Methods, systems, equipment and media for safety management of confined space operations

Abstract

This invention discloses a method, system, equipment, and medium for safety management of confined space operations, relating to the field of industrial safety production technology. The method includes: simultaneously collecting environmental parameters, personnel status, and equipment status to obtain multimodal data; normalizing the multimodal data to construct a fused feature vector containing environmental, physiological, and behavioral characteristics; inputting the fused feature vector into a preset intelligent analysis model for real-time analysis, identifying pre-operation risks, dynamic risks during operation, and predicting personnel safety status, outputting risk identification results and early warning levels for safety status; triggering corresponding graded early warning actions based on the early warning level; matching the risk identification results with a preset accident case database to generate an emergency decision-making plan and execute a coordinated response. This invention solves the problems of insufficient data fusion and collaborative analysis capabilities, low real-time and accuracy of early warnings, and low level of intelligent emergency decision-making in existing technologies.

Description

Technical Field

[0001] This invention relates to the field of industrial safety production technology, and in particular to a method, system, equipment and medium for safety management and control of confined space operations. Background Technology

[0002] Confined space operations (such as cable wells, deep foundation pits, and inside transformers) have always been high-risk aspects of industrial safety production due to their enclosed environment, poor ventilation, and complex risk sources (such as oxygen deficiency, accumulation of toxic and harmful gases, high temperature and humidity, etc.). Accidents are prone to occur and have serious consequences. Traditional safety management models mainly rely on manual periodic inspections, post-accident emergency response, and the deployment of single sensors (such as gas detectors) for monitoring, which makes it difficult to achieve real-time, comprehensive, and accurate control of operational risks.

[0003] Currently, existing technologies mainly focus on single-dimensional monitoring, such as environmental parameter monitoring: detecting oxygen and toxic gas concentrations through gas sensors, but failing to integrate human behavior and health data, and thus unable to correlate the impact of environmental changes on personnel safety. Another example is video surveillance: relying on manual video inspections or simple image recognition, it lacks accuracy in recognizing low-light conditions and complex postures, and cannot link with sensor data for early warning. Yet another example is post-event emergency management: decision-making based on historical cases lacks real-time data-driven intelligent auxiliary systems, leading to blindly formulated rescue strategies and a high risk of secondary accidents. In summary, existing technologies suffer from deficiencies in data fusion and collaborative analysis capabilities, inadequate real-time and accurate early warning systems, and low levels of intelligence in emergency decision-making.

[0004] Therefore, there is an urgent need to design a safety management method for confined space operations. This method should integrate multi-dimensional data on the environment, equipment, and personnel to achieve intelligent identification and graded early warning of violations throughout the entire process before, during, and after the operation; real-time monitoring and accurate prediction of personnel's life safety status; and intelligent auxiliary decision-making for emergencies, thereby reducing accident risks and handling costs. Summary of the Invention

[0005] This invention provides a method, system, equipment, and medium for safety management of confined space operations, which addresses the problems of insufficient data fusion and collaborative analysis capabilities, low real-time and accuracy of early warning, and low level of intelligent emergency decision-making in existing technologies.

[0006] In view of this, the first aspect of the present invention provides a method for safety management and control of confined space operations, the method comprising:

[0007] In confined spaces under control, environmental parameters, personnel status, and equipment status are collected simultaneously to obtain multimodal data.

[0008] The multimodal data is normalized, and a fusion feature vector containing environmental features, physiological features, and behavioral features is constructed based on the normalized data.

[0009] The fused feature vector is input into a preset intelligent analysis model for real-time analysis to identify pre-operation risks, dynamic risks during operation, and predict the safety status of personnel, and output the corresponding risk identification results and the warning level of the safety status.

[0010] Based on the warning level of the safety status, the corresponding graded warning action is triggered. At the same time, the risk identification result is matched with the preset accident case database to generate an emergency decision-making plan and execute a linkage response.

[0011] Optionally, the environmental parameters include oxygen concentration, toxic gas concentration, temperature, and humidity; the personnel status includes physiological indicators obtained through wearable devices and behavioral postures obtained through video recognition technology; the equipment status includes the operating status of ventilation equipment, equipment location information, and the wearing status of personal protective equipment.

[0012] Optionally, the normalization process for the multimodal data includes: using a Min-Max scaling algorithm to normalize the multimodal data to the [0,1] interval.

[0013] Optionally, the step of inputting the fused feature vector into a preset intelligent analysis model for real-time analysis, identifying pre-operation risks and dynamic risks during operation, predicting the safety status of personnel, and outputting the corresponding risk identification results and safety status warning levels includes:

[0014] The fused feature vector is input into the improved density peak clustering model to perform weighted cluster analysis on the ventilation duration and gas concentration parameters before the operation, identify the risks before the operation and generate the corresponding risk identification results.

[0015] The fused feature vector is input into the Laplacian feature map adaptive density peak clustering model to perform dynamic clustering analysis, identify dynamic risks in the operation, and generate corresponding risk identification results.

[0016] The fused feature vector is input into a backpropagation neural network model optimized by particle swarm optimization to predict the safety status of personnel and output the warning level of the safety status, which includes four levels: no warning, minor warning, medium warning, and severe warning.

[0017] Optionally, the Laplacian eigenmap adaptive density peak clustering model uses a sliding window mechanism to update the data neighborhood and dynamically adjusts the weights of different feature parameters based on real-time collected environmental parameters.

[0018] Optionally, the tiered early warning action includes:

[0019] The corresponding level four response to the absence of alarms is: data archiving and periodic inspection;

[0020] The three-level response corresponding to the aforementioned minor alarm is: triggering a voice prompt and initiating a device retest;

[0021] The secondary response corresponding to the aforementioned central alarm is: triggering an audible and visual alarm and activating the ventilation system;

[0022] The first-level response corresponding to the aforementioned serious alarm is as follows: trigger the buzzer alarm, start the emergency ventilation, and simultaneously send an alarm to the medical system.

[0023] Optionally, the step of matching the risk identification results with a preset accident case database to generate an emergency decision-making plan and execute a coordinated response includes:

[0024] Using a case-based reasoning method, with the risk identification results as input, and combining entropy weight method, grey relational analysis method, and approximation ideal solution ranking method, similar cases are matched from the preset accident case library to generate the optimal rescue strategy as an emergency decision-making plan. Based on the emergency decision-making plan, response actions such as equipment linkage and information push are executed.

[0025] A second aspect of the present invention provides a confined space operation safety management system, the system comprising:

[0026] The data acquisition unit is used to simultaneously collect environmental parameters, personnel status, and equipment status in a confined space operation site to be managed, thereby obtaining multimodal data.

[0027] The preprocessing unit is used to normalize the multimodal data and construct a fusion feature vector containing environmental features, physiological features and behavioral features based on the normalized data.

[0028] The identification unit is used to input the fused feature vector into a preset intelligent analysis model for real-time analysis, identify pre-operation risks and dynamic risks during operation, predict the safety status of personnel, and output the corresponding risk identification results and the warning level of the safety status.

[0029] The generation unit is used to trigger corresponding graded warning actions based on the warning level of the safety status, and at the same time, to match the risk identification results with a preset accident case library to generate an emergency decision-making plan and execute a linkage response.

[0030] A third aspect of the present invention provides a confined space operation safety management device, the device comprising a processor and a memory:

[0031] The memory is used to store program code and transmit the program code to the processor;

[0032] The processor is used to execute the steps of the confined space operation safety management method as described in the first aspect above, according to the instructions in the program code.

[0033] A fourth aspect of the present invention provides a computer-readable storage medium for storing program code for executing the confined space operation safety management method described in the first aspect above.

[0034] As can be seen from the above technical solutions, the present invention has the following advantages:

[0035] This invention provides a confined space operation safety management method that integrates environmental, personnel, and equipment information through multimodal data fusion to construct a multi-dimensional feature dataset, eliminating the limitations of single-parameter monitoring. In the pre-defined intelligent analysis model, an improved density peak clustering model achieves real-time clustering identification of violations through dynamic weight adjustment of key parameters and a sliding window neighborhood update mechanism of the Laplace eigenmap adaptive density peak clustering model. A backpropagation neural network model optimized by particle swarm optimization avoids local extrema problems with the help of an adaptive mutation operator, significantly improving the generalization performance of the early warning model. A multi-level linkage emergency mechanism integrates a four-level response system and a case-based reasoning decision-making model, automatically generating rescue strategies from anomaly warnings to emergency response, achieving rapid response and effectively reducing the risk of secondary accidents. This solves the problems of insufficient data fusion and collaborative analysis capabilities, low real-time warning accuracy, and low level of intelligent emergency decision-making in existing technologies. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1 A flowchart illustrating a method for safety management of confined space operations provided in an embodiment of the present invention;

[0038] Figure 2 This is a schematic diagram of a confined space operation safety management system provided in an embodiment of the present invention. Detailed Implementation

[0039] To address the shortcomings of existing technologies, such as insufficient data fusion and collaborative analysis capabilities, low real-time and accuracy of early warnings, and low level of intelligent emergency decision-making, this invention provides specific embodiments of a confined space operation safety management method, system, equipment, and medium, as detailed below.

[0040] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0041] Please see Figure 1 The present invention provides a method for safety management of confined space operations, comprising:

[0042] Step 101: At the confined space operation site to be controlled, environmental parameters, personnel status, and equipment status are collected simultaneously to obtain multimodal data;

[0043] It should be noted that in confined space operations, multiple sensors (such as gas sensors, smart bracelets, UWB positioning, and video recognition devices) are integrated to simultaneously collect multimodal data, including environmental parameters (such as oxygen concentration, toxic gas concentration, temperature, and humidity), personnel status (such as heart rate, blood oxygen, and posture), and equipment status (such as ventilation equipment current, positioning information, and safety equipment wearing status). For example, at the hardware layer, portable individual soldier devices are used as carriers to deploy various sensors in narrow working environments, collecting data in real time via Bluetooth or Wi-Fi, and performing preliminary processing by an edge computing module. The collected data is then transmitted to the analysis layer for subsequent normalization and feature fusion. This step achieves simultaneous acquisition and time alignment of multidimensional information.

[0044] Step 102: Normalize the multimodal data and construct a fusion feature vector containing environmental features, physiological features and behavioral features based on the normalized data;

[0045] It should be noted that the collected multimodal data (such as oxygen concentration, heart rate, equipment current, etc., each with different dimensions and numerical ranges) undergoes standardization and scaling. Typically, the Min-Max normalization method is used to uniformly map them to the [0,1] interval to eliminate the impact of magnitude differences on subsequent analysis. Then, based on this, key indicators are extracted and combined according to preset feature dimensions (e.g., 7-dimensional environment, 8-dimensional physiology, 8-dimensional behavior) to form a fusion feature vector (e.g., a 23-dimensional vector) containing environmental, physiological, and behavioral information. For example, the minimum and maximum values ​​are first calculated for each type of data, and then a linear transformation is performed on each original data to obtain a normalized value. Afterwards, the normalized environmental, physiological, and behavioral features are sequentially concatenated to construct a fusion feature vector that can be directly input into the intelligent analysis model. This step transforms heterogeneous, multi-source data into a structured representation with a unified scale, preserving key information from each dimension while improving the training efficiency and prediction accuracy of subsequent clustering, classification, and other algorithm models.

[0046] Step 103: Input the fused feature vector into the preset intelligent analysis model for real-time analysis, identify pre-operation risks and dynamic risks during operation, predict the safety status of personnel, and output the corresponding risk identification results and the warning level of the safety status.

[0047] It should be noted that the constructed fusion feature vector is input into a pre-trained intelligent analysis model. The model performs real-time calculations and inferences to identify pre-operation risks (such as insufficient ventilation and abnormal gas concentration), dynamic risks during operation (such as boundary violations and sudden environmental changes), and predict the current safety status of personnel (such as abnormal heart rate and postural collapse). Finally, it outputs specific risk identification results (such as violation type and risk level) and corresponding safety status warning levels (such as no warning, minor warning, moderate warning, and severe warning). For example: First, an improved density peak clustering model (A-DPC) is used to perform weighted clustering of pre-operation ventilation duration and gas concentration parameters to determine whether the safety access conditions are met. Then, a Laplace eigenmapped adaptive density peak clustering model (LA-DPC) combined with a sliding window mechanism and dynamic weight adjustment is used to monitor the dynamic characteristics of position, environment, and behavior during the operation in real time, identifying boundary violations or abnormal events. Simultaneously, a backpropagation neural network optimized by particle swarm optimization (PSO-BP) is used to perform nonlinear mapping of personnel's physiological and behavioral characteristics, outputting four safety status levels. This step enables automatic risk identification across the entire lifecycle and multiple dimensions from pre-operation to operation, shortening early warning response time and distinguishing between different levels of threats.

[0048] Step 104: Based on the warning level of the safety status, trigger the corresponding graded warning action, and at the same time, match the risk identification results with the preset accident case library to generate an emergency decision plan and execute the linkage response.

[0049] It should be noted that, based on the safety status warning level (e.g., no alarm, minor alarm, moderate alarm, severe alarm) output in step 103, the corresponding level of warning action is automatically triggered. Simultaneously, the risk identification results (e.g., accident type, abnormal parameters) are matched with a pre-set accident case database to generate a targeted emergency decision-making plan, and the on-site equipment is linked to execute the corresponding response. The specific process consists of two parallel stages: first, the execution of the warning action. According to the four-level response system, in the absence of an alarm, only data archiving and periodic inspections are performed; in the case of a minor alarm, a voice prompt is triggered and equipment retesting is initiated; in the case of a moderate alarm, an audible and visual alarm is activated and the ventilation system is linked; in the case of a severe alarm, a buzzer alarm is triggered, emergency ventilation is automatically activated, and alarm information is simultaneously pushed to the medical system and monitoring terminal. Second, emergency decision generation employs a case-based reasoning approach. Using current risk identification results (such as gas concentration, personnel status, and accident type) as input, it combines entropy weighting to determine the weights of each decision indicator (e.g., a response timeliness weight of 0.3), grey relational analysis to calculate similarity with historical cases, and the Top-Approximation-Ideal-Solution Ranking (TOPSIS) method to rank candidate rescue strategies. This selects the optimal rescue strategy (such as ventilation, rescue routes, and medical coordination) and automatically executes equipment linkage and information push notifications. This step achieves fully automated closed-loop control from risk warning to tiered response, and from accident identification to intelligent rescue, shortening the overall emergency response time and reducing delays in manual decision-making and the risk of secondary accidents.

[0050] In one embodiment, in step 101, the environmental parameters include oxygen concentration, toxic gas concentration, temperature, and humidity; the personnel status includes physiological indicators obtained through wearable devices and behavioral postures obtained through video recognition technology; and the equipment status includes the operating status of ventilation equipment, the location information of the equipment, and the wearing status of personal protective equipment.

[0051] It should be noted that the environmental parameters specifically include: (Normal range: 18%-23%) (Normal range: 0-50ppm) (Normal range: 0-5% Vol), CO (Normal range: 0-500 ppm) (0-5000ppm), temperature (normal range: -20℃~50℃), humidity (normal range: 0%~100%RH), SF6 (normal range: 0-1000ppm), etc., are collected in real time by a gas sensor (accuracy ±1%).

[0052] The specific status of the personnel includes: heart rate (normal range: 60-120 bpm), blood oxygen saturation (normal range: 90%-100%), and body temperature, which are obtained through a smart bracelet (sampling frequency 10Hz); posture (normal range: standing / lying down 10 states), which are obtained through video recognition using YOLO-pose video recognition technology.

[0053] The specific equipment status includes: ventilation equipment current (normal range: 0-20A), UWB (Ultra-Wideband) positioning coordinates (accuracy ±10cm), and safety equipment wearing status (such as safety helmet, respirator, etc.).

[0054] In practical implementation, the aforementioned sensors used to acquire multimodal data can be integrated into a portable individual soldier device, supporting Bluetooth / Wi-Fi data transmission, thus adapting to deployment in confined spaces.

[0055] It is understood that the environmental parameters in this embodiment cover key indicators such as oxygen, hydrogen sulfide, methane, carbon monoxide, carbon dioxide, sulfur hexafluoride, temperature, and humidity. These are monitored in real time by gas sensors with an accuracy of ±1%, and the normal ranges for each parameter are provided. Personnel status includes physiological indicators such as heart rate (60-120 bpm), blood oxygen saturation (90%-100%), and body temperature, which are acquired by a smart bracelet with a sampling frequency of 10 Hz. YOLO-pose video recognition technology is used to identify 10 body postures, including standing and lying down. Equipment status involves ventilation equipment current (0-20A), high-precision positioning coordinates based on UWB (±10cm), and the wearing status of personal protective equipment such as helmets and respirators. In actual deployment, the above sensors can be integrated into a lightweight, portable individual soldier device that supports Bluetooth or Wi-Fi wireless data transmission, allowing for flexible deployment in narrow, high-risk confined spaces such as cable wells and transformer interiors.

[0056] Note: YOLO-Pose is a heatmap-free 2D multi-person real-time pose estimation technique based on the YOLO object detection framework. It abandons the traditional two-stage process of first generating a heatmap and then predicting keypoints, adopting a single-stage end-to-end design. It directly regresses the human bounding box and skeletal keypoint coordinates within the model and directly optimizes this by introducing the Target Keypoint Similarity (OKS) loss function. This design allows the model to simultaneously complete human detection and pose recognition in a single forward propagation, combining the advantages of both top-down and bottom-up methods. It significantly improves inference speed, eliminates complex post-processing grouping steps, and offers high accuracy and ease of deployment, making it particularly suitable for real-time video stream analysis and edge computing scenarios.

[0057] In one embodiment, step 102, normalizing the multimodal data, includes: using a Min-Max scaling algorithm to normalize the multimodal data to the [0,1] interval.

[0058] It should be noted that, firstly, regarding environmental parameters (such as...) concentration, Multimodal data of different types, such as concentration, personnel physiological indicators (e.g., heart rate, blood oxygen), and equipment status (e.g., fan current), are processed to determine the minimum and maximum values ​​of each feature dimension. Then, a linear transformation is performed on the original data to ensure the transformed data values ​​fall within the range [0,1]. This eliminates the influence of differences in dimensions and magnitudes between different data sources, achieving standardized processing of multi-source data and providing a unified data foundation for feature engineering and algorithm model input in the subsequent intelligent fusion analysis layer. Next, a fusion feature vector containing environmental, physiological, and behavioral features is constructed based on the normalized data. For example, a 23-dimensional feature vector is constructed (7 dimensions of environmental features + 8 dimensions of physiological features + 8 dimensions of behavioral features, combined sequentially to form a vector containing 23 feature values).

[0059] Understandably, this embodiment first employs a Min-Max scaling algorithm to determine the minimum and maximum values ​​for each feature dimension of different types of multimodal data, such as environmental parameters (e.g., oxygen concentration), personnel physiological indicators (e.g., heart rate, blood oxygen), and equipment status (e.g., fan current). Then, a linear transformation is used to uniformly map the original data to the [0,1] interval, thereby eliminating the influence of differences in scale and magnitude and making multi-source data comparable. Based on this, the normalized data is combined according to preset dimensions, for example, constructing a 23-dimensional fusion feature vector, where environmental features occupy 7 dimensions (e.g., gas concentration, temperature, humidity), physiological features occupy 8 dimensions (e.g., heart rate, blood oxygen), and behavioral features occupy 8 dimensions (e.g., posture, positioning), and these are sequentially concatenated to form a complete feature vector for subsequent model input. This process achieves standardized processing and structured fusion of multimodal data.

[0060] Note: The Min-Max scaling algorithm is a linear data normalization method that performs a linear transformation on the original data, proportionally compressing the values ​​of each feature dimension to a preset target range (usually [0,1]). Specifically, the algorithm determines the scaling ratio based on the minimum and maximum values ​​of the feature dimension, mapping the minimum value in the original data to the lower limit of the range, the maximum value to the upper limit, and the remaining values ​​proportionally distributed within the range. This algorithm effectively eliminates the influence of differences in units and orders of magnitude between different features, making multi-source data comparable while preserving the original data distribution. It is suitable for data scenarios with known boundaries or relatively stable data, and is often used in the preprocessing stage before multimodal data fusion.

[0061] In one embodiment, step 103 includes:

[0062] Step 1031: Input the fused feature vector into the improved density peak clustering model, perform weighted cluster analysis on the ventilation duration and gas concentration parameters before the operation, identify the risks before the operation and generate the corresponding risk identification results;

[0063] It should be noted that, firstly, data on ventilation duration (e.g., whether it is less than 30 minutes) and gas concentration (e.g., whether the oxygen concentration is below 19.5% and the hydrogen sulfide concentration is above 10 ppm) are weighted, with oxygen concentration given a 1.5x weight to highlight its importance. Next, an improved density peak clustering model (A-DPC algorithm) is used to perform cluster analysis on the weighted data to identify pre-operation risks such as insufficient ventilation and abnormal gas concentrations. Furthermore, targeted rectification suggestions can be generated based on the clustering results, such as extending ventilation time and inspecting gas detection equipment, to ensure that the pre-operation environment meets safety requirements.

[0064] Note: The Improved Density Peak Clustering Model (A-DPC algorithm) is an optimized version developed based on the traditional Density Peak Clustering (DPC) algorithm. Traditional DPC identifies cluster centers by calculating local density and relative distance, but it struggles to adapt to differences in feature importance and complex data distributions. The improved model (A-DPC) in this invention introduces a dynamic weight adjustment mechanism, which assigns higher weights (e.g., 1.5 times) to key feature parameters (such as oxygen concentration) and adaptively adjusts the weight allocation according to actual scenario requirements. This allows for weighted cluster analysis of data such as ventilation duration and gas concentration during pre-operation risk assessment, effectively identifying risks such as insufficient ventilation and abnormal gas concentrations, thus improving clustering accuracy and sensitivity to key safety indicators.

[0065] Step 1032: Input the fused feature vector into the Laplacian Eigenmap adaptive density peak clustering model to perform dynamic clustering analysis, identify dynamic risks in the operation and generate corresponding risk identification results; wherein, the Laplacian Eigenmap adaptive density peak clustering model uses a sliding window mechanism to update the data neighborhood and dynamically adjusts the weights of different feature parameters according to the environmental parameters collected in real time.

[0066] It should be noted that the fused feature vectors are input into the Laplacian feature map adaptive density peak clustering model. This model uses a sliding window mechanism to update the data neighborhood (sliding window mechanism: setting a fixed-size window (such as a 30-second data window) in terms of time or space, and only performing neighborhood calculations on the latest data within the window to reduce redundant computation). At the same time, the weights of different feature parameters are dynamically adjusted according to the environmental parameters collected in real time, thereby performing dynamic clustering analysis to identify dynamic risks in the operation and generate corresponding risk identification results.

[0067] For example, the feature weights can be dynamically adjusted according to the environment (e.g., the weights of relevant parameters are increased by 30% during high summer temperatures), and the weighted data can be clustered in real time using the Laplace eigenmap adaptive density peak clustering model (LA-DPC). This allows for continuous monitoring of whether the personnel's UWB (Ultra-Wideband) positioning coordinates exceed the electronic fence, thereby achieving dynamic control of risks during operations.

[0068] Step 1033: Input the fused feature vector into the backpropagation neural network model optimized by the particle swarm optimization algorithm to predict the safety status of personnel and output the warning level of the safety status. The warning level of the safety status includes four levels: no warning, light warning, medium warning and heavy warning.

[0069] It should be noted that, firstly, the fused feature vector is input into the backpropagation neural network model. An improved particle swarm optimization algorithm (introducing an adaptive mutation operator) is used to optimize the weights and thresholds of the backpropagation neural network model to avoid the problem of traditional backpropagation neural network models easily getting trapped in local extrema, thereby improving the generalization ability of the backpropagation neural network model in small sample scenarios. Subsequently, the optimized backpropagation neural network model performs nonlinear mapping and multiple rounds of iterative training on the input fused feature vector, and finally outputs four warning levels of safety status: no alarm, light alarm, medium alarm, and heavy alarm, realizing real-time monitoring and early warning of personnel life safety status.

[0070] Note: The improved particle swarm optimization (PSO) algorithm is an optimized version based on the standard PSO algorithm, incorporating an adaptive mutation operator. Standard PSO updates particle positions and velocities using individual and group optimum information by simulating bird flock foraging behavior, but it is prone to getting trapped in local optima. The improved algorithm in this invention introduces an adaptive mutation operator, dynamically adjusting the mutation probability based on the particle swarm's aggregation level during iteration. This allows particles to jump when trapped in local optima, enhancing global search capabilities and avoiding premature convergence. This algorithm is used to optimize the initial weights and thresholds of a backpropagation neural network (BPNN), improving the model's generalization ability and prediction accuracy in small sample scenarios.

[0071] Understandably, this embodiment first uses an improved density peak clustering model (A-DPC) to perform weighted clustering of ventilation duration and gas concentration before operation (oxygen concentration weighted by 1.5 times), identifying risks such as insufficient ventilation and abnormal gas and outputting rectification suggestions; then, it uses a Laplace eigenmapped adaptive density peak clustering model (LA-DPC) with the help of a sliding window (e.g., 30 seconds) and dynamic weight adjustment (e.g., increasing the weight by 30% in summer high temperatures) to monitor dynamic risks such as exceeding boundaries during operation in real time; finally, it uses a BP neural network optimized by particle swarm optimization (PSO-BP) to avoid local extrema through an adaptive mutation operator, outputting four categories of personnel safety status: no alarm, light alarm, moderate alarm, and severe alarm, thereby completing risk identification and early warning throughout the entire cycle before, during, and after operation.

[0072] In one embodiment, step 104 includes the following:

[0073] The corresponding level four response to the absence of alarms is: data archiving and periodic inspection;

[0074] The three-level response corresponding to the aforementioned minor alarm is: triggering a voice prompt and initiating a device retest;

[0075] The secondary response corresponding to the aforementioned central alarm is: triggering an audible and visual alarm and activating the ventilation system;

[0076] The first-level response corresponding to the aforementioned serious alarm is as follows: trigger the buzzer alarm, start the emergency ventilation, and simultaneously send an alarm to the medical system.

[0077] It should be noted that the criteria for no alarm obtained in step 103 can be all parameters being normal; the criteria for a minor alarm can be heart rate >100 bpm or blood oxygen <95%; the criteria for a moderate alarm can be CO >500 ppm or a person crossing the boundary for 30 seconds; and the criteria for a severe alarm can be a person fainting or Less than 18%; the above-mentioned criteria can be set by those skilled in the art as needed, and will not be elaborated here.

[0078] It is understood that this embodiment defines a four-level early warning mechanism: no alarm (all parameters are normal) corresponds to a level four response, which only involves data archiving and periodic inspections; a minor alarm (such as heart rate > 100 bpm or blood oxygen < 95%) corresponds to a level three response, which triggers a voice prompt and initiates equipment retesting; a moderate alarm (such as CO > 500 ppm or personnel crossing the boundary for 30 seconds) corresponds to a level two response, which activates an audible and visual alarm and links the ventilation system; a severe alarm (such as personnel fainting or oxygen < 18%) corresponds to a level one response, which triggers a buzzer alarm, automatically activates emergency ventilation, and simultaneously sends an alarm to the medical system, thereby executing differentiated automatic control and alarm actions according to the risk level.

[0079] In one embodiment, step 104 involves matching the risk identification results with a pre-set accident case library to generate an emergency decision-making plan and execute a coordinated response. This includes: using a case-based reasoning method, taking the risk identification results as input, and combining entropy weighting, grey relational analysis, and approximation of ideal solution ranking methods to match similar cases from the pre-set accident case library, generating an optimal rescue strategy as an emergency decision-making plan, and executing equipment linkage and information push response actions according to the emergency decision-making plan.

[0080] It should be noted that, firstly, a case library is constructed. Specifically, a database containing over 50 typical confined space accident cases is established. Each case includes environmental parameters (such as gas concentration, temperature and humidity), personnel status (heart rate, posture), accident type (poisoning / hypoxia / collapse, etc.), and response plan (such as ventilation equipment activation and rescue route planning).

[0081] Next, the weights of the indicators are determined. Specifically, key decision indicators (such as response timeliness, rescue resource input, environmental risk level, etc.) are extracted from the case. The objective weight of each indicator is quantified using the entropy weight method. For example, "response timeliness weight 0.3" means that this indicator accounts for 30% of the importance in the decision. It can be understood that the smaller the entropy value of the indicator, the lower the information uncertainty, and the higher the weight.

[0082] Then, match case similarity, specifically: using real-time data of the current accident (such as...) The concentration (18%) and the state of unconsciousness of the personnel are consistent with the historical data in the case database. The correlation between the current accident and each historical case is calculated (value range 0-1). The closer the correlation is to 1, the higher the case similarity. For example, if the current accident is a "cable well..." "If a leak leads to poisoning of personnel," the system will prioritize matching similar cases in the database.

[0083] Finally, the optimal strategy is generated, specifically: selecting the top N cases with the highest similarity for handling.

[0084] Construct "ideal solutions" (optimal values ​​for each indicator) and "negative ideal solutions" (worst values ​​for each indicator). Calculate the Euclidean distance between each candidate solution and the ideal solution; the smaller the distance, the better the solution. Generate the optimal rescue strategy after ranking.

[0085] Understandably, this embodiment first establishes a case library containing more than 50 typical accidents (including environmental parameters, personnel status, accident type, and response plan); then, it determines the objective weight of each decision indicator using the entropy weight method (e.g., response timeliness weight of 0.3); next, it uses grey relational analysis to calculate the similarity between the current accident and historical cases (the closer the value is to 1, the more similar the case); finally, it uses the Top-Approximation-Ideal-Solution Ranking (TOPSIS) method to rank the response plans of similar cases, selects the plan closest to the ideal solution as the optimal rescue strategy, and executes equipment linkage and information push accordingly to generate emergency decisions.

[0086] In summary, the confined space operation safety management method provided by this invention integrates environmental, personnel, and equipment information through multimodal data fusion, constructing a multi-dimensional feature dataset to eliminate the limitations of single-parameter monitoring. In the pre-defined intelligent analysis model, the improved density peak clustering model achieves real-time clustering identification of violations through dynamic weight adjustment of key parameters and the sliding window neighborhood update mechanism of the Laplace eigenmap adaptive density peak clustering model. The backpropagation neural network model optimized by the particle swarm optimization algorithm avoids local extrema problems with the help of an adaptive mutation operator, significantly improving the generalization performance of the early warning model. The multi-level linkage emergency mechanism integrates a four-level response system and a case-based reasoning decision-making model, automatically generating rescue strategies from anomaly warnings, achieving rapid response and effectively reducing the risk of secondary accidents. This solves the problems of insufficient data fusion and collaborative analysis capabilities, low real-time warning accuracy, and low level of intelligent emergency decision-making in existing technologies.

[0087] The above is a confined space operation safety management method provided in the embodiments of the present invention. The following is a confined space operation safety management system provided in the embodiments of the present invention.

[0088] Please see Figure 2 The confined space operation safety management system provided in this embodiment of the invention includes:

[0089] The data acquisition unit 201 is used to simultaneously collect environmental parameters, personnel status, and equipment status at the confined space operation site to be managed, and obtain multimodal data.

[0090] It should be noted that in confined space operations, the data acquisition unit 201 integrates multiple sensors (such as gas sensors, smart bracelets, UWB positioning, and video recognition devices) to simultaneously collect multimodal data, including environmental parameters (such as oxygen concentration, toxic gas concentration, temperature, and humidity), personnel status (such as heart rate, blood oxygen, and posture), and equipment status (such as ventilation equipment current, positioning information, and safety equipment wearing status). For example, the hardware layer uses portable individual soldier devices as carriers, deploying various sensors in narrow working environments to collect data in real time via Bluetooth or Wi-Fi, and performing preliminary processing by the edge computing module. The collected data is then transmitted to the analysis layer for subsequent normalization and feature fusion. This unit achieves simultaneous acquisition and time alignment of multidimensional information.

[0091] The preprocessing unit 202 is used to normalize the multimodal data and construct a fusion feature vector containing environmental features, physiological features and behavioral features based on the normalized data.

[0092] It should be noted that the preprocessing unit 202 performs standardization and scaling on the collected multimodal data (such as oxygen concentration, heart rate, equipment current, etc., each with different dimensions and numerical ranges). Typically, the Min-Max normalization method is used to uniformly map them to the [0,1] interval to eliminate the impact of magnitude differences on subsequent analysis. Then, based on this, key indicators are extracted and combined according to preset feature dimensions (e.g., 7-dimensional environmental, 8-dimensional physiological, and 8-dimensional behavioral), forming a fusion feature vector (e.g., a 23-dimensional vector) containing environmental, physiological, and behavioral information. For example, the minimum and maximum values ​​are first calculated for each type of data, and then a linear transformation is performed on each original data to obtain a normalized value. Afterward, the normalized environmental, physiological, and behavioral features are sequentially concatenated to construct a fusion feature vector that can be directly input into the intelligent analysis model.

[0093] The identification unit 203 is used to input the fused feature vector into the preset intelligent analysis model for real-time analysis, identify pre-operation risks and dynamic risks during operation, predict the safety status of personnel, and output the corresponding risk identification results and the warning level of the safety status.

[0094] It should be noted that the identification unit 203 inputs the constructed fusion feature vector into the pre-trained intelligent analysis model. Through multiple algorithm modules within the model, it performs real-time calculations and inferences to identify pre-operation risks (such as insufficient ventilation, abnormal gas concentration), dynamic risks during operation (such as boundary crossing behavior, sudden environmental changes), and predict the current safety status of personnel (such as abnormal heart rate, postural collapse, etc.). Finally, it outputs specific risk identification results (such as violation type, risk level) and corresponding safety status warning levels (such as no alarm, minor alarm, moderate alarm, severe alarm). For example: First, the improved density peak clustering model (A-DPC) is used to perform weighted clustering of pre-operation ventilation duration and gas concentration parameters to determine whether the safety access conditions are met; then, the Laplace eigenmapped adaptive density peak clustering model (LA-DPC) is used in combination with a sliding window mechanism and dynamic weight adjustment to monitor the dynamic characteristics of position, environment, and behavior during the operation in real time and identify boundary crossing or abnormal events; at the same time, a backpropagation neural network (PSO-BP) optimized by particle swarm optimization is used to perform nonlinear mapping of the physiological and behavioral characteristics of personnel, outputting four safety status levels. This unit enables automatic risk identification across the entire lifecycle and multiple dimensions from pre-operation to operation, shortening early warning response time and distinguishing between different levels of threats.

[0095] The generation unit 204 is used to trigger corresponding graded warning actions based on the warning level of the safety status, and at the same time, it matches the risk identification results with the preset accident case library to generate an emergency decision plan and execute the linkage response.

[0096] It should be noted that, based on the safety status warning level output by unit 103 (such as no alarm, minor alarm, moderate alarm, and severe alarm), the corresponding level of warning action is automatically triggered. At the same time, the risk identification results (such as accident type, abnormal parameters, etc.) are matched with the preset accident case database to generate a targeted emergency decision-making plan and link the on-site equipment to execute the corresponding response. The specific process is divided into two parallel stages: the first is the execution of the warning action, that is, according to the four-level response system, when there is no alarm, only data archiving and timed inspection are performed; when there is a minor alarm, a voice prompt is triggered and the equipment is retested; when there is a moderate alarm, an audible and visual alarm is activated and the ventilation system is linked; when there is a severe alarm, a buzzer alarm is triggered, emergency ventilation is automatically turned on, and the alarm information is simultaneously pushed to the medical system and monitoring terminal. Second, emergency decision generation employs a case-based reasoning approach. Using current risk identification results (such as gas concentration, personnel status, and accident type) as input, it combines entropy weighting to determine the weights of each decision indicator (e.g., a response timeliness weight of 0.3), grey relational analysis to calculate similarity with historical cases, and the Top-Approximation-Ideal-Solution Ranking (TOPSIS) method to rank candidate rescue strategies. This selects the optimal rescue strategy (such as ventilation, rescue routes, and medical coordination) and automatically executes equipment linkage and information push notifications. This unit achieves fully automated closed-loop control from risk warning to tiered response, and from accident identification to intelligent rescue, shortening the overall emergency response time and reducing delays in manual decision-making and the risk of secondary accidents.

[0097] In one embodiment, the environmental parameters in the data acquisition unit 201 include oxygen concentration, toxic gas concentration, temperature, and humidity; the personnel status includes physiological indicators obtained through wearable devices and behavioral postures obtained through video recognition technology; and the equipment status includes the operating status of ventilation equipment, the location information of the equipment, and the wearing status of personal protective equipment.

[0098] It should be noted that the environmental parameters specifically include: (Normal range: 18%-23%) (Normal range: 0-50ppm) (Normal range: 0-5% Vol), CO (Normal range: 0-500 ppm) (0-5000ppm), temperature (normal range: -20℃~50℃), humidity (normal range: 0%~100%RH), SF6 (normal range: 0-1000ppm), etc., are collected in real time by a gas sensor (accuracy ±1%).

[0099] The specific status of the personnel includes: heart rate (normal range: 60-120 bpm), blood oxygen saturation (normal range: 90%-100%), and body temperature, which are obtained through a smart bracelet (sampling frequency 10Hz); posture (normal range: standing / lying down 10 states), which are obtained through video recognition using YOLO-pose video recognition technology.

[0100] The specific equipment status includes: ventilation equipment current (normal range: 0-20A), UWB (Ultra-Wideband) positioning coordinates (accuracy ±10cm), and safety equipment wearing status (such as safety helmet, respirator, etc.).

[0101] In practical implementation, the aforementioned sensors used to acquire multimodal data can be integrated into a portable individual soldier device, supporting Bluetooth / Wi-Fi data transmission, thus adapting to deployment in confined spaces.

[0102] It is understood that the environmental parameters in this embodiment cover key indicators such as oxygen, hydrogen sulfide, methane, carbon monoxide, carbon dioxide, sulfur hexafluoride, temperature, and humidity. These are monitored in real time by gas sensors with an accuracy of ±1%, and the normal ranges for each parameter are provided. Personnel status includes physiological indicators such as heart rate (60-120 bpm), blood oxygen saturation (90%-100%), and body temperature, which are acquired by a smart bracelet with a sampling frequency of 10 Hz. YOLO-pose video recognition technology is used to identify 10 body postures, including standing and lying down. Equipment status involves ventilation equipment current (0-20A), high-precision positioning coordinates based on UWB (±10cm), and the wearing status of personal protective equipment such as helmets and respirators. In actual deployment, the above sensors can be integrated into a lightweight, portable individual soldier device that supports Bluetooth or Wi-Fi wireless data transmission, allowing for flexible deployment in narrow, high-risk confined spaces such as cable wells and transformer interiors.

[0103] Note: YOLO-Pose is a heatmap-free 2D multi-person real-time pose estimation technique based on the YOLO object detection framework. It abandons the traditional two-stage process of first generating a heatmap and then predicting keypoints, adopting a single-stage end-to-end design. It directly regresses the human bounding box and skeletal keypoint coordinates within the model and directly optimizes this by introducing the Target Keypoint Similarity (OKS) loss function. This design allows the model to simultaneously complete human detection and pose recognition in a single forward propagation, combining the advantages of both top-down and bottom-up methods. It significantly improves inference speed, eliminates complex post-processing grouping steps, and offers high accuracy and ease of deployment, making it particularly suitable for real-time video stream analysis and edge computing scenarios.

[0104] In one embodiment, the preprocessing unit 202 performs normalization processing on the multimodal data, including: using the Min-Max scaling algorithm to normalize the multimodal data to the [0,1] interval.

[0105] It should be noted that, firstly, regarding environmental parameters (such as...) concentration, Multimodal data of different types, such as concentration, personnel physiological indicators (e.g., heart rate, blood oxygen), and equipment status (e.g., fan current), are processed to determine the minimum and maximum values ​​of each feature dimension. Then, a linear transformation is performed on the original data to ensure the transformed data values ​​fall within the range [0,1]. This eliminates the influence of differences in dimensions and magnitudes between different data sources, achieving standardized processing of multi-source data and providing a unified data foundation for feature engineering and algorithm model input in the subsequent intelligent fusion analysis layer. Next, a fusion feature vector containing environmental, physiological, and behavioral features is constructed based on the normalized data. For example, a 23-dimensional feature vector is constructed (7 dimensions of environmental features + 8 dimensions of physiological features + 8 dimensions of behavioral features, combined sequentially to form a vector containing 23 feature values).

[0106] Understandably, this embodiment first employs a Min-Max scaling algorithm to determine the minimum and maximum values ​​for each feature dimension of different types of multimodal data, such as environmental parameters (e.g., oxygen concentration), personnel physiological indicators (e.g., heart rate, blood oxygen), and equipment status (e.g., fan current). Then, a linear transformation is used to uniformly map the original data to the [0,1] interval, thereby eliminating the influence of differences in scale and magnitude and making multi-source data comparable. Based on this, the normalized data is combined according to preset dimensions, for example, constructing a 23-dimensional fusion feature vector, where environmental features occupy 7 dimensions (e.g., gas concentration, temperature, humidity), physiological features occupy 8 dimensions (e.g., heart rate, blood oxygen), and behavioral features occupy 8 dimensions (e.g., posture, positioning), and these are sequentially concatenated to form a complete feature vector for subsequent model input. This process achieves standardized processing and structured fusion of multimodal data.

[0107] Note: The Min-Max scaling algorithm is a linear data normalization method that performs a linear transformation on the original data, proportionally compressing the values ​​of each feature dimension to a preset target range (usually [0,1]). Specifically, the algorithm determines the scaling ratio based on the minimum and maximum values ​​of the feature dimension, mapping the minimum value in the original data to the lower limit of the range, the maximum value to the upper limit, and the remaining values ​​proportionally distributed within the range. This algorithm effectively eliminates the influence of differences in units and orders of magnitude between different features, making multi-source data comparable while preserving the original data distribution. It is suitable for data scenarios with known boundaries or relatively stable data, and is often used in the preprocessing stage before multimodal data fusion.

[0108] In one embodiment, the identification unit 203 is specifically used for:

[0109] First, the fused feature vector is input into the improved density peak clustering model to perform weighted clustering analysis on the ventilation duration and gas concentration parameters before the operation, identify the risks before the operation and generate the corresponding risk identification results;

[0110] It should be noted that, firstly, data on ventilation duration (e.g., whether it is less than 30 minutes) and gas concentration (e.g., whether the oxygen concentration is below 19.5% and the hydrogen sulfide concentration is above 10 ppm) are weighted, with oxygen concentration given a 1.5x weight to highlight its importance. Next, an improved density peak clustering model (A-DPC algorithm) is used to perform cluster analysis on the weighted data to identify pre-operation risks such as insufficient ventilation and abnormal gas concentrations. Furthermore, targeted rectification suggestions can be generated based on the clustering results, such as extending ventilation time and inspecting gas detection equipment, to ensure that the pre-operation environment meets safety requirements.

[0111] Note: The Improved Density Peak Clustering Model (A-DPC algorithm) is an optimized version developed based on the traditional Density Peak Clustering (DPC) algorithm. Traditional DPC identifies cluster centers by calculating local density and relative distance, but it struggles to adapt to differences in feature importance and complex data distributions. The improved model (A-DPC) in this invention introduces a dynamic weight adjustment mechanism, which assigns higher weights (e.g., 1.5 times) to key feature parameters (such as oxygen concentration) and adaptively adjusts the weight allocation according to actual scenario requirements. This allows for weighted cluster analysis of data such as ventilation duration and gas concentration during pre-operation risk assessment, effectively identifying risks such as insufficient ventilation and abnormal gas concentrations, thus improving clustering accuracy and sensitivity to key safety indicators.

[0112] Then, the fused feature vector is input into the Laplacian eigenmap adaptive density peak clustering model for dynamic clustering analysis to identify dynamic risks in the operation and generate corresponding risk identification results. The Laplacian eigenmap adaptive density peak clustering model uses a sliding window mechanism to update the data neighborhood and dynamically adjusts the weights of different feature parameters according to the environmental parameters collected in real time.

[0113] It should be noted that the fused feature vectors are input into the Laplacian feature map adaptive density peak clustering model. This model uses a sliding window mechanism to update the data neighborhood (sliding window mechanism: setting a fixed-size window (such as a 30-second data window) in terms of time or space, and only performing neighborhood calculations on the latest data within the window to reduce redundant computation). At the same time, the weights of different feature parameters are dynamically adjusted according to the environmental parameters collected in real time, thereby performing dynamic clustering analysis to identify dynamic risks in the operation and generate corresponding risk identification results.

[0114] For example, the feature weights can be dynamically adjusted according to the environment (e.g., the weights of relevant parameters are increased by 30% during high summer temperatures), and the weighted data can be clustered in real time using the Laplace eigenmap adaptive density peak clustering model (LA-DPC). This allows for continuous monitoring of whether the personnel's UWB (Ultra-Wideband) positioning coordinates exceed the electronic fence, thereby achieving dynamic control of risks during operations.

[0115] Finally, the fused feature vectors are input into the backpropagation neural network model optimized by the particle swarm optimization algorithm to predict the safety status of personnel and output the warning level of the safety status. The warning level of the safety status includes four levels: no warning, minor warning, medium warning, and severe warning.

[0116] It should be noted that, firstly, the fused feature vector is input into the backpropagation neural network model. An improved particle swarm optimization algorithm (introducing an adaptive mutation operator) is used to optimize the weights and thresholds of the backpropagation neural network model to avoid the problem of traditional backpropagation neural network models easily getting trapped in local extrema, thereby improving the generalization ability of the backpropagation neural network model in small sample scenarios. Subsequently, the optimized backpropagation neural network model performs nonlinear mapping and multiple rounds of iterative training on the input fused feature vector, and finally outputs four warning levels of safety status: no alarm, light alarm, medium alarm, and heavy alarm, realizing real-time monitoring and early warning of personnel life safety status.

[0117] Note: The improved particle swarm optimization (PSO) algorithm is an optimized version based on the standard PSO algorithm, incorporating an adaptive mutation operator. Standard PSO updates particle positions and velocities using individual and group optimum information by simulating bird flock foraging behavior, but it is prone to getting trapped in local optima. The improved algorithm in this invention introduces an adaptive mutation operator, dynamically adjusting the mutation probability based on the particle swarm's aggregation level during iteration. This allows particles to jump when trapped in local optima, enhancing global search capabilities and avoiding premature convergence. This algorithm is used to optimize the initial weights and thresholds of a backpropagation neural network (BPNN), improving the model's generalization ability and prediction accuracy in small sample scenarios.

[0118] Understandably, this embodiment first uses an improved density peak clustering model (A-DPC) to perform weighted clustering of ventilation duration and gas concentration before operation (oxygen concentration weighted by 1.5 times), identifying risks such as insufficient ventilation and abnormal gas and outputting rectification suggestions; then, it uses a Laplace eigenmapped adaptive density peak clustering model (LA-DPC) with the help of a sliding window (e.g., 30 seconds) and dynamic weight adjustment (e.g., increasing the weight by 30% in summer high temperatures) to monitor dynamic risks such as exceeding boundaries during operation in real time; finally, it uses a BP neural network optimized by particle swarm optimization (PSO-BP) to avoid local extrema through an adaptive mutation operator, outputting four categories of personnel safety status: no alarm, light alarm, moderate alarm, and severe alarm, thereby completing risk identification and early warning throughout the entire cycle before, during, and after operation.

[0119] In one embodiment, in the generation unit 204: the graded early warning action includes:

[0120] The corresponding level four response to the absence of alarms is: data archiving and periodic inspection;

[0121] The three-level response corresponding to the aforementioned minor alarm is: triggering a voice prompt and initiating a device retest;

[0122] The secondary response corresponding to the aforementioned central alarm is: triggering an audible and visual alarm and activating the ventilation system;

[0123] The first-level response corresponding to the aforementioned serious alarm is as follows: trigger the buzzer alarm, start the emergency ventilation, and simultaneously send an alarm to the medical system.

[0124] It should be noted that the criteria for no alarm obtained by the recognition unit 203 can be that all parameters are normal; the criteria for a minor alarm can be that the heart rate is >100 bpm or the blood oxygen is <95%; the criteria for a moderate alarm can be that CO is >500 ppm or a person has crossed the boundary for 30 seconds; and the criteria for a severe alarm can be that a person has fainted or Less than 18%; the above-mentioned criteria can be set by those skilled in the art as needed, and will not be elaborated here.

[0125] It is understood that this embodiment defines a four-level early warning mechanism: no alarm (all parameters are normal) corresponds to a level four response, which only involves data archiving and periodic inspections; a minor alarm (such as heart rate > 100 bpm or blood oxygen < 95%) corresponds to a level three response, which triggers a voice prompt and initiates equipment retesting; a moderate alarm (such as CO > 500 ppm or personnel crossing the boundary for 30 seconds) corresponds to a level two response, which activates an audible and visual alarm and links the ventilation system; a severe alarm (such as personnel fainting or oxygen < 18%) corresponds to a level one response, which triggers a buzzer alarm, automatically activates emergency ventilation, and simultaneously sends an alarm to the medical system, thereby executing differentiated automatic control and alarm actions according to the risk level.

[0126] In one embodiment, the generation unit 204: based on the risk identification results and a preset accident case library, an emergency decision-making plan is generated and a linkage response is executed, including: using a case reasoning method, taking the risk identification results as input, and combining the entropy weight method, grey relational analysis method and the approximation ideal solution ranking method, matching similar cases from the preset accident case library, generating the optimal rescue strategy as an emergency decision-making plan, and executing equipment linkage and information push response actions according to the emergency decision-making plan.

[0127] It should be noted that, firstly, a case library is constructed. Specifically, a database containing over 50 typical confined space accident cases is established. Each case includes environmental parameters (such as gas concentration, temperature and humidity), personnel status (heart rate, posture), accident type (poisoning / hypoxia / collapse, etc.), and response plan (such as ventilation equipment activation and rescue route planning).

[0128] Next, the weights of the indicators are determined. Specifically, key decision indicators (such as response timeliness, rescue resource input, environmental risk level, etc.) are extracted from the case. The objective weight of each indicator is quantified using the entropy weight method. For example, "response timeliness weight 0.3" means that this indicator accounts for 30% of the importance in the decision. It can be understood that the smaller the entropy value of the indicator, the lower the information uncertainty, and the higher the weight.

[0129] Then, match case similarity, specifically: using real-time data of the current accident (such as...) The concentration (18%) and the state of unconsciousness of the personnel are consistent with the historical data in the case database. The correlation between the current accident and each historical case is calculated (value range 0-1). The closer the correlation is to 1, the higher the case similarity. For example, if the current accident is a "cable well..." "If a leak leads to poisoning of personnel," the system will prioritize matching similar cases in the database.

[0130] Finally, the optimal strategy is generated, specifically: selecting the top N cases with the highest similarity for handling.

[0131] Construct "ideal solutions" (optimal values ​​for each indicator) and "negative ideal solutions" (worst values ​​for each indicator). Calculate the Euclidean distance between each candidate solution and the ideal solution; the smaller the distance, the better the solution. Generate the optimal rescue strategy after ranking.

[0132] Understandably, this embodiment first establishes a case library containing more than 50 typical accidents (including environmental parameters, personnel status, accident type, and response plan); then, it determines the objective weight of each decision indicator using the entropy weight method (e.g., response timeliness weight of 0.3); next, it uses grey relational analysis to calculate the similarity between the current accident and historical cases (the closer the value is to 1, the more similar the case); finally, it uses the Top-Approximation-Ideal-Solution Ranking (TOPSIS) method to rank the response plans of similar cases, selects the plan closest to the ideal solution as the optimal rescue strategy, and executes equipment linkage and information push accordingly to generate emergency decisions.

[0133] In summary, the confined space operation safety management system provided by this invention integrates environmental, personnel, and equipment information through multimodal data fusion, constructing a multi-dimensional feature dataset to eliminate the limitations of single-parameter monitoring. In the pre-defined intelligent analysis model, the improved density peak clustering model achieves real-time clustering identification of violations through dynamic weight adjustment of key parameters and the sliding window neighborhood update mechanism of the Laplace feature mapping adaptive density peak clustering model. The backpropagation neural network model optimized by the particle swarm optimization algorithm avoids local extrema problems with the help of an adaptive mutation operator, significantly improving the generalization performance of the early warning model. The multi-level linkage emergency mechanism integrates a four-level response system and a case-based reasoning decision-making model, automatically generating rescue strategies from anomaly warnings, achieving rapid response and effectively reducing the risk of secondary accidents. This solves the problems of insufficient data fusion and collaborative analysis capabilities, low real-time warning accuracy, and low level of intelligent emergency decision-making in existing technologies.

[0134] Furthermore, this embodiment of the invention also provides a confined space operation safety management device, the device including a processor and a memory:

[0135] The memory is used to store program code and transmit the program code to the processor;

[0136] The processor is used to execute the steps of the confined space operation safety management method as described in the above method embodiments, according to the instructions in the program code.

[0137] It should be noted that this embodiment provides a confined space operation safety management device, consisting of a processor and a memory. The memory stores the program code implementing the safety management method and transmits the code to the processor. The processor executes the instructions in the code to fully implement the steps described in the aforementioned method embodiment, including synchronously collecting multimodal data of the environment, personnel, and equipment; normalizing the data and constructing a fusion feature vector; using intelligent analysis models such as improved density peak clustering, Laplace feature mapping adaptive density peak clustering, and particle swarm optimization BP neural network to identify risks and predict personnel safety status before and during operations; outputting warning levels and triggering graded responses (such as voice prompts, audible and visual alarms, ventilation linkage, medical alarms, etc.); and simultaneously generating the optimal emergency decision-making scheme and executing linkage control based on case reasoning and entropy weight-grey relational-TOPSIS methods. Through the collaborative work of the processor and memory, this device can automatically complete the full-cycle, real-time safety management of confined space operations, improving the accuracy of risk identification, the speed of warning response, and the scientific nature of emergency decision-making, thereby effectively reducing the accident rate and secondary risks.

[0138] Furthermore, this embodiment of the invention also provides a computer-readable storage medium for storing program code, which is used to execute the confined space operation safety management method described in the above method embodiments.

[0139] It should be noted that this embodiment provides a computer-readable storage medium, which can be any entity capable of storing program code, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk. The storage medium stores program code for executing the aforementioned confined space operation safety management method. When this program code is read and executed by a processor (such as a computer, server, embedded device, etc.), the following steps are performed sequentially: synchronously collecting multimodal data on the environment, personnel, and equipment at the confined space site; normalizing the data and constructing a fusion feature vector; identifying risks before and during the operation and predicting the personnel safety status through intelligent analysis models such as improved density peak clustering, Laplace feature mapping adaptive density peak clustering, and particle swarm optimization BP neural network; outputting warning levels and triggering corresponding graded responses (such as voice prompts, audible and visual alarms, ventilation linkage, medical alarms, etc.); and simultaneously matching the optimal emergency decision-making scheme and executing linkage control based on case reasoning and the entropy weight-grey relational-TOPSIS method. By loading and running the program code in this storage medium, any device with computing power can achieve fully automated safety management of confined space operations, thereby improving the real-time nature and accuracy of risk warnings and the level of emergency decision-making.

[0140] In the embodiments provided by this invention, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0141] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0142] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0143] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0144] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for safety management and control of confined space operations, characterized in that, include: In confined spaces under control, environmental parameters, personnel status, and equipment status are collected simultaneously to obtain multimodal data. The multimodal data is normalized, and a fusion feature vector containing environmental features, physiological features, and behavioral features is constructed based on the normalized data. The fused feature vector is input into a preset intelligent analysis model for real-time analysis to identify pre-operation risks, dynamic risks during operation, and predict the safety status of personnel, and output the corresponding risk identification results and the warning level of the safety status. Based on the warning level of the safety status, the corresponding graded warning action is triggered. At the same time, the risk identification result is matched with the preset accident case database to generate an emergency decision-making plan and execute a linkage response.

2. The confined space operation safety management method according to claim 1, characterized in that, The environmental parameters include oxygen concentration, toxic gas concentration, temperature, and humidity; the personnel status includes physiological indicators obtained through wearable devices and behavioral postures obtained through video recognition technology; the equipment status includes the operating status of ventilation equipment, equipment location information, and the wearing status of personal protective equipment.

3. The confined space operation safety management method according to claim 1, characterized in that, The normalization process for the multimodal data includes: using a Min-Max scaling algorithm to normalize the multimodal data to the [0,1] interval.

4. The confined space operation safety management method according to claim 1, characterized in that, The process of inputting the fused feature vector into a preset intelligent analysis model for real-time analysis, identifying pre-operation risks and dynamic risks during operation, predicting personnel safety status, and outputting corresponding risk identification results and safety status warning levels includes: The fused feature vector is input into the improved density peak clustering model to perform weighted cluster analysis on the ventilation duration and gas concentration parameters before the operation, identify the risks before the operation and generate the corresponding risk identification results. The fused feature vector is input into the Laplacian feature map adaptive density peak clustering model to perform dynamic clustering analysis, identify dynamic risks in the operation, and generate corresponding risk identification results. The fused feature vector is input into a backpropagation neural network model optimized by particle swarm optimization to predict the safety status of personnel and output the warning level of the safety status, which includes four levels: no warning, minor warning, medium warning, and severe warning.

5. The confined space operation safety management method according to claim 4, characterized in that, The Laplacian eigenmap adaptive density peak clustering model uses a sliding window mechanism to update the data neighborhood and dynamically adjusts the weights of different feature parameters based on real-time collected environmental parameters.

6. The confined space operation safety management method according to claim 5, characterized in that, The tiered early warning actions include: The corresponding level four response to the absence of alarms is: data archiving and periodic inspection; The three-level response corresponding to the aforementioned minor alarm is: triggering a voice prompt and initiating a device retest; The secondary response corresponding to the aforementioned central alarm is: triggering an audible and visual alarm and activating the ventilation system; The first-level response corresponding to the aforementioned serious alarm is as follows: trigger the buzzer alarm, start the emergency ventilation, and simultaneously send an alarm to the medical system.

7. The confined space operation safety management method according to claim 1, characterized in that, The process of matching the risk identification results with a pre-set accident case database to generate an emergency decision-making plan and execute a coordinated response includes: Using a case-based reasoning method, with the risk identification results as input, and combining entropy weight method, grey relational analysis method, and approximation ideal solution ranking method, similar cases are matched from the preset accident case library to generate the optimal rescue strategy as an emergency decision-making plan. Based on the emergency decision-making plan, response actions such as equipment linkage and information push are executed.

8. A confined space operation safety management system, characterized in that, include: The data acquisition unit is used to simultaneously collect environmental parameters, personnel status, and equipment status in a confined space operation site to be managed, thereby obtaining multimodal data. The preprocessing unit is used to normalize the multimodal data and construct a fusion feature vector containing environmental features, physiological features and behavioral features based on the normalized data. The identification unit is used to input the fused feature vector into a preset intelligent analysis model for real-time analysis, identify pre-operation risks and dynamic risks during operation, predict the safety status of personnel, and output the corresponding risk identification results and the warning level of the safety status. The generation unit is used to trigger corresponding graded warning actions based on the warning level of the safety status, and at the same time, to match the risk identification results with a preset accident case library to generate an emergency decision-making plan and execute a linkage response.

9. A confined space operation safety management and control device, characterized in that, The device includes a processor and a memory: The memory is used to store program code and transmit the program code to the processor; The processor is used to execute the confined space operation safety management method according to any one of claims 1-7 according to the instructions in the program code.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store program code for executing the confined space operation safety management method according to any one of claims 1-7.

Images