Fire starting and leaving person early warning method and system based on behavior track and hot area correlation analysis

Abstract

The present application relates to a fire starting person leaving early warning method and system based on behavior track and heat field correlation analysis, belongs to the technical field of electric data analysis and processing, the early warning method comprises: obtaining the personnel position track, heat field distribution and equipment operation state of the fire starting operation area Multidimensional perception data, after pretreatment, realize data fusion with space-time alignment algorithm;Subsequently, a behavior-heat field correlation feature model based on deep learning is constructed, combined with transfer learning and adversarial training mechanism, the correlation feature parameters of the distance change rate and the stay time of the personnel track and the heat field are accurately extracted;According to the danger level of fire starting operation and the real-time environmental parameter, the threshold interval is dynamically set, the feature parameter comparison triggers the graded early warning, and the equipment response and secondary verification mechanism are linked. The present application solves the problems of low early warning accuracy, weak anti-interference ability, unassociated heat field risk and poor timeliness, realizes the accurate and real-time early warning of fire starting person leaving risk, and provides reliable protection for high-risk operation safety.

Description

Technical Field

[0001] This invention belongs to the field of electrical data analysis and processing technology, and specifically relates to a method and system for early warning of hot work and personnel leaving the site based on behavioral trajectory and thermal domain correlation analysis. Background Technology

[0002] Hot work is a common and highly dangerous type of work in industrial production, construction and oil and gas extraction. During the operation, high-temperature flames, sparks and other ignition sources will be generated. If there are flammable and explosive materials in the vicinity (such as oil and gas, dust and flammable materials), it is very easy to cause fires, explosions and other safety accidents, resulting in casualties, equipment damage and major economic losses.

[0003] According to statistics from emergency management departments, nearly 30% of industrial safety accidents in recent years are related to violations of hot work regulations. Among these, "unsupervised personnel leaving their posts during hot work" (i.e., workers leaving their posts without authorization, leaving the ignition source unattended) is one of the core contributing factors to accidents. Therefore, achieving accurate and real-time early warning of personnel leaving their posts during hot work is of vital practical significance for ensuring operational safety and reducing the accident rate.

[0004] To address the safety hazards posed by unauthorized hot work operations, existing technologies have developed several related early warning solutions, which can be broadly categorized into three types: The first type is personnel location-based early warning schemes. These involve wearing positioning terminals (such as RFID tags or GPS locators) on workers, setting up positioning base stations in the hot work area, and acquiring real-time personnel location information. An early warning is triggered when a worker's location exceeds a pre-defined work area. The second type is image recognition schemes based on video surveillance. These utilize surveillance cameras deployed at the work site to collect image data and identify workers using target detection algorithms. If no worker is detected in the surveillance footage within a set time, an early warning is issued. The third type is trigger-based early warning schemes based on personnel-carried equipment. For example, workers are required to carry emergency stop buttons or sensors. When a worker leaves the work area, the signal between the sensor and the equipment is interrupted, triggering equipment shutdown and an early warning. However, current unauthorized hot work operations early warning technologies suffer from low accuracy, weak anti-interference capabilities, inability to correlate with thermal risks, and poor timeliness, making it difficult to effectively ensure on-site safety during hot work operations. Summary of the Invention

[0005] This invention provides a method and system for early warning of hot work leaving personnel based on behavioral trajectory and thermal domain correlation analysis. It addresses the technical problems of existing hot work leaving personnel early warning technologies, such as low accuracy, weak anti-interference capability, inability to correlate with thermal domain risks, and poor timeliness. Based on a deep learning-based behavioral-thermal domain correlation feature model, combined with transfer learning and adversarial training mechanisms, it accurately extracts correlation feature parameters of the distance change rate and dwell time between personnel trajectories and thermal domains. Threshold ranges are dynamically set according to the hazard level of the hot work and real-time environmental parameters. Tiered early warnings are triggered through feature parameter comparison, and equipment response and secondary verification mechanisms are linked. This solves the problems of low accuracy, weak anti-interference capability, lack of correlation with thermal domain risks, and poor timeliness, achieving accurate and real-time early warning of hot work leaving personnel risks.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solution:

[0007] The method for early warning of hot work leaving personnel based on behavioral trajectory and thermal domain correlation analysis includes the following steps:

[0008] Acquire multi-dimensional sensing data of the hot work area. The multi-dimensional sensing data should include at least personnel location trajectory data, thermal distribution data, and hot work equipment operation status data.

[0009] The spatiotemporal alignment algorithm is used to fuse personnel location trajectory data with thermal distribution data to obtain a spatiotemporal fusion dataset. The spatiotemporal alignment algorithm dynamically adjusts the alignment granularity by introducing a hot work operation time window, so that the timestamp matching error between trajectory data and thermal data is less than 50ms, which enables accurate analysis of spatiotemporal data misalignment.

[0010] Construct a behavior-heat domain association feature model and extract the association feature parameters between personnel trajectory and heat domain in the spatiotemporal fusion dataset. The association feature parameters include at least the distance change rate between personnel trajectory and the core area of ​​heat domain, the duration of personnel stay within the coverage area of ​​heat domain, and the dynamic response coefficient of heat domain morphology with personnel trajectory.

[0011] Set a threshold range for associated features in hot work scenarios, compare the extracted associated feature parameters with the threshold range, and trigger a personnel evacuation warning signal when the associated feature parameters exceed the threshold range for a preset duration.

[0012] Furthermore, the specific steps for acquiring multi-dimensional sensing data of the hot work area include:

[0013] The system employs a combined acquisition method of lidar and infrared thermal imager, where lidar acquires the three-dimensional position coordinates of personnel in real time to form trajectory data.

[0014] The infrared thermal imager acquires thermal images at a preset frame rate and converts them into thermal distribution data. At the same time, it collects operational status data such as start / stop status and flame temperature through the built-in sensors of the hot work equipment.

[0015] The collected multidimensional data is preprocessed, and the preprocessing steps include: noise removal of trajectory data based on Kalman filtering, background removal of thermal domain data based on adaptive threshold segmentation, and abnormal data completion based on data integrity verification, so as to achieve high-quality acquisition and preprocessing of multidimensional data.

[0016] Furthermore, the specific steps of the linkage acquisition method between lidar and infrared thermal imager include: realizing time synchronization between lidar and infrared thermal imager based on time synchronization protocol, and determining the coordinate transformation relationship between the two through spatial calibration algorithm;

[0017] The three-dimensional position coordinates collected by the lidar are mapped to the image coordinate system of the infrared thermal imager to achieve precise spatial matching between personnel position and thermal distribution.

[0018] When the lidar detects that a person's trajectory is obstructed, the location of the person is predicted based on historical trajectory data and the trend of thermal distribution changes, ensuring the continuity of the trajectory data under obstruction.

[0019] Furthermore, the specific steps for constructing the behavior-heat domain correlation feature model include:

[0020] A network for extracting related features is constructed based on deep learning networks. The input layer of the network receives a spatiotemporal fusion dataset, the hidden layer strengthens the association weight between personnel trajectory and the core area of ​​the hot zone through an attention mechanism, and the output layer outputs multi-dimensional related feature parameters.

[0021] By employing transfer learning, the parameters of the personnel-environment association model in general scenarios are transferred to the hot work operation scenario. Then, fine-tuning and optimization are performed using a hot work operation sample dataset. The sample dataset contains labeled data under different hot work types, different personnel working postures, and different environmental interferences, thereby improving the model's feature extraction accuracy in the hot work scenario.

[0022] Furthermore, the training process of the association feature extraction network includes the following steps:

[0023] An adversarial training mechanism is introduced to construct a generator and a discriminator. The generator is used to generate simulated behavior-hot domain correlation features, and the discriminator is used to distinguish between real correlation features and generated simulated features.

[0024] The network parameters are optimized through adversarial training between the generator and the discriminator to improve the network's robustness to noise interference and changes in human posture.

[0025] Cross-validation is used to evaluate the performance of the trained network, with feature extraction accuracy and misclassification rate as evaluation metrics. When the evaluation metrics fail to reach the preset threshold, the model is retrained by expanding the sample dataset to ensure the reliability of the model.

[0026] Furthermore, the specific steps for setting the threshold range of associated features in hot work scenarios are as follows:

[0027] Different threshold ranges are divided based on the hazard level of hot work operations. The hazard level is determined according to the type of hot work operation, the density of combustibles in the work area, and the ventilation conditions of the work environment.

[0028] A dynamic threshold adjustment mechanism is adopted to collect environmental parameters such as wind speed, temperature and humidity in the working environment in real time. The threshold range of related features is dynamically corrected through a preset threshold correction formula, so that the threshold range can adapt to the hot work requirements under different environmental conditions and the fixed threshold can adapt to complex environmental changes.

[0029] Furthermore, after the human absence warning signal is triggered, the warning classification process is as follows: based on the degree and duration of the related characteristic parameters exceeding the threshold range, the warning signal is divided into three levels: Level 1 warning, Level 2 warning, and Level 3 warning.

[0030] Differentiated early warning response measures are implemented for different early warning levels. At the first level of early warning, local audible and visual alarms are activated and early warning information is pushed to the on-site monitor's terminal.

[0031] A Level II warning, based on a Level I warning, requires hot work equipment to operate at reduced power.

[0032] The Level 3 warning triggers an emergency shutdown of hot work equipment and activates fire emergency equipment in a pre-prepared state, building upon the Level 2 warning, thus achieving precise grading of the warning and efficient response.

[0033] Furthermore, the verification and tracing steps after the warning are as follows: after the warning signal is triggered, images of the hot work area are collected in real time through video monitoring equipment;

[0034] Image recognition algorithms are used to perform secondary verification of whether people have left the area, avoiding false alarms caused by sensor misdetection;

[0035] Establish an early warning and traceability database to record the time, warning level, associated characteristic parameters, on-site environmental parameters, response measures, and handling results of each early warning, forming a complete early warning and traceability chain to provide data support for subsequent operational safety optimization.

[0036] Furthermore, the steps for acquiring personnel location trajectory data include:

[0037] Personnel identity association involves obtaining the identity information of workers through UWB positioning tags or RFID identification cards, binding the identity information with the corresponding personnel location trajectory data, and forming a trajectory dataset with identity identification.

[0038] When a person is unaccounted for, the warning message includes the person's identity information and is simultaneously pushed to that person's mobile device, enabling precise identity association and targeted reminders for unaccounted individuals.

[0039] A hot work evacuation early warning system based on behavioral trajectory and thermal domain correlation analysis includes:

[0040] The multi-dimensional sensing module includes a lidar, an infrared thermal imager, equipment sensors, and an identity recognition device, used to collect multi-dimensional sensing data and personnel identity information.

[0041] The data fusion processing module is used to perform spatiotemporal alignment, data preprocessing, and identity-trajectory binding;

[0042] The association feature analysis module is used to construct association feature models and extract feature parameters;

[0043] The early warning determination and response module is used for threshold comparison, early warning classification, and execution of response measures;

[0044] The data storage module is used to store sensing data, model parameters, and early warning record data;

[0045] Each module is used to execute the steps of the above-mentioned hot work and personnel evacuation early warning method based on behavior trajectory and thermal domain correlation analysis. Data interaction is achieved through a high-speed communication bus to ensure the real-time performance and reliability of the system.

[0046] As can be seen from the above, this invention breaks through the limitations of traditional early warning systems that rely solely on location or image recognition by constructing a behavior-thermal domain correlation feature model. By analyzing the dynamic correlation parameters between personnel trajectories and the core area of ​​the thermal domain, it accurately identifies the safety risks caused by leaving personnel during hot work, making the early warning logic more aligned with the high-risk characteristics of hot work operations, and realizing intelligent early warning of associated thermal domain risks.

[0047] This invention utilizes a spatiotemporal alignment algorithm to control timestamp matching errors within 50ms, and combines it with a high-speed communication bus to achieve real-time data interaction, ensuring rapid triggering of early warning signals. It adopts a dynamic threshold adjustment mechanism to adaptively correct the threshold range based on the operational hazard level and real-time environmental parameters. At the same time, it establishes a three-level graded early warning response system, and implements differentiated measures such as audible and visual alarms, equipment power reduction, emergency shutdown, and fire linkage to achieve efficient risk management and has timely and responsive early warning capabilities. Attached Figure Description

[0048] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0049] 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, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1 This is a schematic diagram of the workflow of the present invention;

[0051] Figure 2 This is a schematic diagram of the internal structure of the system of the present invention;

[0052] Figure 3 This is a schematic diagram of the internal structure of the multidimensional sensing module of the present invention.

[0053] The following are the reference numerals: 10, Multi-dimensional perception module; 20, Data fusion processing module; 30, Correlation feature analysis module; 40, Early warning judgment and response module; 50, Data storage module; 100, LiDAR; 101, Infrared thermal imager; 102, Equipment sensor and identification device.

[0054] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0055] The technical solutions of the present invention will now be clearly and completely described with reference to the accompanying drawings.

[0056] Traditional methods of hot work absence warnings primarily rely on personnel location, video surveillance, or equipment triggering. In practical applications, these methods suffer from low accuracy, weak anti-interference capabilities, inability to effectively correlate with thermal risks, and poor timeliness. These limitations pose a serious challenge to safety assurance at hot work sites, making it difficult to effectively prevent major accidents such as fires and explosions caused by personnel leaving their posts.

[0057] To address the aforementioned issues, current hot work evasion warning technologies suffer from low accuracy, weak anti-interference capabilities, inability to correlate with thermal risks, and poor timeliness, making it difficult to effectively ensure on-site safety during hot work operations. This invention proposes a hot work evasion warning method and system based on behavioral trajectory and thermal domain correlation analysis. This method offers advantages such as accurate identification and real-time warning of personnel leaving their posts during hot work operations, effectively reducing false alarm rates and improving operational safety.

[0058] Therefore, as Figure 1As shown in the figure, this embodiment proposes a method for early warning of hot work leaving personnel based on behavioral trajectory and thermal domain correlation analysis, including the following steps:

[0059] Step S001: Obtain multi-dimensional sensing data of the hot work area. The multi-dimensional sensing data includes at least personnel location trajectory data, thermal distribution data, and hot work equipment operating status data.

[0060] Step S002: Based on the spatiotemporal alignment algorithm, the personnel location trajectory data and the thermal distribution data are fused to obtain the spatiotemporal fusion dataset;

[0061] Step S003: Construct a behavior-heat domain association feature model and extract the association feature parameters between personnel trajectory and heat domain in the spatiotemporal fusion dataset. The association feature parameters include at least the distance change rate between personnel trajectory and the core area of ​​heat domain, the duration of personnel stay within the coverage area of ​​heat domain, and the dynamic response coefficient of heat domain morphology with personnel trajectory.

[0062] Step S004: Set the threshold range of associated features in the hot work scenario, compare the extracted associated feature parameters with the threshold range, and trigger the personnel departure warning signal when the associated feature parameters exceed the threshold range and continue for a preset time.

[0063] For ease of understanding, the following explains some key terms in this embodiment:

[0064] Multidimensional sensing data refers to various types of data collected from the hot work area, with the aim of comprehensively reflecting personnel activities, heat source distribution, and equipment operation status at the work site.

[0065] Personnel location trajectory data describes the real-time location information of workers in the hot work area and their path changing over time, usually presented in the form of a series of coordinate points and timestamps.

[0066] Thermal distribution data reflects the spatial distribution of the temperature field within the hot work area. It is usually acquired through thermal imaging technology and used to identify high-temperature areas and their extent.

[0067] The operating status data of hot work equipment includes key operating parameters such as the start-up and shutdown status of the hot work equipment (e.g., welding machine, cutting machine) and flame temperature, which are used to determine the actual progress of hot work.

[0068] Spatiotemporal alignment algorithms are used to address the problem of inconsistent timestamps and spatial coordinates between different sensors or data sources. Their purpose is to accurately match data from different sources in both time and space dimensions for subsequent fusion analysis.

[0069] A spatiotemporal fusion dataset refers to a unified dataset that integrates personnel location trajectory data and thermal distribution data after being processed by a spatiotemporal alignment algorithm. This dataset contains synchronous information on personnel behavior and thermal changes.

[0070] The behavior-heat domain correlation feature model is a mathematical or computational model used to analyze the interaction between personnel behavior and heat domains. Its purpose is to extract key features from spatiotemporal fusion data that can characterize this correlation. Correlation feature parameters are numerical indicators extracted from the behavior-heat domain correlation feature model to quantify the degree of correlation between personnel behavior and heat domains; these parameters directly reflect the risk of personnel leaving their posts.

[0071] The core area of ​​the hot zone refers to the region with the highest temperature and most concentrated energy in the hot zone distribution. It is usually the direct point of action of hot work or the area near it, and has the highest risk.

[0072] The rate of change of distance describes how quickly the distance between a person's location and the core of the hot zone changes over time. This parameter can be used to determine whether a person is moving closer to or further away from the danger zone.

[0073] Duration of stay refers to the length of time a person remains within the coverage area of ​​a hot zone. This parameter can be used to assess the exposure risk of a person in a hazardous area.

[0074] The dynamic response coefficient reflects the degree of dynamic adjustment or influence of the thermal morphology (e.g., size and intensity) as the trajectory of personnel changes. This parameter can be used to assess the impact of personnel behavior on the diffusion or control of heat sources.

[0075] The threshold range for associated features refers to a safe range set for associated feature parameters. When the associated feature parameters exceed this range, it indicates a potential risk of human departure.

[0076] The preset duration refers to the length of time that the associated feature parameters continuously exceed the threshold range. This is used to avoid false alarms caused by instantaneous fluctuations and to ensure the stability of the warning.

[0077] The evacuation warning signal is an alarm message issued by the system when the warning conditions are met, and its purpose is to remind relevant personnel to take intervention measures.

[0078] To address this, the present invention further proposes specific steps for acquiring multidimensional sensing data of the hot work area, including: using a combined acquisition method of lidar and infrared thermal imager, where the lidar acquires the three-dimensional position coordinates of personnel in real time to form trajectory data; the infrared thermal imager acquires thermal images at a preset frame rate and converts them into thermal domain distribution data, while simultaneously acquiring operational status data such as start / stop status and flame temperature through the built-in sensors of the hot work equipment; and preprocessing the acquired multidimensional data, the preprocessing steps including: noise removal of trajectory data based on Kalman filtering, background removal of thermal domain data based on adaptive threshold segmentation, and abnormal data completion based on data integrity verification, to achieve high-quality acquisition and preprocessing of multidimensional data.

[0079] The use of a combined LiDAR and infrared thermal imager for data acquisition aims to combine the advantages of different sensor types to obtain more comprehensive and accurate multidimensional data. For example, hardware-level linkage can be achieved by installing LiDAR and infrared thermal imager on the same physical platform and coordinating their activation through a unified control unit; or software-level linkage can be achieved by deploying a unified data acquisition gateway at the data acquisition end to timestamp and integrate data streams from different sensors.

[0080] Infrared thermal imagers detect the infrared radiation energy emitted from the surface of objects and convert it into visualized thermal images, thus reflecting the temperature distribution of the work area. The device continuously acquires thermal images at a preset fixed frame rate (e.g., 10 or 30 frames per second), ensuring the continuity and timeliness of the thermal data. Subsequently, image processing techniques, such as temperature threshold-based binarization or region growing algorithms, are used to identify and extract areas in the thermal images with temperatures exceeding a specific threshold as thermal distribution data, thereby clearly defining the range and shape of high-temperature areas.

[0081] Sensors integrated into hot work equipment (such as welding and cutting machines) can monitor the equipment's operating status in real time. For example, current sensors can detect the equipment's operating current to determine if it is in a startup state; or pressure sensors, flow sensors, and other sensors can monitor key operating parameters to indirectly determine its start-up or shutdown status. Simultaneously, flame temperature sensors (such as thermocouples or infrared temperature measurement modules) can directly measure the real-time temperature of the flame, providing crucial data for assessing the hazard level of hot work operations. Combining this operational status data with personnel behavior data allows for a more comprehensive assessment of the risks associated with hot work operations.

[0082] Data preprocessing is a crucial step in improving the quality of raw sensory data. It aims to eliminate noise, redundancy, inconsistencies, and missing values, ensuring the data meets the input requirements of subsequent analytical models. For example, it may include common processing steps such as data format standardization, timestamp alignment, and outlier detection and labeling, providing a clean and reliable data foundation for subsequent feature extraction and model building.

[0083] Kalman filtering is a highly efficient recursive filter, particularly suitable for estimating the state of a dynamic system from a series of incomplete or noisy measurements. In personnel trajectory data, the acquired position data often contains random noise due to sensor accuracy limitations or environmental interference. By constructing a personnel motion model (e.g., a uniform motion model or a uniformly accelerated motion model), Kalman filtering can optimally estimate the true position of the personnel based on the previous time-to-time estimate and the current time-to-time measurement, thereby effectively filtering out random noise in the trajectory data and improving the smoothness and accuracy of the trajectory. Besides Kalman filtering, other signal processing methods such as moving average filtering and Gaussian filtering can also be used for noise removal.

[0084] In thermal domain data, besides the core heat source generated by hot work, there may also be environmental background heat sources (such as direct sunlight and heat dissipation from other equipment). These background heat sources can interfere with the identification of the true thermal domain. Adaptive threshold segmentation technology can dynamically determine the segmentation threshold based on the grayscale or temperature characteristics of local regions in the thermal image, rather than using a single fixed threshold. For example, local Otsu's algorithm or adaptive segmentation methods based on region growing can be used to more accurately separate the boundaries of the thermal domain from the complex background, thereby effectively removing irrelevant background information and focusing on the core thermal domain generated by the hot work. In addition, methods based on morphological operations (such as opening and closing operations) can also be used to remove small noise regions or thermal domains with broken connections.

[0085] Through the above technical solution, this invention effectively solves the problems of noise interference, background influence, and data incompleteness that may exist in the acquisition process of multi-dimensional sensing data, significantly improving data quality and reliability. Specifically, by adopting a linkage acquisition method of lidar and infrared thermal imager, combined with the built-in sensors of hot work equipment, comprehensive, real-time, and synchronous acquisition of personnel location, thermal domain distribution, and equipment operating status data is achieved, providing rich and complementary data sources for subsequent behavior-thermal domain correlation analysis. On this basis, noise removal of trajectory data based on Kalman filtering ensures the smoothness and accuracy of personnel trajectories; background removal of thermal domain data based on adaptive threshold segmentation makes the identification of core thermal domains more accurate and effectively eliminates environmental interference; and abnormal data completion based on data integrity verification ensures the continuity and integrity of the data stream. These refined data acquisition and preprocessing steps together lay a solid data foundation for constructing a behavior-thermal domain correlation feature model, greatly improving the accuracy and robustness of correlation feature parameter extraction, thereby significantly improving the accuracy and reliability of hot work personnel departure warning, effectively reducing the risk of false alarms and missed alarms, and providing more reliable protection for hot work safety.

[0086] In some of the solutions described above in this invention, a method of linking lidar and infrared thermal imager for acquiring multi-dimensional sensing data of the hot work area was proposed. However, when the lidar detects obstruction in the personnel trajectory, the personnel trajectory data is interrupted, affecting data continuity. To address this, this invention further proposes specific steps for the linked acquisition method of lidar and infrared thermal imager to solve the above problem and improve data continuity.

[0087] Specifically, this coordinated data acquisition method includes: achieving time synchronization between the lidar and the infrared thermal imager based on a time synchronization protocol, and determining the coordinate transformation relationship between the two through a spatial calibration algorithm. The time synchronization protocol aims to ensure that different sensors have a unified time reference during data acquisition, thereby guaranteeing the temporal consistency of the acquired data. For example, standard protocols such as Network Time Protocol (NTP) or Precision Time Protocol (PTP) can be used, with a synchronization server sending timestamps to each sensor to calibrate its internal clock; or a hardware triggering mechanism can be used, where a main control unit simultaneously sends synchronization trigger signals to the lidar and the infrared thermal imager, causing them to start or complete data frame acquisition at the same time. Spatial calibration algorithms are used to establish the spatial correspondence between the three-dimensional coordinate system of the lidar and the two-dimensional image coordinate system of the infrared thermal imager, enabling data from different sensors to be fused and analyzed under a unified spatial reference system. For example, a feature-point-based calibration method can be used, placing a calibration board with a known geometric shape in the scene, and calculating the rotation and translation matrix between them by identifying the projections of feature points on the calibration board into the two sensors; or a motion-based calibration method can be used, moving an object with a known geometric shape and simultaneously recording its data in the two sensors, and using optimization algorithms to solve for the external parameters between the sensors.

[0088] When lidar detects obstruction in a person's trajectory, the system predicts the person's location based on historical trajectory data and thermal distribution trends, ensuring the continuity of trajectory data even under obstruction. Person location prediction aims to infer the person's current possible location by utilizing previously collected trajectory information and dynamic changes in thermal distribution when lidar cannot acquire their real-time location due to obstruction, thus maintaining the continuity of trajectory data. For example, a Kalman filter or extended Kalman filter can be used, combined with the person's historical motion state (e.g., velocity, acceleration) and thermal diffusion / contraction trends, to predict the person's future location; alternatively, machine learning-based methods, such as recurrent neural networks (RNNs) or long short-term memory networks (LSTMs), can be employed to predict the person's movement path during obstruction by learning from a large amount of historical trajectory data and thermal change patterns. Through these prediction mechanisms, even when sensor data is temporarily interrupted or incomplete, the integrity and continuity of personnel location trajectory data can be maintained, avoiding system failure or misjudgment due to data loss.

[0089] Through the above technical solution, this invention effectively solves the problem of interrupted personnel trajectory data due to lidar obstruction in hot work scenarios, affecting data continuity. Specifically, based on a time synchronization protocol, it ensures the temporal consistency of data acquisition between lidar and infrared thermal imager, avoiding data fusion errors caused by time misalignment and laying the foundation for subsequent accurate matching. Through a spatial calibration algorithm, a precise transformation relationship is established between the lidar's three-dimensional coordinate system and the infrared thermal imager's image coordinate system, making it possible to map the three-dimensional position coordinates of personnel acquired by lidar to the infrared thermal imager's image coordinate system. This achieves precise spatial matching of personnel position and thermal domain distribution, greatly improving the accuracy of personnel behavior and thermal domain correlation analysis. More importantly, when lidar detects obstruction in personnel trajectory, this solution can intelligently predict based on historical trajectory data and the changing trends of thermal domain distribution, effectively compensating for data loss caused by obstruction and ensuring the continuity and integrity of personnel trajectory data. This enables the early warning system to continuously and stably acquire personnel location information even in complex and ever-changing hot work environments, avoiding early warning failures or misjudgments caused by trajectory interruptions. This significantly improves the accuracy and reliability of hot work personnel departure warnings, further ensuring the safety of hot work operations.

[0090] In some of the embodiments of the present invention described above, a behavior-thermal domain correlation feature model is proposed to extract the correlation feature parameters between personnel trajectory and thermal domain. However, in its implementation, the model has insufficient accuracy in feature extraction under hot work scenarios and needs to be optimized for different hot work types, personnel working postures and environmental interference.

[0091] To address this, the present invention further proposes specific steps for constructing a behavior-thermal domain association feature model, including: constructing an association feature extraction network based on a deep learning network, wherein the input layer of the network receives a spatiotemporal fusion dataset, the hidden layer strengthens the association weight between personnel trajectory and the core area of ​​the thermal domain through an attention mechanism, and the output layer outputs multi-dimensional association feature parameters; employing transfer learning, the parameters of the personnel-environment association model in general scenarios are transferred to the hot work operation scenario, and then fine-tuned and optimized using a hot work operation sample dataset, wherein the sample dataset contains labeled data under different hot work types, different personnel working postures, and different environmental interferences, thereby improving the accuracy of feature extraction of the model in the hot work scenario.

[0092] Specifically, deep learning networks are multi-layered neural networks capable of automatically extracting complex features by learning from large amounts of data. Their advantage lies in their ability to handle high-dimensional, non-linear data and automatically discover latent patterns within the data. When constructing a network for feature extraction, various deep learning architectures can be employed. For example, convolutional neural networks (CNNs) can be used to process local and time-series features in spatiotemporally fused datasets, or recurrent neural networks (RNNs) or Transformers can be used to capture longer-term temporal dependencies between personnel trajectories and heat domains. The input layer of the network is the starting part of the deep learning network, responsible for receiving raw or preprocessed data. Here, the input layer receives a spatiotemporally fused dataset, which has spatiotemporally aligned and fused personnel location trajectory data with heat domain distribution data, ensuring the integrity and relevance of the input data. This allows the network to directly learn the complex relationship between personnel behavior and heat domains from the fused data, avoiding information loss or mismatch issues. Hidden layers are the core part of deep learning networks for feature learning and transformation. Attention mechanisms are techniques that allow the network to dynamically focus on important parts of the input data. By introducing an attention mechanism into the hidden layers, the network can adaptively allocate weights to different regions or time steps based on the importance of the current task (i.e., related feature extraction). For example, a self-attention mechanism can be used to capture the correlation between different time points within a person's trajectory, or a cross-attention mechanism can be used to explicitly model the interaction between the person's trajectory and the core area of ​​the hotspot, thereby more accurately identifying key interaction patterns and effectively suppressing background noise and irrelevant information. The output layer is the final part of the deep learning network, responsible for converting the features learned by the hidden layers into the final output. Here, the output layer outputs multi-dimensional related feature parameters, which are quantitative indicators learned and refined by the network that can characterize the complex relationship between the person's trajectory and the hotspot. These parameters may include, but are not limited to, the rate of change of distance between the person's trajectory and the core area of ​​the hotspot, the duration of the person's stay within the coverage area of ​​the hotspot, and the dynamic response coefficient of the hotspot morphology with the person's trajectory, providing a comprehensive and detailed basis for subsequent departure warning determination.

[0093] To further optimize the feature extraction accuracy of the model in hot work scenarios, this invention employs transfer learning to transfer the parameters of a personnel-environment association model from a general scenario to a hot work operation scenario. Transfer learning is a machine learning method that aims to apply a model (or some of its parameters) already trained on a source task to a target task. Here, through transfer learning, the general feature representation capabilities learned by the model pre-trained in a general personnel-environment association scenario can be utilized, and its parameters can be used as the initialization parameters of the association feature extraction network in a hot work operation scenario. This approach can effectively utilize existing knowledge, accelerate the convergence speed of the model in the specific and potentially data-scarce hot work operation scenario, and improve the model's generalization ability. Based on this, fine-tuning and optimization are then performed using a hot work operation sample dataset. Fine-tuning and optimization refer to the process in transfer learning of further training and adjusting the parameters of the pre-trained model on a specific dataset for the target task. Fine-tuning the transferred model using a hot work operation sample dataset allows the model to better adapt to the unique characteristics and complexity of hot work operation scenarios. During fine-tuning, a relatively small learning rate is typically used to preserve the general knowledge learned by the pre-trained model while enabling it to learn fine-grained features specific to hot work scenarios. The quality and diversity of the sample dataset are crucial to the model's training effectiveness. This dataset is constructed to fully consider the complexity and diversity of hot work operations, including labeled data under different hot work types, different personnel postures, and different environmental disturbances. This diverse data ensures that the model learns more robust and generalizable features, thus enabling it to accurately extract relevant features in real-world hot work environments, even in complex and changing situations, avoiding misjudgments or omissions.

[0094] Through the above technical solution, this invention effectively solves the problem of insufficient accuracy in feature extraction from behavior-heat domain association feature models in hot work scenarios. Specifically, by constructing an association feature extraction network based on a deep learning network, the powerful nonlinear modeling capabilities of deep learning enable it to automatically learn and extract complex association features between personnel trajectories and heat domains from spatiotemporal fusion datasets, avoiding the limitations of manually designing features in traditional methods. The attention mechanism introduced in the hidden layer allows the network to dynamically focus on key interaction areas between personnel trajectories and the core area of ​​the heat domain, effectively strengthening the association weights and thus suppressing noise interference in complex environments, improving the accuracy of feature extraction. Furthermore, by employing transfer learning, the parameters of the personnel-environment association model pre-trained in general scenarios are transferred to hot work scenarios, and fine-tuned and optimized using a hot work sample dataset containing different hot work types, personnel working postures, and environmental interference, enabling the model to fully utilize general knowledge and quickly adapt to the diversity of hot work operations. This significantly improves the accuracy of feature extraction in hot work scenarios, ensures the reliability of the extracted associated feature parameters, and provides a more accurate and robust basis for subsequent hot work and personnel evacuation warnings, thereby effectively reducing the risk of false alarms and missed alarms.

[0095] In some of the embodiments of the present invention described above, a training process for an associated feature extraction network is proposed to improve the accuracy of feature extraction. However, during its implementation, the network is easily affected by noise interference and changes in human posture, resulting in insufficient model robustness and affecting the reliability of early warning.

[0096] To address this, the present invention further proposes a training process for the associated feature extraction network, including the following steps:

[0097] An adversarial training mechanism is introduced to construct a generator and a discriminator. The generator is used to generate simulated behavior-hot domain correlation features, and the discriminator is used to distinguish between real correlation features and generated simulated features.

[0098] The network parameters are optimized through adversarial training between the generator and the discriminator to improve the network's robustness to noise interference and changes in human posture.

[0099] Cross-validation is used to evaluate the performance of the trained network, with feature extraction accuracy and misclassification rate as evaluation metrics. When the evaluation metrics fail to reach the preset threshold, the model is retrained by expanding the sample dataset to ensure the reliability of the model.

[0100] Specifically, when introducing adversarial training mechanisms, a Generative Adversarial Network (GAN) architecture can be constructed. The generator can be a deep neural network, such as a structure based on a Convolutional Neural Network (CNN) or a Recurrent Neural Network (RNN). Its input can be a random noise vector or real behavior-heat domain correlation features. It learns the data distribution to generate simulated features similar to real features. The discriminator can also be a deep neural network, such as a Multilayer Perceptron or a CNN. It receives real behavior-heat domain correlation features or simulated features generated by the generator as input and outputs a probability value indicating whether the input feature is real or simulated. In this way, the generator continuously attempts to generate more realistic simulated features to "deceive" the discriminator, while the discriminator continuously improves its ability to distinguish between real and simulated features. Another implementation is that the generator can be designed to apply specific perturbations to real behavior-heat domain correlation features (e.g., adding simulated noise, performing pose changes, etc.), thereby generating simulated features with noise or pose changes. The discriminator then learns to recognize these perturbated features, thus forcing the feature extraction network to learn feature representations robust to these perturbations.

[0101] Through the above technical solution, this invention effectively solves the problem of insufficient robustness of the correlation feature extraction network under noise interference and personnel posture changes. The introduction of an adversarial training mechanism enables the network to actively learn and adapt to various potential interference factors, thereby extracting more stable and discriminative behavior-heat domain correlation features. Through adversarial optimization between the generator and discriminator, the network parameters are finely adjusted, significantly improving its adaptability to dynamic environmental changes and reducing feature extraction bias caused by environmental noise or the diversity of personnel movements. Furthermore, the performance evaluation mechanism using cross-validation combined with iterative feedback ensures the model's generalization ability on different data subsets, and through continuous sample dataset expansion and retraining, the model performance is continuously optimized, effectively reducing the false alarm rate and false negative rate of the warning. This enables the hot work away from personnel warning system to provide more accurate and reliable warning signals in actual operation scenarios, thereby significantly improving the safety of hot work operations.

[0102] In some of the embodiments of the present invention described above, a threshold range of associated features in hot work scenarios is proposed to determine whether personnel have left their posts and trigger an early warning. However, in its implementation, the fixed threshold cannot adapt to changes in different environmental conditions, resulting in a decrease in the accuracy of the early warning.

[0103] In response, this invention further proposes specific steps for setting the associated feature threshold range under hot work scenarios: different threshold ranges are divided based on the hazard level of the hot work, which is determined according to the type of hot work, the density of combustibles in the work area, and the ventilation conditions of the work environment; a dynamic threshold adjustment mechanism is adopted to collect environmental parameters such as wind speed, temperature, and humidity in the work environment in real time, and the associated feature threshold range is dynamically corrected through a preset threshold correction formula, so that the threshold range adapts to the hot work requirements under different environmental conditions, and the fixed threshold can adapt to complex environmental changes.

[0104] The system categorizes hot work operations into different threshold ranges based on their inherent risk levels, aiming to provide an initial, graded judgment standard for the early warning system. Hazard levels are a quantitative assessment of the inherent risks of hot work operations, with the goal of introducing risk-based management into the early warning mechanism. Hazard levels can be determined in various ways. For example, based on industry safety standards, internal company operating procedures, and the experience of safety experts, a comprehensive assessment can be conducted on different types of hot work (e.g., welding, cutting, grinding, and baking), the density of combustibles in the work area (e.g., the presence of flammable liquids, gases, dust, and wood), and the ventilation conditions of the work environment (e.g., open spaces, semi-enclosed spaces, and enclosed spaces). Multiple hazard levels, such as high, medium, and low, can be pre-defined, and an initial set of associated characteristic threshold ranges can be configured for each level. Alternatively, a risk matrix can be established, with the horizontal axis representing the combination of hot work type and combustible density, and the vertical axis representing ventilation conditions. By quantifying the risk of each cell in the matrix, the corresponding hazard level can be determined, and different threshold ranges can be defined accordingly. For example, high-risk levels correspond to narrower and stricter threshold ranges, meaning that a warning is triggered if personnel deviate slightly or the thermal field becomes slightly abnormal; medium-risk levels correspond to moderate threshold ranges; and low-risk levels correspond to broader threshold ranges.

[0105] Furthermore, a dynamic threshold adjustment mechanism is employed to enable the early warning system's judgment criteria to adapt to subtle changes in the work environment in real time, thereby improving the accuracy and robustness of the early warning. This mechanism dynamically adjusts the threshold range of related features by collecting environmental parameters such as wind speed, temperature, and humidity in real time and using a preset threshold correction formula. Real-time collected environmental parameters include: wind speed, obtained through a wind speed sensor, which affects heat diffusion and spark distance; temperature, obtained through a temperature sensor, which affects the ignition point of combustibles and the diffusion rate of the heat zone; and humidity, obtained through a humidity sensor, which affects the moisture content in the air and may affect the spark extinguishing rate. The threshold correction formula can be a predefined mathematical model. For example, when wind speed increases, the spark distance may increase, in which case the lower limit of the threshold for the rate of change of distance between personnel and the core area of ​​the heat zone can be appropriately tightened; when the ambient temperature rises, the ignition point of combustibles may decrease, in which case the upper limit of the threshold for the duration of personnel stay within the heat zone coverage area can be appropriately tightened. The correction formula can be expressed as: Correction threshold = Base threshold + f (wind speed) + g (temperature) + h (humidity), where f, g, and h are correction functions determined based on experimental data or experience. Alternatively, a model can be trained using machine learning algorithms (such as regression models or neural networks) by collecting a large amount of hot work operation data under different environmental parameters, including normal operations, critical states, and evacuation warning events, and labeling the corresponding environmental parameters and related feature parameters. This model can predict the current optimal range of related feature thresholds based on real-time environmental parameters such as wind speed, temperature, and humidity.

[0106] Through the above technical solution, this invention enables refined management of the risks associated with hot work unattended by personnel. It employs a dynamic threshold adjustment mechanism to further consider the real-time changes in environmental parameters (wind speed, temperature, humidity) during operations. These parameters directly affect key safety factors such as heat diffusion, spark splashing, and the ignition point of combustibles. By correcting thresholds in real time, the system can dynamically adapt to these changes. Combining hazard level classification and dynamic threshold adjustment, this invention optimizes the early warning threshold from two dimensions: static operational risk assessment and dynamic environmental changes. This significantly improves the accuracy, timeliness, and robustness of early warnings, effectively reducing the safety risks of hot work operations and ensuring the safety of personnel and property.

[0107] In some of the embodiments of the present invention described above, a triggering warning signal for leaving personnel is proposed to detect personnel leaving their posts. However, in this process, the warning signal lacks a hierarchical mechanism, resulting in a single response measure. It is impossible to respond accurately and efficiently to personnel leaving events of different severity levels, which may lead to resource waste or insufficient response and affect the overall safety of hot work operations.

[0108] To address this, the present invention further proposes a pre-warning grading process after the triggering of a vacancy warning signal. This process categorizes the warning signal into three levels—Level 1, Level 2, and Level 3—based on the degree and duration of the related characteristic parameters exceeding the threshold range. The pre-warning grading process involves classifying the warning event into different risk levels according to its severity and urgency, and setting corresponding processing procedures and response measures for each level. This allows the system to respond more flexibly and economically to vacancy events of varying risk levels. The degree to which related characteristic parameters exceed the threshold range refers to the deviation of parameters such as the rate of change of distance between the person's trajectory and the core area of ​​the hotspot, the duration of the person's stay within the hotspot's coverage area, and the dynamic response coefficient of the hotspot's morphology following the person's trajectory from the normal threshold range. The greater the deviation, the higher the risk of vacancy. The duration refers to the duration of the state where the related characteristic parameters exceed the threshold range. The longer the duration, the more certain the vacancy behavior, and the higher the accumulated risk. These two indicators provide an objective and quantitative basis for pre-warning grading, ensuring the scientific accuracy of the pre-warning level determination. Level 1, Level 2, and Level 3 warnings are three levels of risk classification for leaving people unattended while hot work, from low to high. Level 1 warnings usually correspond to minor risks of leaving people unattended, Level 2 warnings correspond to moderate risks of leaving people unattended, and Level 3 warnings correspond to severe risks of leaving people unattended.

[0109] Differentiated early warning response measures are implemented for different warning levels. These differentiated measures refer to the system automatically executing or suggesting different types and intensities of intervention actions based on the warning level, ensuring that the response measures match the risk level and avoiding over- or under-response. Specifically, a Level 1 warning activates a local audible and visual alarm and pushes the warning information to the on-site supervisor's terminal. The local audible and visual alarm refers to equipment installed near the hot work area that alerts on-site personnel by emitting sound and light, for example, by sending a trigger signal to a connected audible and visual alarm via a control module. Pushing the warning information to the on-site supervisor's terminal means sending detailed information about the warning event to the mobile device held by the on-site safety supervisor via a wireless communication network.

[0110] Through the above technical solution, this invention can finely classify vacant warning signals based on the degree and duration of the correlation feature parameters exceeding the threshold range, thereby solving the problem of a lack of a classification mechanism for warning signals, leading to a single response measure. This classification mechanism makes the warning system no longer a simple "yes / no" judgment, but can provide multi-level, progressive responses based on subtle changes in vacant behavior and the accumulation of potential risks. For example, for minor vacant risks, only local audible and visual alarms are activated and the supervisor is notified, avoiding unnecessary production interruptions; for medium risks, hot work equipment is further linked to reduce power operation, maintaining some operational continuity while reducing risk; and for severe vacant risks, emergency equipment shutdown is immediately triggered and fire emergency equipment is activated to a pre-ready state, eliminating ignition sources to the greatest extent possible and preparing for the fastest possible response to possible fire accidents. This precise classification avoids overreacting to minor vacant events, reduces unnecessary production interruptions and resource waste, and at the same time, can quickly escalate the response to severe vacant events, maximizing operational safety and effectively reducing the probability of major accidents such as fires and explosions. By using the correlation feature parameters obtained from the fusion and analysis of multi-dimensional sensing data such as personnel behavior trajectories, thermal distribution, and equipment operating status as the basis for classification, the scientific nature and accuracy of classification judgment are ensured, and the intelligence and reliability of the entire early warning system are improved.

[0111] In the above-mentioned solution of the present invention, a graded warning processing method is proposed to achieve accurate grading and efficient response. However, in this process, false warnings may be caused by sensor misdetection, and there is a lack of verification and traceability mechanisms for warning events, which makes it impossible to ensure the accuracy of the warning and provide data support to optimize the safety of subsequent operations.

[0112] To address this, the present invention further proposes a verification and tracing step after the early warning is issued. Specifically, after the early warning signal is triggered, images of the hot work area are collected in real time using video surveillance equipment. The video surveillance equipment can be a high-definition network camera deployed in the hot work area, transmitting the video stream in real time to the central processing unit for analysis via Ethernet or a wireless network; alternatively, the video surveillance equipment can be a smart camera with edge computing capabilities, performing preliminary processing of the image data locally and uploading key frames or event video clips to the system.

[0113] Subsequently, an image recognition algorithm is used for secondary verification to determine whether personnel have left the area, in order to avoid false alarms caused by sensor misdetection. The image recognition algorithm can be based on a deep learning convolutional neural network (CNN) model, such as YOLO or Faster R-CNN, trained to recognize personnel engaged in hot work and determine their presence within a pre-defined work area; alternatively, the image recognition algorithm can employ a skeleton keypoint detection (such as OpenPose) combined with behavioral analysis, analyzing the personnel's posture and movements to determine whether they are in a working state or have left the work area.

[0114] Simultaneously, an early warning traceability database is established to record the time, warning level, associated characteristic parameters, on-site environmental parameters, response measures, and processing results of each early warning, thus forming a complete early warning traceability chain and providing data support for subsequent operational safety optimization. The early warning traceability database can use a relational database (such as MySQL or PostgreSQL), with a table structure containing fields such as early warning time, early warning level, associated characteristic parameters, on-site environmental parameters, response measures, and processing results to store data in a structured manner; alternatively, the early warning traceability database can use a non-relational database (such as MongoDB), storing the structured and unstructured data of each early warning in document form for flexible expansion and management. During the recording process, when an early warning signal is triggered, the system can automatically retrieve relevant data from various related modules (such as the early warning judgment and response module and the multi-dimensional sensing module) and package and write it into the database; alternatively, through a message queue mechanism, early warning-related data generated by different modules can be asynchronously sent to the data storage module for unified recording. By using foreign keys or unique identifiers in the database to logically associate different types of data (such as early warning event ID, personnel ID, and equipment ID), a complete early warning traceability chain is formed. To provide data support for subsequent operational safety optimization, statistical analysis of data in the traceability database can be performed regularly to identify high-frequency warning scenarios, false alarm causes, response efficiency, etc., providing a basis for adjusting system parameters and improving operational processes; alternatively, machine learning models can be used to train historical warning data to predict potential risks and optimize warning thresholds or algorithm models.

[0115] Through the above technical solution, this invention effectively solves the problem of false alarms that may occur during the hot work evacuation warning process and establishes a comprehensive event traceability mechanism. Specifically, after the warning signal is triggered, the system can collect images of the hot work area in real time through independent video monitoring equipment, and use image recognition algorithms to perform secondary verification of whether personnel have left the area. This effectively avoids false alarms caused by misdetection from a single sensor, significantly improving the accuracy and reliability of the warning. Simultaneously, by establishing a warning traceability database, the system can comprehensively record the time, warning level, associated characteristic parameters, on-site environmental parameters, response measures, and processing results of each warning, forming a complete warning traceability chain. This not only provides detailed data for subsequent accident analysis and liability determination, but more importantly, through the analysis of this historical data, it is possible to continuously optimize the parameters of the warning model, adjust the warning threshold, and even improve the safety procedures for hot work operations. This achieves continuous improvement and enhancement of operational safety, providing strong data support and decision-making basis for the safety management of hot work operations.

[0116] In some of the solutions described above in this invention, the acquisition of personnel location trajectory data is proposed for early warning of personnel leaving the area. However, in this process, it is impossible to accurately associate the identity information of the personnel leaving the area, resulting in the early warning information not being able to accurately target and remind relevant personnel, thereby reducing the pertinence and response efficiency of the early warning.

[0117] In response, this invention further proposes the following steps for acquiring the aforementioned personnel location trajectory data: personnel identity association, which involves acquiring the identity information of the workers through UWB positioning tags or RFID identification cards, binding the identity information with the corresponding personnel location trajectory data to form a trajectory dataset with identity identification; when a personnel departure warning is triggered, the warning information carries the identity information of the departing personnel and is simultaneously pushed to the personnel's mobile terminal, thereby achieving accurate identity association and targeted reminders for personnel departure warnings.

[0118] Specifically, personnel identification linkage aims to ensure that each monitored personnel location trajectory data can be uniquely linked to a specific worker, which is the foundation for accurate early warning and targeted alerts. This linkage can be achieved in several ways. For example, by pre-entering the worker's biometric information (such as fingerprints and facial features) into the system database, and verifying the worker's identity through a biometric device when entering the work area, the verified identity information is bound to the real-time collected personnel location trajectory data; or, by using QR code or barcode scanning, workers wear QR codes or barcodes containing their identity information, which are scanned before or during work using fixed or mobile scanning devices, and the scanned identity information is linked to the current personnel's location trajectory data.

[0119] Using UWB positioning tags or RFID identification cards to obtain worker identification information provides a wireless, contactless method for automatically identifying and acquiring worker identity information, avoiding the tediousness and errors of manual data entry, and enabling integration with positioning systems. For example, workers can wear UWB positioning tags, which not only provide high-precision positioning data, but their built-in unique IDs also serve as identification. The system receives the tag signal via a UWB base station, parses the ID, and matches it with pre-stored identity information. Alternatively, workers can wear RFID identification cards, which embed RFID chips to store identity information. RFID readers are deployed in the work area; when a worker enters the reader's sensing range, the reader automatically reads the card information to obtain the worker's identity.

[0120] Binding identity information to corresponding personnel location trajectory data aims to establish a logical correspondence between personnel identity and their spatial location and time sequence, ensuring that each segment of trajectory data clearly belongs to a specific person. For example, at the data acquisition end, when a UWB positioning tag or RFID tag identifies a person, this identity information is used as metadata and packaged with the personnel location coordinate data collected by LiDAR (or the UWB positioning system itself) within the same time period to form a composite data packet containing identity and location. Alternatively, at the data processing center, binding can be achieved through timestamp alignment. For instance, if the UWB or RFID system acquires personnel A's identity information at time T1, and the LiDAR acquires personnel A's location coordinates at time T1, the system matches and associates these two independent data streams at time T1 to generate trajectory points with identity identifiers.

[0121] The goal of creating an identified trajectory dataset is to build a structured dataset where each trajectory point or segment explicitly includes the identity information of the person to whom it belongs, providing a data foundation for subsequent precise early warning and tracing. For example, a relational database can be used for storage, with a table structure including fields such as person ID, timestamp, X coordinate, Y coordinate, and Z coordinate. Each time identified location data is collected, it is inserted as a record. Alternatively, a time-series database can be used, with person identification information as labels, location coordinates as fields, and timestamps as indexes, enabling efficient storage and retrieval of time-series trajectory data with identified information.

[0122] Including the identification information of departing personnel in early warning messages aims to ensure that the message not only informs that "someone has left their post," but also clearly identifies "who has left their post," thereby improving the targeting and efficiency of the warnings. For example, when the early warning system determines that a departure event has occurred, it extracts the unique identifier (e.g., employee ID, name) of the departing personnel from the trajectory dataset with identification marks and embeds it into the text content or data structure of the early warning message. Alternatively, the early warning message can be encapsulated in a structured data format such as JSON or XML, which includes a dedicated field for storing the identity information of the departing personnel.

[0123] Through the above technical solution, this invention, based on the acquisition of personnel location trajectory data, further introduces a personnel identity association mechanism. The identity information of the workers is obtained through UWB positioning tags or RFID identification cards and bound to the personnel location trajectory data, thereby forming a trajectory dataset with identity identifiers. When the system triggers a personnel departure warning, the warning message clearly identifies the departing personnel and is directly pushed to their mobile terminal. This allows the warning message to accurately target specific departing personnel, avoiding the problems of missing or ambiguous identity information in traditional warning methods, greatly improving the targeting and response efficiency of the warning. Departing personnel can receive the warning immediately and take corrective measures, effectively reducing the accident risk caused by "unattended personnel during hot work," thus ensuring on-site safety during hot work operations.

[0124] Example 2;

[0125] Based on Example 1, this example proposes a hot work departure warning system based on behavioral trajectory and thermal domain correlation analysis. Through multi-dimensional data perception and intelligent analysis mechanism, it can achieve accurate quantitative assessment of personnel departure risks.

[0126] like Figure 2 As shown, the system includes: a multi-dimensional perception module 10, a data fusion processing module 20, a correlation feature analysis module 30, an early warning judgment and response module 40, and a data storage module 50;

[0127] like Figure 3 As shown, the multi-dimensional sensing module 10 includes: a lidar 100, an infrared thermal imager 101, equipment sensors, and an identification device 102. The multi-dimensional sensing module 10 is used to collect real-time data on personnel location trajectories, thermal distribution data, operating status data of hot work equipment, and personnel identification information within the hot work area. Specifically, the lidar 100 provides a high-precision sequence of personnel location coordinates, the infrared thermal imager 101 captures spatial distribution images of the temperature field, the equipment sensors monitor the start / stop status of the equipment sensors and the identification device 102, and the flame temperature parameters. The identification device uses biometrics or electronic tags to bind personnel identities, thereby ensuring the integrity and synchronization of the multi-dimensional sensing data.

[0128] The data fusion processing module 20 receives the data stream output from the multi-dimensional perception module 10 and performs spatiotemporal alignment, data preprocessing, and identity-trajectory binding operations. The spatiotemporal alignment algorithm, based on timestamp calibration and coordinate system transformation, fuses personnel location trajectory data and thermal distribution data under a unified spatiotemporal reference. Data preprocessing uses sliding window filtering to eliminate sensor noise and interpolation algorithms to fill in any missing or abnormal data. Identity-trajectory binding establishes a unique mapping between personnel identity information and corresponding trajectory data, generating a structured spatiotemporal fusion dataset. Based on this, the data output from the data fusion processing module provides high-quality input for subsequent feature analysis.

[0129] The correlation feature analysis module 30 constructs a correlation feature model based on a spatiotemporal fusion dataset, extracting feature parameters that characterize the dynamic relationship between personnel behavior and the thermal domain. This model analyzes the rate of change in distance between personnel trajectories and the core area of ​​the thermal domain through mathematical operations, quantifying the trend of personnel movement towards the danger zone; it calculates the duration of personnel stay within the thermal domain coverage area to assess the risk of continuous exposure; and it monitors the dynamic response coefficient of the thermal domain morphology to personnel trajectories, reflecting the degree of influence of personnel behavior on heat source diffusion. Due to the adoption of an adaptive weight allocation mechanism, the correlation feature analysis module can effectively suppress environmental interference and improve the robustness of the feature parameters.

[0130] The early warning judgment and response module 40 sets dynamic threshold ranges for hot work scenarios and compares the extracted feature parameters with the thresholds in real time. When the distance change rate exceeds the safety threshold, the dwell time exceeds the allowable range, or the dynamic response coefficient is abnormal, the system activates a graded response mechanism: for a level one warning, it triggers the on-site audible and visual alarm device; for a level two warning, it performs power reduction operation on the hot work equipment; for a level three warning, it automatically cuts off the equipment's energy supply. Because the threshold range is dynamically adjusted according to the working environment, and the preset duration mechanism avoids false alarms due to instantaneous fluctuations, the early warning judgment and response module significantly improves the timeliness and accuracy of early warnings.

[0131] The data storage module 50 is used to persistently store sensing data, model parameters, and early warning record data, supporting historical data tracing and system optimization iteration. The modules communicate with each other via a high-speed communication bus to achieve millisecond-level data exchange, ensuring that the end-to-end latency from data acquisition to early warning response is controlled within a safe threshold. Through the above technical solution, this application achieves deep correlation analysis between personnel behavior trajectories and thermal risk, effectively solving the core problems of low accuracy, weak anti-interference ability, inability to correlate thermal risk, and poor early warning timeliness in existing technologies, providing reliable protection for hot work safety.

[0132] The following example will provide a more detailed explanation of the above technical solution:

[0133] In an industrial production area, such as a pipe welding site at location A, there is a potential risk of fire or explosion caused by hot work, especially when workers leave the area; an unsupervised ignition source could lead to a serious accident. This solution addresses the above problem in the following ways:

[0134] First, the system deploys a multi-dimensional sensing module 10, and a lidar 100 to collect the three-dimensional position coordinates of the operator (e.g., User 1) in real time, forming personnel position trajectory data. Simultaneously, an infrared thermal imager 101 collects thermal images of the work area at a preset frame rate and converts them into thermal distribution data, clearly displaying the range and intensity of heat sources such as flames and high-temperature components. The built-in sensors of the equipment sensors and identification device 102 collect real-time data on the start / stop status of the welding equipment and the operating status of the flame temperature. Furthermore, User 1 wears a UWB positioning tag; their identity information is obtained through the identification device and bound to the personnel position trajectory data, forming a trajectory dataset with an identification tag.

[0135] The collected multidimensional data is transmitted to the data fusion processing module 20 for preprocessing. This module first synchronizes the time of the lidar 100 and the infrared thermal imager 101 based on a time synchronization protocol, and determines their coordinate transformation relationship through a spatial calibration algorithm. Subsequently, the three-dimensional position coordinates collected by the lidar 100 are precisely mapped to the image coordinate system of the infrared thermal imager 101, achieving accurate spatial matching between personnel position and thermal distribution. During data preprocessing, a Kalman filter algorithm is used to remove noise from the trajectory data, adaptive threshold segmentation technology is used to remove background interference from the thermal data, and abnormal data is supplemented based on data integrity verification, ensuring the high quality and continuity of the multidimensional data. For example, when the lidar detects a brief occlusion of user 1's trajectory, the system predicts the personnel position based on historical trajectory data and the trend of thermal distribution changes, ensuring the continuity of the trajectory data.

[0136] Next, the correlation feature analysis module 30 constructs a behavior-thermal domain correlation feature model. This model is based on a deep learning network, and its input layer receives a spatiotemporally fused dataset after spatiotemporal alignment and preprocessing. The hidden layer strengthens the correlation weight between personnel trajectory and the core area of ​​the thermal domain through an attention mechanism, thereby capturing key features more accurately. The output layer outputs multi-dimensional correlation feature parameters, including the rate of change of distance between personnel trajectory and the core area of ​​the thermal domain, the duration of personnel stay within the thermal domain coverage area, and the dynamic response coefficient of the thermal domain morphology to personnel trajectory. This model adopts a transfer learning approach, transferring the personnel-environment correlation model parameters from general scenarios to hot work scenarios, and fine-tuning and optimizing it using labeled datasets containing different hot work types, different personnel working postures, and different environmental disturbances, significantly improving the model's feature extraction accuracy in hot work scenarios. During model training, an adversarial training mechanism is introduced, optimizing network parameters through adversarial interaction between the generator and discriminator, improving the network's robustness to noise interference and personnel posture changes.

[0137] The early warning judgment and response module 40 determines the vacancy warning based on the extracted associated feature parameters. The system pre-sets the associated feature threshold range for hot work scenarios. This threshold range is based on the hazard level classification of hot work operations. For example, for welding operations with high combustible density, the threshold range will be set more strictly. At the same time, the system adopts a dynamic threshold adjustment mechanism, which collects environmental parameters such as wind speed, temperature, and humidity in real time, and dynamically adjusts the associated feature threshold range through a preset threshold correction formula, so that the threshold range can adapt to the hot work requirements under different environmental conditions. When the extracted associated feature parameters (e.g., the rate of change of distance between user 1 and the core area of ​​the hot zone continuously exceeds the safety threshold, and the duration of stay within the hot zone coverage area is less than the preset value) exceed the threshold range and continue for a preset duration (e.g., 5 seconds), the system will immediately trigger a vacancy warning signal.

[0138] After an early warning signal is activated, the system performs a graded warning process. Based on the degree to which the associated characteristic parameters exceed the threshold range and the duration of the exceedance, the early warning signal is classified into three levels: Level 1, Level 2, and Level 3.

[0139] After the warning signal is triggered, the system also performs post-warning verification and tracing. Real-time images of the hot work area are collected via video surveillance equipment, and image recognition algorithms are used to perform secondary verification to check whether personnel have left their posts, effectively avoiding false warnings caused by sensor malfunctions.

[0140] Through the above solution, this system overcomes the problems of low early warning accuracy, weak anti-interference ability, inability to correlate with thermal risk, and poor early warning timeliness in existing technologies. Compared with solutions based solely on personnel location or video recognition, this solution achieves a deeper understanding and more accurate judgment of "fire leaving personnel" behavior by integrating personnel trajectory and thermal distribution data and constructing a behavior-thermal-domain correlation feature model, effectively reducing false alarms and missed alarms.

[0141] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope described in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for early warning of hot work leaving personnel based on behavioral trajectory and thermal domain correlation analysis, characterized in that, Includes the following steps: Acquire multi-dimensional sensing data of the hot work area, wherein the multi-dimensional sensing data includes at least personnel location trajectory data, thermal domain distribution data, and hot work equipment operating status data; The specific steps for acquiring multi-dimensional sensing data of the hot work area include: The system employs a combined acquisition method of lidar and infrared thermal imager, where lidar acquires the three-dimensional position coordinates of personnel in real time to form trajectory data. The infrared thermal imager acquires thermal images at a preset frame rate and converts them into thermal distribution data. At the same time, it collects operational status data such as start / stop status and flame temperature through the built-in sensors of the hot work equipment. The collected multidimensional data is preprocessed, and the preprocessing steps include: noise removal of trajectory data based on Kalman filtering, background removal of thermal domain data based on adaptive threshold segmentation, and abnormal data completion based on data integrity verification, so as to achieve high-quality collection and preprocessing of multidimensional data. The personnel location trajectory data and thermal distribution data are fused based on the spatiotemporal alignment algorithm to obtain a spatiotemporal fused dataset; Construct a behavior-heat domain association feature model, and extract the association feature parameters between personnel trajectory and heat domain in the spatiotemporal fusion dataset. The association feature parameters include at least the distance change rate between personnel trajectory and the core area of ​​heat domain, the dwell time of personnel within the coverage area of ​​heat domain, and the dynamic response coefficient of heat domain morphology with personnel trajectory. The specific steps for constructing the behavior-thermal domain correlation feature model include: A network for extracting related features is constructed based on deep learning networks. The input layer of the network receives a spatiotemporal fusion dataset, the hidden layer strengthens the association weight between personnel trajectory and the core area of ​​the hot zone through an attention mechanism, and the output layer outputs multi-dimensional related feature parameters. The transfer learning approach is adopted to transfer the parameters of the personnel-environment association model in general scenarios to the hot work operation scenario. Then, the model is fine-tuned and optimized using a hot work operation sample dataset. The sample dataset contains labeled data under different hot work types, different personnel working postures, and different environmental interferences, thereby improving the feature extraction accuracy of the model in the hot work scenario. Set a threshold range for associated features in hot work scenarios, compare the extracted associated feature parameters with the threshold range, and trigger a personnel evasion warning signal when the associated feature parameters exceed the threshold range for a preset duration. The specific steps of the laser radar and infrared thermal imager linkage acquisition method include: realizing time synchronization between the laser radar and the infrared thermal imager based on a time synchronization protocol, and determining the coordinate transformation relationship between the two through a spatial calibration algorithm; The three-dimensional position coordinates collected by the lidar are mapped to the image coordinate system of the infrared thermal imager to achieve precise spatial matching between personnel position and thermal distribution. When the lidar detects that a person's trajectory is obstructed, the location of the person is predicted based on historical trajectory data and the trend of thermal distribution changes, so as to ensure the continuity of the trajectory data of the person under obstruction. The training process of the associated feature extraction network includes the following steps: An adversarial training mechanism is introduced to construct a generator and a discriminator. The generator is used to generate simulated behavior-hot domain correlation features, and the discriminator is used to distinguish between real correlation features and generated simulated features. The network parameters are optimized through adversarial training between the generator and the discriminator to improve the network's robustness to noise interference and changes in human posture. Cross-validation is used to evaluate the performance of the trained network, with feature extraction accuracy and misclassification rate as evaluation metrics. When the evaluation metrics fail to reach the preset threshold, the model is retrained by expanding the sample dataset to ensure the reliability of the model. The specific steps for setting the threshold range of associated features under hot work operation scenarios are as follows: Different threshold ranges are divided based on the hazard level of hot work operations. The hazard level is determined according to the type of hot work operation, the density of combustibles in the work area, and the ventilation conditions of the work environment. A dynamic threshold adjustment mechanism is adopted to collect environmental parameters such as wind speed, temperature and humidity in the working environment in real time. The threshold range of related features is dynamically corrected through a preset threshold correction formula, so that the threshold range can adapt to the hot work requirements under different environmental conditions and the fixed threshold can adapt to complex environmental changes.

2. The method for early warning of hot work and personnel leaving the premises based on behavioral trajectory and thermal domain correlation analysis according to claim 1, characterized in that, After the triggering of the human absence warning signal is activated, the warning classification processing steps are as follows: based on the degree and duration of the related feature parameters exceeding the threshold range, the warning signal is divided into three levels: Level 1 warning, Level 2 warning, and Level 3 warning. Differentiated early warning response measures are implemented for different early warning levels. At the first level of early warning, local audible and visual alarms are activated and early warning information is pushed to the on-site monitor's terminal. A Level II warning, based on a Level I warning, requires hot work equipment to operate at reduced power. The Level 3 warning triggers an emergency shutdown of hot work equipment and activates fire emergency equipment in a pre-prepared state, building upon the Level 2 warning, thus achieving precise grading of the warning and efficient response.

3. The method for early warning of hot work and personnel leaving the premises based on behavioral trajectory and thermal domain correlation analysis according to claim 2, characterized in that, The verification and traceability steps after the warning are as follows: After the warning signal is triggered, images of the hot work area are collected in real time through video monitoring equipment; Image recognition algorithms are used to perform secondary verification of whether people have left the area, avoiding false alarms caused by sensor misdetection; Establish an early warning and traceability database to record the time, warning level, associated characteristic parameters, on-site environmental parameters, response measures, and handling results of each early warning, forming a complete early warning and traceability chain to provide data support for subsequent operational safety optimization.

4. The method for early warning of hot work and personnel leaving the premises based on behavioral trajectory and thermal domain correlation analysis according to claim 1, characterized in that, The steps for obtaining the personnel location trajectory data include: Personnel identity association involves obtaining the identity information of workers through UWB positioning tags or RFID identification cards, binding the identity information with the corresponding personnel location trajectory data, and forming a trajectory dataset with identity identification. When a person is unaccounted for, the warning message includes the person's identity information and is simultaneously pushed to that person's mobile device, enabling precise identity association and targeted reminders for unaccounted individuals.

5. A hot work evacuation early warning system based on behavioral trajectory and thermal domain correlation analysis, characterized in that, include: The multi-dimensional sensing module includes a lidar, an infrared thermal imager, equipment sensors, and an identity recognition device, used to collect multi-dimensional sensing data and personnel identity information. The data fusion processing module is used to perform spatiotemporal alignment, data preprocessing, and identity-trajectory binding; The association feature analysis module is used to construct association feature models and extract feature parameters; The early warning determination and response module is used for threshold comparison, early warning classification, and execution of response measures; The data storage module is used to store sensing data, model parameters, and early warning record data; Each module is used to execute the steps of the hot work and personnel departure early warning method based on behavioral trajectory and thermal domain correlation analysis as described in any one of claims 1-4, and realizes data interaction through a high-speed communication bus to ensure the real-time performance and reliability of the system.

Images