prediction capability and lack of understanding of a safety mechanism of a production safety management system in the prior art. According to one aspect of the embodiment of the disclosure, a production safety digital management method based on space-time risk field evolution is provided, and is used for training a production safety model and utilizing the production safety model to conduct safety management, and comprises a multi-mode safety data set and a specific data set, wherein the multi-mode safety data set comprises correlated historical sensor data, historical video monitoring data and unstructured safety text data, the specific data set comprises correlated accident chain knowledge patterns and space-time risk field evolution data, the accident chain knowledge patterns are constructed based on historical accidents, risks and operability analysis, based on the multi-mode safety data, a pre-training model with universal safety situation sensing capability is obtained by carrying out multi-mode pre-training on a preset safety base large model, based on the specific data set, the pre-training model is subjected to internal fine adjustment of safety mechanisms so as to obtain the production safety model capable of understanding the safety mechanisms and carrying out risk modeling, the real-time production field real-time data are obtained, the accident chain knowledge patterns are obtained by utilizing the production safety, the accident chain knowledge patterns are obtained by carrying out analysis on historical accidents, risks and the risk level can be controlled by carrying out the analysis, and the inference and the risk level can be controlled by a real-time inference device, and the risk level is controlled by a trace, and a target is obtained. According to another aspect of the embodiments of the present disclosure, there is also provided a storage medium including a stored program, wherein the method described above is performed by a processor when the program is run. According to another aspect of an embodiment of the present disclosure, there is also provided a production security digital management system based on spatio-temporal risk field evolution, including a security data infrastructure layer configured to construct a multimodal security data set and a specific data set for training a production security model; the multi-mode security data set comprises correlated historical sensor data, historical video monitoring data and unstructured security text data; the intelligent early warning and decision interaction layer is configured to acquire multi-mode real-time data of a real-time production site, and utilize the production safety model to carry out industrial reasoning on the multi-mode real-time data to acquire a reasoning result comprising a risk level, an evolution trend and a causal path, and the intelligent safety application layer is configured to generate early warning information, an interpretable risk tracing path and a device control