A safe production risk early warning method and system

Abstract

This invention discloses a method and system for early warning of safety production risks, relating to the field of safety production management technology. The method includes collecting multimodal data from the production site, preprocessing and fusing the multimodal data to form a fused feature dataset; analyzing the fused feature dataset to determine risk characteristics, and calibrating these risk characteristics, including parametric and descriptive calibration, and analyzing the calibrated content to determine risk hazard values; determining the preset risk hazard value range to which the risk hazard value belongs, and further determining the early warning level. This invention achieves deep fusion of multimodal data and intelligent risk identification, significantly improving the accuracy, timeliness, and interpretability of safety production risk early warning.

Description

Technical Field

[0001] This invention relates to the field of safety production management technology, and in particular to a safety production risk early warning method and system. Background Technology

[0002] With the acceleration of industrialization and the expansion of production scale, workplace safety has become a crucial issue concerning the safety of people's lives and property and the stable development of the economy and society. Especially in high-risk industries such as coal mining, hazardous chemicals, and industrial and commercial enterprises, the complex production environment and intertwined risk factors, including equipment failure, abnormal parameters, and human error, lead to frequent workplace accidents, causing huge losses to the country and enterprises. Existing workplace risk monitoring and early warning technologies mainly rely on sensor monitoring systems, video surveillance systems, and equipment operation log analysis. Sensor monitoring can collect data on parameters such as temperature, pressure, and gas concentration in real time, but it often uses threshold-based alarms, making it difficult to capture the dynamic trends of parameter curves and potential associated risks. While video surveillance can provide on-site images, traditional methods rely on manual inspections, which are inefficient and prone to oversights. Summary of the Invention

[0003] The purpose of this invention is to provide a method for early warning of safety production risks, comprising: Step S100: Collect multimodal data from the production site, and preprocess and fuse the multimodal data to form a fused feature dataset; Step S200: Analyze the fused feature dataset to determine the risk features, and label the risk features, including parametric labeling and descriptive labeling, and analyze the labeled content to determine the risk hazard value; Step S300: Determine the preset risk hazard value range to which the risk hazard value belongs, and then determine the warning level.

[0004] In some embodiments disclosed in this invention, multimodal data includes sensor monitoring data, video image data, and device operation log data.

[0005] In some embodiments disclosed in this invention, the method for analyzing the fused feature dataset includes: Analysis of parametric feature data and analysis of video image data; The methods for analyzing parametric feature data include: Step S201: Convert the parametric feature data into parametric curves, analyze the variation characteristics of each parametric curve, and extract the parametric curve segments that meet the preset monitoring standards. Step S202: The combination of parameter curve segments corresponding to several parameter feature data is identified as parameter curve feature group, and the production status corresponding to the parameter curve feature group within a preset time range is analyzed to obtain a status keyword group used to describe the production status. Methods for analyzing video image data include: Step S203: Construct a production site location map for the production site, determine the production site block to which each video surveillance belongs, and delineate the corresponding monitoring range block on the production site location map; Step S204: Identify the human body in the video image data and analyze the human body's posture, movement and position. Based on the analysis results, drive the preset simplified human body movement model to be configured at the corresponding position on the production site positioning map. Step S205: Analyze and judge the behavior of the simplified human action model on the production site positioning map, determine several corresponding behavior description keywords, and sort the behavior description keywords based on the probability of the behavior occurring corresponding to the behavior description keywords to form a behavior description keyword sequence.

[0006] In some embodiments disclosed in this invention, the method for analyzing the production status corresponding to the parameter curve feature group within a preset time range includes: Step S2021: Match and compare the parameter curve feature group with the preset production status knowledge base, and extract several production status descriptions with the highest similarity to the parameter curve feature group. Step S2022: Based on the extracted production status description and combined with the abnormal event annotations in historical production data, risk correlation analysis is performed on the matched production status description to identify potential risk types. Step S2023: Extract keywords from the descriptions corresponding to the identified risk types, and sort and filter them according to the frequency and importance of the keywords in the production status description to form status keyword groups; Methods for calculating the similarity between feature groups of parametric curves include: Step S20211: Normalize the parametric curve feature group, including amplitude scaling and time alignment. Step S20212: The Dynamic Time Warping (DTW) algorithm is used to calculate the distance between corresponding parameter curve segments in the parameter curve feature group, and the similarity between individual parameter curve segments is obtained. Step S20213: Perform a weighted average or comprehensive calculation on the similarity of multiple parametric curve segments to form the overall similarity of the parametric curve feature group; Step S20214: Based on the overall similarity threshold, extract several production status descriptions with the highest overall similarity.

[0007] In some embodiments disclosed in this invention, the method for configuring a preset simplified human motion model at a corresponding position on a production site positioning map includes: Step S2041: The simplified human motion model includes arm mapping lines, leg mapping lines, torso mapping lines, and human orientation mapping lines. Based on the analysis of human posture and motion, the changes of arm mapping lines, leg mapping lines, driving mapping lines, and human orientation mapping lines are driven frame by frame. Based on the analysis of the human body's location, the position change of the simplified human motion model on the production site positioning map is determined.

[0008] In some embodiments disclosed in this invention, the method for judging and analyzing the behavior of a simplified human motion model on a production site positioning map includes: Step S20411: Further define the content of the monitoring range blocks, including defining the work blocks, channel blocks, rest blocks and equipment blocks, and construct the model behavior judgment library corresponding to each block. The model behavior judgment library includes several model behavior judgment standards, and set the behavior description keywords corresponding to each model behavior judgment standard, and set the occurrence probability of the behavior corresponding to the behavior description keywords. Step S20412: Perform differential behavior feature analysis on the model behavior judgment criteria in the model behavior judgment library, generate several key equivalent segments at each level, and classify the model behavior judgment criteria at each level based on the key equivalent segments. Step S20413: Determine the specific location coordinates of the specific block within the monitoring range area to which the simplified human motion model belongs, and analyze the line change sequences of the arm mapping polyline, leg mapping polyline, driving mapping polyline, and human orientation mapping line of the simplified human motion model. Determine the key equivalent segment corresponding to the simplified human motion model, and call the corresponding model behavior judgment standard according to the key equivalent segment. Among them, the key equivalent segment must ensure that the distance between the specific location coordinates of the simplified human motion models is less than or equal to the preset value. Step S20414: Compare the performance of the several model behavior judgment criteria called with the simplified human action model to determine the degree of similarity between the two. Based on the degree of similarity, sort the model behavior judgment criteria and determine the model behavior judgment criteria with the highest degree of similarity. According to the sorting, compare the difference between the degree of similarity of other model behavior judgment criteria and the highest degree of similarity. If the difference meets the preset standard, the corresponding model behavior judgment criteria and the model behavior judgment criteria corresponding to the highest degree of similarity are selected and recorded as reference model behavior judgment criteria. Step S20415: Sort the behavioral description keywords corresponding to the behavioral judgment criteria of the reference model according to their occurrence probability to obtain a sequence of behavioral description keywords.

[0009] In some embodiments disclosed in this invention, the method for determining key equivalent sections includes: Step S204121: Spatial comparison is performed on the line change sequence in the model behavior judgment criteria. If there are line overlaps for more than a preset number of consecutive frames, and the number of model behavior judgment criteria with the same condition at the same time meets the preset criteria, then the line overlap for a preset number of consecutive frames is identified as a key equivalent segment. The line overlap includes the overlap of arm mapping polyline, leg mapping polyline, driving mapping polyline, and human body orientation mapping line. The overlap identification criteria are that the angle between the lines is less than or equal to a preset value, and the distance between the endpoints is less than or equal to a preset value.

[0010] In some embodiments disclosed in this invention, the method for calculating the approximation between the model behavior judgment criteria and the simplified human motion model includes: Step S204141: Compare the line body change sequence between the model behavior judgment standard and the performance of the simplified human motion model, including comparing the intersection angle and endpoint distance between relative lines between each relative frame, and determining the degree of approximation between the two based on the comparison results. ; Where J represents the degree of approximation. Let x be the matching parameter determination function between the x-th relative line bodies in the i-th frame. Based on the preset parameter ranges to which the intersection angle and endpoint distance belong, the corresponding approximate parameters are output. n is the number of relative frames participating in the comparison, N is the number of relative line bodies participating in the comparison in the i-th frame, D is the number of frames determined to be non-approximate relative frames among the relative frames participating in the comparison, and the method for determining non-approximate relative frames includes judging whether the sum of the approximate parameters of all relative line bodies in the same frame is higher than the preset value, K is the adjustment coefficient for the impact of continuous frame interruption, and b is the adjustment constant for the impact of continuous frame interruption.

[0011] In some embodiments disclosed in this invention, the method for analyzing parametric calibration and descriptive calibration to determine risk hazard values ​​includes: Step S206: Compare the parametric and descriptive calibration contents with the preset hazard values ​​in the hazard comparison database to determine the matching hazard contents, and output the risk hazard values ​​corresponding to the corresponding contents.

[0012] In some embodiments disclosed in this invention, a safety production risk early warning system is also disclosed, comprising: The first module is used to collect multimodal data from the production site, and to preprocess and fuse the multimodal data to form a fused feature dataset. The second module is used to analyze the fused feature dataset, identify risk features, and label the risk features, including parametric labeling and descriptive labeling. The labeled content is then analyzed to determine the risk hazard value. The third module is used to determine the preset risk hazard value range to which the risk hazard value belongs, and then determine the warning level.

[0013] This invention discloses a method and system for early warning of safety production risks, relating to the field of safety production management technology. The method includes collecting multimodal data from the production site, preprocessing and fusing the multimodal data to form a fused feature dataset; analyzing the fused feature dataset to determine risk characteristics, and calibrating these risk characteristics, including parametric and descriptive calibration, and analyzing the calibrated content to determine risk hazard values; determining the preset risk hazard value range to which the risk hazard value belongs, and further determining the early warning level. This invention achieves deep fusion of multimodal data and intelligent risk identification, significantly improving the accuracy, timeliness, and interpretability of safety production risk early warning.

[0014] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the steps of a safety production risk early warning method disclosed in this invention. Detailed Implementation

[0016] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0017] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings and specific embodiments. It should be understood that the preferred embodiments described herein are only for illustration and explanation of the present invention and should not be construed as limiting the scope of protection of the present invention. Those skilled in the art can make some non-essential improvements and adjustments based on the following content of the present invention. In the present invention, unless otherwise expressly specified and limited, the technical terms used in the present invention should have the ordinary meaning understood by those skilled in the art.

[0018] Example: The purpose of this invention is to provide a method for early warning of safety production risks, comprising: Step S100: Collect multimodal data from the production site, and preprocess and fuse the multimodal data to form a fused feature dataset.

[0019] Step S100 is the foundational step in the entire safety production risk early warning method. Its core lies in achieving comprehensive coverage of risk information at the production site through multimodal data acquisition, and improving data quality and usability through preprocessing and fusion. The multimodal data at the production site mainly includes sensor monitoring data (such as real-time values ​​of temperature, pressure, gas concentration, etc.), video image data (including visual information of personnel actions and equipment status), and equipment operation log data (text-based operation records, fault logs, etc.).

[0020] Step S200: Analyze the fused feature dataset to determine the risk features, and calibrate the risk features, including parametric calibration and descriptive calibration. Analyze the calibration content to determine the risk hazard value.

[0021] Step S200 is the core processing step for risk identification and quantification. It transforms raw features into risk hazard values ​​through in-depth analysis of the fused feature dataset. First, the fused feature dataset undergoes separate or joint analysis: parametric features are processed using curve manipulation, DTW dynamic time warping similarity matching, and comparison with a production status knowledge base to extract status keyword groups; video image features are processed using a simplified model driven by human posture recognition (composed of arm, leg, torso, and orientation lines), combined with block division, key equivalent segment classification, and a custom approximation formula (integrating inter-frame angles, endpoint distances, and interruption adjustments) to determine behavior and generate a sorted sequence of behavior description keywords. These analysis results form parametric calibrations (quantifying curve anomalies and deviation values) and descriptive calibrations (keywords and their probability of occurrence). Subsequently, the two types of calibrations are comprehensively analyzed, combined with a preset hazard value comparison database or a weighted fusion model, to correlate parameter anomalies with unsafe behaviors, ultimately outputting quantified risk hazard values. This step organically unifies numerical and behavioral risks, overcoming the shortcomings of existing technologies such as the separation of parameters and video analysis and subjective risk labeling, and achieving more accurate and interpretable risk feature identification and hazard quantification.

[0022] Step S300: Determine the preset risk hazard value range to which the risk hazard value belongs, and then determine the warning level.

[0023] Step S300 is the final step in early warning decision-making. Its principle is to achieve graded early warning based on the quantitative results of risk hazard values, so as to ensure that the early warning response matches the severity of the risk.

[0024] In some embodiments disclosed in this invention, multimodal data includes sensor monitoring data, video image data, and device operation log data.

[0025] In some embodiments disclosed in this invention, the method for analyzing the fused feature dataset includes: Analysis of parametric feature data and analysis of video image data; The methods for analyzing parametric feature data include: Step S201: Convert the parametric feature data into parametric curves, analyze the variation characteristics of each parametric curve, and extract the parametric curve segments that meet the preset monitoring standards.

[0026] Step S202: The combination of parameter curve segments corresponding to several parameter feature data is identified as parameter curve feature group, and the production status corresponding to the parameter curve feature group within a preset time range is analyzed to obtain a status keyword group used to describe the production status.

[0027] Methods for analyzing video image data include: Step S203: Construct a production site location map for the production site, determine the production site block to which each video surveillance belongs, and delineate the corresponding monitoring range block on the production site location map.

[0028] Step S204: Identify the human body in the video image data and analyze the human body's posture, movement and position. Based on the analysis results, drive the preset simplified human body movement model to be configured at the corresponding position on the production site positioning map.

[0029] Step S205: Analyze and judge the behavior of the simplified human action model on the production site positioning map, determine several corresponding behavior description keywords, and sort the behavior description keywords based on the probability of the behavior occurring corresponding to the behavior description keywords to form a behavior description keyword sequence.

[0030] In some embodiments disclosed in this invention, the method for analyzing the production status corresponding to the parameter curve feature group within a preset time range includes: Step S2021: Match and compare the parameter curve feature group with the preset production status knowledge base, and extract several production status descriptions with the highest similarity to the parameter curve feature group.

[0031] Step S2021 is a crucial matching step in parametric risk feature analysis. Its core principle lies in achieving precise comparison between the current parameter curve feature set and the historical production status knowledge base through similarity calculation, thereby mapping the abstract curve pattern to known production status descriptions. The parameter curve feature set is composed of multiple extracted parameter curve segments. After normalization processing (amplitude scaling and time alignment), the distance between each curve segment is calculated using the Dynamic Time Warping (DTW) algorithm. This algorithm effectively handles the nonlinear scaling and offset of time series, obtaining the similarity of individual curve segments, which is then weighted averaged or comprehensively calculated to form the overall similarity. This overall similarity reflects the degree of matching between the current parameter change pattern and the historical production status (such as normal operation, load fluctuation, abnormal temperature rise, etc.) stored in the knowledge base. Subsequently, based on a preset threshold, the most similar production status descriptions are extracted.

[0032] Step S2022: Based on the extracted production status description and combined with the abnormal event annotations in historical production data, risk correlation analysis is performed on the matched production status description to identify potential risk types.

[0033] The principle of step S2022 is to use historical anomaly event annotations as supervisory signals to perform risk association mining on the high-similarity production status descriptions extracted in step S2021, thereby realizing the inference transformation from "similar status" to "potential risk type". The extracted production status descriptions themselves only represent the similarity of the parameter curve patterns, but do not directly carry risk information. Therefore, it is necessary to combine them with the anomaly events (such as equipment failure, leakage, explosion, etc.) already annotated in historical production data for association analysis.

[0034] Step S2023: Extract keywords from the descriptions corresponding to the identified risk types, and sort and filter them according to the frequency and importance of the keywords in the production status description to form status keyword groups.

[0035] The principle behind step S2023 lies in extracting interpretable descriptive features from the identified risk types to form structured condition keyword groups, providing high-quality textual feature input for subsequent descriptive labeling and risk hazard value calculation. Identified potential risk types typically correspond to multiple historical descriptive texts. First, natural language processing techniques (such as TF-IDF, TextRank, or keyword extraction models) are used to extract high-frequency and high-importance words (such as "overload," "leakage," and "fatigue") from these descriptions. Then, these words are sorted and filtered according to their frequency of occurrence in the risk type descriptions (reflecting prevalence) and importance weight (reflecting discriminative power), ultimately forming ordered condition keyword groups. These keyword groups not only retain the core semantics of the risk but also reflect the priority or severity of the risk through their ranking. This dimensionality reduction and structuring process from risk type to keywords enhances the interpretability and computability of descriptive labeling, facilitates integration with parametric labeling, and provides standardized input for matching in the hazard value comparison database, significantly improving the accuracy and traceability of the overall risk assessment.

[0036] Methods for calculating the similarity between feature groups of parametric curves include: Step S20211: Normalize the parametric curve feature group, including amplitude scaling and time alignment.

[0037] Step S20212: The Dynamic Time Warping (DTW) algorithm is used to calculate the distance between corresponding parameter curve segments in the parameter curve feature group, and the similarity between individual parameter curve segments is obtained.

[0038] Step S20213: Perform a weighted average or comprehensive calculation on the similarity of multiple parametric curve segments to form the overall similarity of the parametric curve feature group.

[0039] Step S20214: Based on the overall similarity threshold, extract several production status descriptions with the highest overall similarity.

[0040] In some embodiments disclosed in this invention, the method for configuring a preset simplified human motion model at a corresponding position on a production site positioning map includes: Step S2041: The simplified human motion model includes arm mapping lines, leg mapping lines, torso mapping lines, and human orientation mapping lines. Based on the analysis of human posture and motion, the changes of arm mapping lines, leg mapping lines, driving mapping lines, and human orientation mapping lines are driven frame by frame. Based on the analysis of the human body's location, the position change of the simplified human motion model on the production site positioning map is determined.

[0041] In some embodiments disclosed in this invention, the method for judging and analyzing the behavior of a simplified human motion model on a production site positioning map includes: Step S20411: Further define the content of the monitoring range blocks, including defining the work blocks, channel blocks, rest blocks and equipment blocks, and construct the model behavior judgment library corresponding to each block. The model behavior judgment library includes several model behavior judgment standards, and set the behavior description keywords corresponding to each model behavior judgment standard, and set the occurrence probability of the behavior corresponding to the behavior description keywords. Step S20412: Perform differential behavior feature analysis on the model behavior judgment criteria in the model behavior judgment library, generate several key equivalent segments at each level, and classify the model behavior judgment criteria at each level based on the key equivalent segments. Step S20413: Determine the specific location coordinates of the specific block within the monitoring range area to which the simplified human motion model belongs, and analyze the line change sequences of the arm mapping polyline, leg mapping polyline, driving mapping polyline, and human orientation mapping line of the simplified human motion model. Determine the key equivalent segment corresponding to the simplified human motion model, and call the corresponding model behavior judgment standard according to the key equivalent segment. Among them, the key equivalent segment must ensure that the distance between the specific location coordinates of the simplified human motion models is less than or equal to the preset value. Step S20414: Compare the performance of the several model behavior judgment criteria called with the simplified human action model to determine the degree of similarity between the two. Based on the degree of similarity, sort the model behavior judgment criteria and determine the model behavior judgment criteria with the highest degree of similarity. According to the sorting, compare the difference between the degree of similarity of other model behavior judgment criteria and the highest degree of similarity. If the difference meets the preset standard, the corresponding model behavior judgment criteria and the model behavior judgment criteria corresponding to the highest degree of similarity are selected and recorded as reference model behavior judgment criteria. Step S20415: Sort the behavioral description keywords corresponding to the behavioral judgment criteria of the reference model according to their occurrence probability to obtain a sequence of behavioral description keywords.

[0042] In some embodiments disclosed in this invention, the method for determining key equivalent sections includes: Step S204121: Spatial comparison is performed on the line change sequence in the model behavior judgment criteria. If there are line overlaps for more than a preset number of consecutive frames, and the number of model behavior judgment criteria with the same condition at the same time meets the preset criteria, then the line overlap for a preset number of consecutive frames is identified as a key equivalent segment. The line overlap includes the overlap of arm mapping polyline, leg mapping polyline, driving mapping polyline, and human body orientation mapping line. The overlap identification criteria are that the angle between the lines is less than or equal to a preset value, and the distance between the endpoints is less than or equal to a preset value.

[0043] In some embodiments disclosed in this invention, the method for calculating the approximation between the model behavior judgment criteria and the simplified human motion model includes: Step S204141: Compare the line body change sequence between the model behavior judgment standard and the performance of the simplified human motion model, including comparing the intersection angle and endpoint distance between relative lines between each relative frame, and determining the degree of approximation between the two based on the comparison results. ; Where J represents the degree of approximation. Let x be the matching parameter determination function between the x-th relative line bodies in the i-th frame. Based on the preset parameter ranges to which the intersection angle and endpoint distance belong, the corresponding approximate parameters are output. n is the number of relative frames participating in the comparison, N is the number of relative line bodies participating in the comparison in the i-th frame, D is the number of frames determined to be non-approximate relative frames among the relative frames participating in the comparison, and the method for determining non-approximate relative frames includes judging whether the sum of the approximate parameters of all relative line bodies in the same frame is higher than the preset value, K is the adjustment coefficient for the impact of continuous frame interruption, and b is the adjustment constant for the impact of continuous frame interruption.

[0044] Wherein, K: the adjustment coefficient for the impact of continuous frame interruptions (penalty strength coefficient). The larger K is, the stronger the penalty for non-approximate frames, emphasizing the continuity and stability of the action sequence.

[0045] b: Continuous frame interruptions affect the adjustment constant (bias term). This constant term is used to fine-tune the penalty base, preventing the penalty from being too small or too large when D=0.

[0046] In some embodiments disclosed in this invention, the method for analyzing parametric calibration and descriptive calibration to determine risk hazard values ​​includes: Step S206: Compare the parametric and descriptive calibration contents with the preset hazard values ​​in the hazard comparison database to determine the matching hazard contents, and output the risk hazard values ​​corresponding to the corresponding contents.

[0047] In some embodiments disclosed in this invention, a safety production risk early warning system is also disclosed, comprising: The first module is used to collect multimodal data from the production site, and to preprocess and fuse the multimodal data to form a fused feature dataset. The second module is used to analyze the fused feature dataset, identify risk features, and label the risk features, including parametric labeling and descriptive labeling. The labeled content is then analyzed to determine the risk hazard value. The third module is used to determine the preset risk hazard value range to which the risk hazard value belongs, and then determine the warning level.

[0048] Through the above description of the embodiments, those skilled in the art can clearly understand that the present invention can be implemented in hardware or by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of the present invention can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, external hard drive, etc.) and includes several instructions to cause a computer device (such as a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0049] This invention discloses a method and system for early warning of safety production risks, relating to the field of safety production management technology. The method includes collecting multimodal data from the production site, preprocessing and fusing the multimodal data to form a fused feature dataset; analyzing the fused feature dataset to determine risk characteristics, and calibrating these risk characteristics, including parametric and descriptive calibration, and analyzing the calibrated content to determine risk hazard values; determining the preset risk hazard value range to which the risk hazard value belongs, and further determining the early warning level. This invention achieves deep fusion of multimodal data and intelligent risk identification, significantly improving the accuracy, timeliness, and interpretability of safety production risk early warning.

[0050] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for early warning of safety production risks, characterized in that, include: Step S100: Collect multimodal data from the production site, and preprocess and fuse the multimodal data to form a fused feature dataset; Step S200: Analyze the fused feature dataset to determine the risk features, and label the risk features, including parametric labeling and descriptive labeling, and analyze the labeled content to determine the risk hazard value; Step S300: Determine the preset risk hazard value range to which the risk hazard value belongs, and then determine the warning level.

2. The method for early warning of safety production risks according to claim 1, characterized in that, Multimodal data includes sensor monitoring data, video image data, and equipment operation log data.

3. The method for early warning of safety production risks according to claim 1, characterized in that, Methods for analyzing fused feature datasets include: Analysis of parametric feature data and analysis of video image data; The methods for analyzing parametric feature data include: Step S201: Convert the parametric feature data into parametric curves, analyze the variation characteristics of each parametric curve, and extract the parametric curve segments that meet the preset monitoring standards. Step S202: The combination of parameter curve segments corresponding to several parameter feature data is identified as parameter curve feature group, and the production status corresponding to the parameter curve feature group within a preset time range is analyzed to obtain a status keyword group used to describe the production status. Methods for analyzing video image data include: Step S203: Construct a production site location map for the production site, determine the production site block to which each video surveillance belongs, and delineate the corresponding monitoring range block on the production site location map; Step S204: Identify the human body in the video image data and analyze the human body's posture, movement and position. Based on the analysis results, drive the preset simplified human body movement model to be configured at the corresponding position on the production site positioning map. Step S205: Analyze and judge the behavior of the simplified human action model on the production site positioning map, determine several corresponding behavior description keywords, and sort the behavior description keywords based on the probability of the behavior occurring corresponding to the behavior description keywords to form a behavior description keyword sequence.

4. The method for early warning of safety production risks according to claim 3, characterized in that, Methods for analyzing the production status corresponding to the characteristic group of the parameter curve within a preset time range include: Step S2021: Match and compare the parameter curve feature group with the preset production status knowledge base, and extract several production status descriptions with the highest similarity to the parameter curve feature group. Step S2022: Based on the extracted production status description and combined with the abnormal event annotations in historical production data, risk correlation analysis is performed on the matched production status description to identify potential risk types. Step S2023: Extract keywords from the descriptions corresponding to the identified risk types, and sort and filter them according to the frequency and importance of the keywords in the production status description to form status keyword groups; Methods for calculating the similarity between feature groups of parametric curves include: Step S20211: Normalize the parametric curve feature group, including amplitude scaling and time alignment. Step S20212: The Dynamic Time Warping (DTW) algorithm is used to calculate the distance between corresponding parameter curve segments in the parameter curve feature group, and the similarity between individual parameter curve segments is obtained. Step S20213: Perform a weighted average or comprehensive calculation on the similarity of multiple parametric curve segments to form the overall similarity of the parametric curve feature group; Step S20214: Based on the overall similarity threshold, extract several production status descriptions with the highest overall similarity.

5. A method for early warning of safety production risks according to claim 3, characterized in that, Methods for configuring a pre-defined simplified human motion model at its corresponding position on a production site positioning map include: Step S2041: The simplified human motion model includes arm mapping lines, leg mapping lines, torso mapping lines, and human orientation mapping lines. Based on the analysis of human posture and motion, the changes of arm mapping lines, leg mapping lines, driving mapping lines, and human orientation mapping lines are driven frame by frame. Based on the analysis of the human body's location, the position change of the simplified human motion model on the production site positioning map is determined.

6. A method for early warning of safety production risks according to claim 5, characterized in that, Methods for judging and analyzing the behavior of simplified human motion models on production site positioning maps include: Step S20411: Further define the content of the monitoring range blocks, including defining the work blocks, channel blocks, rest blocks and equipment blocks, and construct the model behavior judgment library corresponding to each block. The model behavior judgment library includes several model behavior judgment standards, and set the behavior description keywords corresponding to each model behavior judgment standard, and set the occurrence probability of the behavior corresponding to the behavior description keywords. Step S20412: Perform differential behavior feature analysis on the model behavior judgment criteria in the model behavior judgment library, generate several key equivalent segments at each level, and classify the model behavior judgment criteria at each level based on the key equivalent segments. Step S20413: Determine the specific location coordinates of the specific block within the monitoring range area to which the simplified human motion model belongs, and analyze the line change sequences of the arm mapping polyline, leg mapping polyline, driving mapping polyline, and human orientation mapping line of the simplified human motion model. Determine the key equivalent segment corresponding to the simplified human motion model, and call the corresponding model behavior judgment standard according to the key equivalent segment. Among them, the key equivalent segment must ensure that the distance between the specific location coordinates of the simplified human motion models is less than or equal to the preset value. Step S20414: Compare the performance of the several model behavior judgment criteria called with the simplified human action model to determine the degree of similarity between the two. Based on the degree of similarity, sort the model behavior judgment criteria and determine the model behavior judgment criteria with the highest degree of similarity. According to the sorting, compare the difference between the degree of similarity of other model behavior judgment criteria and the highest degree of similarity. If the difference meets the preset standard, the corresponding model behavior judgment criteria and the model behavior judgment criteria corresponding to the highest degree of similarity are selected and recorded as reference model behavior judgment criteria. Step S20415: Sort the behavioral description keywords corresponding to the behavioral judgment criteria of the reference model according to their occurrence probability to obtain a sequence of behavioral description keywords.

7. A method for early warning of safety production risks according to claim 6, characterized in that, Methods for identifying key equivalent segments include: Step S204121: Spatial comparison is performed on the line change sequence in the model behavior judgment criteria. If there are line overlaps for more than a preset number of consecutive frames, and the number of model behavior judgment criteria with the same condition at the same time meets the preset criteria, then the line overlap for a preset number of consecutive frames is identified as a key equivalent segment. The line overlap includes the overlap of arm mapping polyline, leg mapping polyline, driving mapping polyline, and human body orientation mapping line. The overlap identification criteria are that the angle between the lines is less than or equal to a preset value, and the distance between the endpoints is less than or equal to a preset value.

8. A method for early warning of safety production risks according to claim 6, characterized in that, Methods for calculating the approximation between the behavioral judgment criteria of the computational model and the simplified human motion model include: Step S204141: Compare the line body change sequence between the model behavior judgment standard and the performance of the simplified human motion model, including comparing the intersection angle and endpoint distance between relative lines between each relative frame, and determining the degree of approximation between the two based on the comparison results. ； Where J represents the degree of approximation. Let x be the matching parameter determination function between the x-th relative line bodies in the i-th frame. Based on the preset parameter ranges to which the intersection angle and endpoint distance belong, the corresponding approximate parameters are output. n is the number of relative frames participating in the comparison, N is the number of relative line bodies participating in the comparison in the i-th frame, D is the number of frames determined to be non-approximate relative frames among the relative frames participating in the comparison, and the method for determining non-approximate relative frames includes judging whether the sum of the approximate parameters of all relative line bodies in the same frame is higher than the preset value, K is the adjustment coefficient for the impact of continuous frame interruption, and b is the adjustment constant for the impact of continuous frame interruption.

9. A method for early warning of safety production risks according to claim 1, characterized in that, Methods for determining risk hazard values ​​by analyzing parametric and descriptive calibrations include: Step S206: Compare the parametric and descriptive calibration contents with the preset hazard values ​​in the hazard comparison database to determine the matching hazard contents, and output the risk hazard values ​​corresponding to the corresponding contents.

10. A safety production risk early warning system, characterized in that, include: The first module is used to collect multimodal data from the production site, and to preprocess and fuse the multimodal data to form a fused feature dataset. The second module is used to analyze the fused feature dataset, identify risk features, and label the risk features, including parametric labeling and descriptive labeling. The labeled content is then analyzed to determine the risk hazard value. The third module is used to determine the preset risk hazard value range to which the risk hazard value belongs, and then determine the warning level.

Images