Real-time estimation method and system for automobile fire heat release rate based on infrared temperature field

Abstract

The application discloses a real-time estimation method and system for the heat release rate of automobile fire based on an infrared temperature field, and belongs to the technical field of fire safety monitoring and computer vision.The application adopts a two-stage processing framework, first extracts the flame projection area and average temperature features by accurately segmenting the flame area from a 16-bit original infrared image through an improved Mask R-CNN model, and then constructs a double-parameter simplified physical model based on the heat transfer law to realize real-time estimation of the heat release rate after full-scale fire experiment calibration.The application takes into account the feature extraction capability of deep learning and the explainability of the physical model, can realize non-contact and high-precision monitoring of the heat release rate of automobile fire, effectively removes the heat background interference, and is suitable for early warning and intelligent fire-fighting linkage of automobile fire in limited spaces such as underground garages, tunnels and charging stations.

Description

Technical Field

[0001] This invention relates to the fields of fire safety monitoring and computer vision technology, and in particular to a method and system for real-time estimation of heat release rate in automotive fires based on infrared temperature fields. Background Technology

[0002] With the rapid growth in the number of new energy vehicles, car fires are becoming increasingly frequent, especially in enclosed spaces such as underground garages, urban tunnels, and charging stations. Car fires are characterized by rapid heat release, highly toxic smoke, and high difficulty in firefighting, easily causing significant casualties and property damage. Heat Release Rate (HRR) is a core parameter characterizing fire intensity. Its value directly reflects the scale and trend of energy release in a fire, providing irreplaceable guidance for personnel evacuation planning, building structural safety assessment, and firefighting and rescue decisions.

[0003] Currently, the mainstream heat release rate measurement methods in the industry are mainly divided into two categories: one is the calorimetric method based on the oxygen consumption principle, which is the gold standard for laboratory measurement, but it relies on large-scale flue gas collection and analysis equipment. The system is bulky and has high deployment costs, and it cannot achieve real-time, non-contact measurement on site. It is completely unsuitable for engineering deployment in scenarios such as underground garages and tunnels. The other is the measurement method based on radiation heat flow meters, which can achieve non-contact measurement, but it is easily affected by continuous radiation interference from high-temperature flue gas, heated vehicle bodies, and other thermal backgrounds. In particular, it will show obvious lag and overestimation during the intermittent flame extinction phase, and cannot accurately reflect the real-time open flame status.

[0004] With the development of infrared imaging and deep learning technologies, existing technologies have included research on fire monitoring and HRR estimation based on infrared images. However, significant shortcomings remain: First, existing studies mostly use compressed 8-bit display images from infrared cameras, losing the original high dynamic range temperature information of the sensors. This results in insufficient sensitivity for detecting early, weak flames, failing to meet the accuracy requirements for quantitative analysis. Second, existing end-to-end deep learning HRR regression models are mostly black-box models, lacking physical interpretability and generalization ability, and their output results cannot support fire dynamics analysis. Third, existing methods cannot effectively remove thermal background radiation interference, leading to significant estimation errors during flame decay and intermittent extinction stages, thus failing to provide reliable data for fire early warning and rescue. Summary of the Invention

[0005] The purpose of this invention is to propose a real-time estimation method and system for the heat release rate of automotive fires based on infrared temperature fields. Through a two-stage processing framework that combines high-fidelity infrared image processing with physical-guided modeling, it takes into account both the high-precision feature extraction capability of deep learning and the interpretability of physical models, and achieves real-time, non-contact, and high-precision estimation of the heat release rate of automotive fires. It effectively removes thermal background interference and provides key technical support for early warning, dynamic intervention, and accident investigation of automotive fires in confined spaces.

[0006] To achieve the above objectives, this invention provides a real-time estimation method for the heat release rate of automotive fires based on infrared temperature fields, comprising the following steps: Step S1: Infrared image acquisition and preprocessing: Infrared thermal imager with original 16-bit data output capability is used to continuously acquire temperature field data of fire scene, and the original temperature matrix is ​​preprocessed to obtain image sequence that can be input to deep learning model; Step S2: Flame instance segmentation and feature extraction: The improved Mask R-CNN instance segmentation model is used to infer the preprocessed image and output the pixel-level segmentation mask of the flame instance. Based on the segmentation mask and the original temperature matrix, the geometric and radiation features of the flame are extracted. Step S3: Construction of a physical heuristic heat release rate model: Based on the fundamental laws of heat transfer and combined with the characteristics of radiative heat transfer in automobile fires, a simplified two-parameter engineering model is constructed to establish the physical mapping relationship between flame characteristics and heat release rate. Step S4: Model parameter fitting and real-time estimation: The global parameters of the simplified model are calibrated using full-scale automobile fire test data. The calibrated parameters are then fixed in the system. Steps S1-S2 are repeated for the real-time acquired infrared images, and the real-time heat release rate value is calculated by substituting them into the model. Step S5: Visualization and Early Warning Linkage: Real-time display of fire monitoring data and heat release rate time-series curves, triggering alarms based on preset multi-level heat release rate thresholds, and linking with external fire protection systems.

[0007] Preferably, in step S1, the preprocessing steps specifically include: Step S11: Convert the CSV format temperature matrix output by the infrared thermal imager into a 16-bit single-channel infrared image; Step S12: Process the image using a percentile-based adaptive normalization method, the formula is as follows: ; in, These are the pixel values ​​of the original 16-bit image. The 1st percentile of pixel values ​​in a single frame of an image. The 99th percentile of pixel values ​​in a single frame of an image. For truncation functions, exceeding The values ​​in the interval are truncated to 0 and 1 respectively; Step S13: Scale the normalized image to 512×512 pixels using bilinear interpolation, and copy the single-channel image three times to construct a pseudo-RGB three-channel image.

[0008] Preferably, in step S2, the specific steps of the improved Mask R-CNN instance segmentation model are as follows: Step S211: Use ResNet-50 as the backbone network and combine it with the Feature Pyramid Network (FPN) to construct multi-scale feature representations; Step S212: Set up three anchor frames with aspect ratios of 1:2, 1:1, and 2:1 to adapt to different flame shapes; Step S213: Replace the original RoIPooling operation with RoIAlign to eliminate spatial quantization error of the feature map; Step S214: Employ a multi-task joint training paradigm, with the overall loss function as follows: ; in, The proposed network loss for the region includes foreground or background classification loss and bounding box regression loss. To detect head damage, This is the loss due to mask segmentation.

[0009] Preferably, in step S2, the specific steps for extracting the geometric and radiation features of the flame include: Step S221: Count the number of foreground pixels in the segmentation mask and calculate the projected area of ​​the flame. The formula is: ; in, For the first A flame instance in Projected area at time, This represents the number of foreground pixels in the mask. The actual area conversion factor corresponding to a pre-calibrated single pixel; Step S222: Index all pixels covered by the mask in the original 16-bit temperature matrix, calculate the arithmetic mean of the temperature values ​​of these pixels, and obtain the average absolute temperature of the flame surface. Step S223: If there are multiple flame instances in a single frame image, sum the projected areas of all instances, and average the temperatures by area to obtain the overall feature value of the frame.

[0010] Preferably, in step S3, the formula for the two-parameter simplified engineering model is as follows: ; in, for The predicted heat release rate at any time for The projected area of ​​the flame at any given moment. for Average absolute temperature of the flame surface at any given time To comprehensively measure the calibration coefficients, flame emissivity, Stefan-Boltzmann constant, convective heat transfer coefficient, ambient temperature, and system error are included. It is a dimensionless temperature index, with a theoretical range of values ​​of [value missing]. This corresponds to the heat transfer mechanism from pure convection to pure radiation.

[0011] Preferably, in step S4, the model parameter fitting specifically includes: Step S41: Align the flame feature time series output in step S2 with the true heat release rate time series obtained by the radiation heat flux meter or oxygen consumption calorimeter, and use linear interpolation to interpolate the feature data to the sampling time of the true heat release rate. Step S42: Construct a constrained nonlinear least squares optimization problem with the objective of minimizing the mean square error between the model's predicted values ​​and the true values: ; in, The number of aligned sample points is given by the following constraint: and ; Step S43: Use the L-BFGS-B algorithm to solve for the globally optimal parameters. .

[0012] Preferably, in step S5, the visualization and early warning linkage specifically includes: Real-time display of raw infrared or pseudo-color temperature images, flame segmentation results, flame projection area and average temperature value in the current frame, and heat release rate time-series curve; The system has three preset alarm thresholds for heat release rate: a warning threshold, an action threshold, and a danger threshold. When the real-time heat release rate exceeds any of these thresholds, an audible and visual alarm is triggered and the system displays a high-brightness indicator. Automatically capture image sequences and heat release rate data for preset durations before and after the alarm, and save them to a local database; The system sends linkage signals to the automatic fire alarm system and the fine water mist fire extinguishing system via Modbus, relay or network protocol.

[0013] A real-time estimation system for the heat release rate of automotive fires based on infrared temperature fields includes: Infrared thermal imager module: used to acquire raw 16-bit temperature matrix data; Image preprocessing module: used to perform adaptive normalization, resizing, and pseudo-RGB conversion; Flame segmentation module: Deploys a pre-trained improved Mask R-CNN model and outputs the mask, bounding box, and confidence score of flame instances; Feature extraction module: Calculates the projected area and average temperature of the flame based on the segmentation results and the original temperature matrix; HRR estimation module: Built-in calibrated model parameters, calculates heat release rate value in real time; Data storage and display module: used to display monitoring data and store raw data, intermediate results and heat release rate records; Early warning and linkage module: used to monitor the heat release rate value and trigger alarms and linkage with external systems.

[0014] Preferably, the modules communicate with each other through shared memory, message queues, or REST APIs.

[0015] Preferably, the flame segmentation module adopts an asynchronous pipeline design, with a processing speed matching the acquisition frame rate of the infrared thermal imager and an inference speed of not less than 11.5 FPS.

[0016] Therefore, the present invention employs the above-mentioned real-time estimation method and system for automotive fire heat release rate based on infrared temperature field, which has the following advantages: (1) High-fidelity temperature sensing and high detection accuracy: 16-bit raw temperature data is used to completely preserve the temperature information of high dynamic range. Combined with the adaptive normalization method, the detection sensitivity and segmentation accuracy of weak flames in the early stage of fire are significantly improved, laying the foundation for quantitative analysis. (2) Strong physical interpretability and good generalization ability: Based on the heat transfer law, a simplified physical model is constructed. The parameters have clear physical meanings, eliminating the uninterpretable defects of the black box model. Only a small amount of experimental data is needed to complete the calibration. It has excellent cross-scenario generalization ability. (3) Excellent real-time performance, meeting engineering deployment requirements: The improved Mask R-CNN model adopts an asynchronous pipeline design, and the measured inference speed can reach 11.5 FPS. Combined with a lightweight physical model, it can realize real-time calculation of heat release rate frame by frame, meeting the real-time requirements of fire dynamic monitoring. (4) Strong anti-interference ability and accurate estimation results: HRR estimation is based only on the real-time open flame area, which can effectively remove the radiation interference from the thermal background such as high-temperature smoke and heated vehicle body, solve the problem of lag and overestimation of traditional methods, and can truly reflect the real-time open flame intensity. (5) Good versatility and wide range of applicable scenarios: It is specifically optimized for the combustion characteristics of automobile fires and can be widely adapted to various restricted spaces such as underground garages, urban tunnels, and new energy vehicle charging stations. It can be transferred to fire monitoring of different models and different types of fire sources, and has strong engineering applicability. (6) High system integration and easy engineering promotion: The system integrates data collection, processing, modeling, estimation and early warning into an integrated software and hardware system, supports seamless linkage with existing fire protection systems, is easy to deploy and maintain, and has extremely high engineering promotion value and results transformation prospects.

[0017] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0018] Figure 1 This is an overall flowchart of the real-time estimation method for the heat release rate of a car fire based on an infrared temperature field in an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of the improved Mask R-CNN flame instance segmentation model in an embodiment of the present invention; Figure 3 This is a flowchart of infrared image preprocessing and feature extraction in an embodiment of the present invention; Figure 4 This is a comparison chart of HRR fitting effects in experimental verification of embodiments of the present invention, including training set and validation set; Figure 5 This is an example diagram of flame segmentation results in an embodiment of the present invention, including the initial, middle, and later stages of a fire; Figure 6 This is a schematic diagram of a typical application scenario of the system in an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0020] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0021] Example like Figure 1-6 As shown in the figure, this embodiment proposes a method and system for real-time estimation of heat release rate in automotive fires based on infrared temperature fields. The specific implementation steps are as follows: Step 1: Infrared Image Acquisition and Preprocessing Infrared thermal imagers, as core sensing devices, can directly sense the thermal radiation of object surfaces without being obstructed by visible light or affected by smoke, making them particularly suitable for non-contact measurements in fire scenarios. This invention employs an infrared thermal imager with raw data output capabilities to continuously acquire temperature field data from fire scenes. Unlike images automatically compressed by ordinary infrared cameras, the raw data fully preserves the high dynamic range temperature information captured by the sensor (typically covering 20℃ to 950℃), laying the foundation for subsequent quantitative analysis. During acquisition, the thermal imager continuously outputs the temperature matrix for each frame at a preset frame rate (e.g., 1 Hz or higher), typically stored in a CSV file or a proprietary format.

[0022] First, the CSV file is converted into a 16-bit infrared image. Because 16-bit images have an extremely high dynamic range and significant differences in temperature distribution across different fire stages, directly inputting them into a deep learning model may lead to training instability. Therefore, this invention employs a percentile-based adaptive normalization method to preprocess the original image. Specifically, for each frame, the 1st percentile and 99th percentile of its pixel values ​​are first calculated, denoted as _____. and Then, all pixel values ​​in the image are compressed to the [0,1] interval using a linear mapping, and values ​​outside this range are truncated, as shown in the following formula: ; This method can effectively suppress the interference of extreme low or high temperature anomalies, adaptively enhance the effective temperature dynamic range of each frame image, make the contrast of the flame area more prominent, and facilitate the learning of subsequent segmentation models.

[0023] After normalization, the image is scaled to a fixed size (e.g., 512×512 pixels) using bilinear interpolation to maintain spatial continuity. Since conventional convolutional neural networks typically accept three-channel input, this invention copies the single-channel normalized image three times to construct a pseudo-RGB format image, allowing direct transfer learning using model weights pre-trained on ImageNet. The preprocessed image sequence becomes the input data for subsequent deep learning models.

[0024] Step 2: Flame instance segmentation and feature extraction: To accurately extract the geometric and radiation features of flames from infrared images, this invention constructs an improved Mask R-CNN instance segmentation model. Mask R-CNN is a classic two-stage instance segmentation framework that can simultaneously output the target's bounding box, class label, and pixel-level segmentation mask. This invention optimizes the model specifically for the characteristics of infrared flame detection tasks by employing ResNet-50 as the backbone network and combining it with a Feature Pyramid Network (FPN) to construct multi-scale feature representations. The residual structure of ResNet-50 is beneficial for training deep networks, while FPN can fuse semantic information from different levels, enhancing the model's ability to detect flames with drastic scale changes (from small flames to large plumes). Based on statistical data of flame morphology, three aspect ratio anchor boxes (1:2, 1:1, 2:1) are set to accommodate different flame shapes such as vertical plumes, circular fire cores, and horizontal spread. RoIAlign is used instead of the original RoIPooling to avoid spatial quantization errors in feature maps, thereby improving the boundary accuracy of the segmentation mask, which is crucial for accurately calculating the flame area. The model outputs classification, bounding box regression, and segmentation masks in parallel, with the overall loss function being the sum of the region proposal network loss, the detector head loss, and the mask loss. Through end-to-end joint optimization, the model achieves the best balance between detection and segmentation tasks.

[0025] The improved Mask R-CNN model employs a multi-task joint training paradigm, and the overall loss function is defined as: ; in, The region proposal network loss includes foreground / background classification loss and bounding box regression loss; To detect head loss, it is used to optimize the classification and localization of flame instances; This is the loss due to mask segmentation.

[0026] The model training process is based on a self-built high-quality dataset containing at least 2,000 16-bit infrared images extracted from full-scale car fire experiments, with pixel-level flame instance annotations by experts. The dataset is divided into training and validation sets in an 8:2 ratio, and the AdamW optimizer, step-down learning rate decay, and early stopping mechanism are used to ensure efficient model convergence and prevent overfitting.

[0027] During the inference phase, the trained model performs forward propagation on each preprocessed image frame, outputting a set of flame instances and their corresponding pixel-level segmentation masks. Based on these masks, the system further extracts two key physical features for each flame instance: by counting the number of foreground pixels in the mask and multiplying it by the pre-calibrated actual area corresponding to a single pixel (obtained through calibration using a reference object of known size), the projected area of ​​the instance on the image plane is obtained (unit: m²). 2 Since a flame is a three-dimensional body, although the projected area cannot completely represent the actual surface area, it still has physical significance as an indicator of relative change over time and is easily obtained from two-dimensional images. Using the original 16-bit temperature matrix, all pixels covered by the index mask are indexed, and the arithmetic mean (unit: K) of the temperature values ​​of these pixels is calculated. This value reflects the average temperature level of the flame's radiating surface and is a key input for subsequent estimation of the heat release rate.

[0028] Based on the flame instance segmentation mask, the projected area is obtained by multiplying the number of foreground pixels by the actual area corresponding to a single pixel: ; in, For the first A flame instance in Projected area at time (m²) 2 ), This represents the number of foreground pixels in the mask. Pre-calibrated conversion factor for actual pixel area.

[0029] The features of all flame instances and their corresponding timestamps are summarized and output as a structured CSV file. Each row records the area and temperature values ​​of an instance at a given time. If multiple flame instances exist in a single frame (such as jet fire and chassis fire coexisting), the system can sum the areas of all instances and take an area-weighted average of the temperatures to obtain the overall feature value of that frame. This feature sequence becomes the standardized data interface connecting the first and second stages.

[0030] Step 3: Construction of a physical heuristic heat release rate model: The heat release rate is a core parameter describing fire intensity, and its physical essence is the thermal power of the chemical energy released by fuel combustion. According to heat transfer theory, the heat released by the fire source is transferred to the surrounding environment through both radiation and convection. The radiative component follows the Stefan-Boltzmann law, meaning that the radiative heat flux is proportional to the effective radiating area of ​​the flame and the fourth power of its surface temperature; the convective component follows Newton's law of cooling, meaning that the convective heat flux is proportional to the product of the flame surface area and its temperature difference with the environment. Therefore, the total heat release rate can be expressed as: ; in, For flame emissivity, The Stefan-Boltzmann constant is 5.67 × 10⁻⁸ W·m⁻¹. -2 ·K -4 ), for The effective radiation area of ​​the flame at any given time (m²) 2 ), The average absolute temperature of the flame surface (K). The convective heat transfer coefficient (W·m) -2 ·K -1 ), The absolute temperature of the environment (K).

[0031] In real-world fire scenarios, flame temperatures are typically much higher than ambient temperatures (e.g., T > 800 K, T0 ≈ 300 K), and The growth rate of the item is much faster than The result indicates that radiative heat transfer dominates the total heat release. Furthermore, flame emissivity... convective heat transfer coefficient These parameters are affected by complex factors such as flame pulsation, smoke obstruction, and viewing angle factors, making them difficult to measure accurately through experiments, and theoretical calculations are also extremely challenging. To improve the model's practicality while maintaining physical rationality, this invention simplifies the complete model by: ignoring the relative contribution of the linear temperature term in the convection term, or merging it with the constant in the radiation term; and simplifying all constants related to material properties and the environment (…). , , , Including potential perspective factors and measurement system errors, these are integrated into a comprehensive calibration coefficient. Introduce an undetermined index. To replace the fourth power relationship, it can more flexibly fit the nonlinear dependence between temperature and heat release rate in real data.

[0032] This leads to the simplified two-parameter engineering model: ; in, (Unit: kW·m) -2 ·K -n () is a comprehensive calibration coefficient that encapsulates all constants and uncertainties; (Dimensionless) is the temperature index, its theoretical range should be 1~4, corresponding to pure convective heat transfer ( ) and pure radiative heat transfer ( This is an extreme case. In real car fires, radiation is the dominant factor, therefore... The value should be close to 4. This simplified model retains the core physical relationships (the strong correlation between HRR and area and temperature), while simplifying the multi-parameter identification problem to the calibration of two parameters, greatly reducing the difficulty of model calibration and enhancing its engineering applicability.

[0033] Step 4: Model parameter fitting and real-time estimation: To determine the global parameters in the simplified model above and This invention utilizes existing full-scale automotive fire experimental data for parameter fitting. During the experiment, infrared video and heat release rate data calculated based on radiation heat flow meter (or oxygen consumption calorimetry) were simultaneously acquired, and flame characteristic time-series data were obtained through steps one and two. The parameter fitting process is as follows: The flame feature time series output in step two Compared with the true time series of HRR obtained by traditional methods such as radiation heat flow meters Time alignment is performed. Since the sampling times of the two data sources may not be completely consistent, a linear interpolation method is used to interpolate the feature data to the sampling time of the HRR ground truth, obtaining a data pair for fitting. .

[0034] The goal is to minimize the mean squared error between the model's predicted values ​​and the true values, and then solve for the optimal parameters. The optimization problem can be formulated as a constrained nonlinear least squares problem: ; in This represents the number of aligned sample points. Constraints. and The physical rationality of the parameters was ensured: A positive value indicates a positive correlation; The range corresponds to the theoretical boundary of the heat transfer mechanism.

[0035] The L-BFGS-B algorithm (a quasi-Newton method suitable for bounded constraints) is used for solving the problem. The initial value can be set to... This is based on the physical assumption that deflection is dominant. After the algorithm iteratively converges, the globally optimal parameters are obtained. Experiments show that, on typical full-size car fire data, the fitted n* is usually between 3.0 and 3.8, which is consistent with the expectation of radiation dominance and verifies the effectiveness of the model.

[0036] After obtaining the global parameters, they are embedded into the system's HRR estimation module. During the real-time monitoring phase, for each newly acquired image frame, the system repeats steps one and two to extract the flame projection area at the current moment. and average temperature Then, by substituting the values ​​into the calibrated simplified model, the current heat release rate can be calculated in real time. ; As time progresses, the system calculates and accumulates HRR values ​​frame by frame, forming a continuous HRR time-series curve that visually demonstrates the evolution of fire intensity.

[0037] Step 5: Visualization and Early Warning To enhance the system's usability and human-computer interaction capabilities, this invention designs a visualization and early warning module. This module receives and displays the following information in real time: the original infrared image or pseudo-color temperature image; flame segmentation results (with a mask and bounding box superimposed on the image); the flame projection area and average temperature value of the current frame; a real-time plotted HRR time-series curve (showing the trend over the past few minutes); alarm status indicator lights and a threshold setting interface.

[0038] The system presets multiple HRR alarm thresholds (such as warning threshold, action threshold, and danger threshold). When the real-time HRR exceeds any of these thresholds, the module triggers an audible and visual alarm and highlights it on the interface. Simultaneously, the system automatically captures image sequences and HRR data for a period before and after the alarm, saving them to a local database for subsequent accident analysis and model optimization. This module can also be linked with automatic fire alarm systems and fine water mist fire suppression systems via a standard communication interface to achieve intelligent fire response based on HRR.

[0039] To implement the above method, this invention constructs a real-time estimation system for the heat release rate of automotive fires based on infrared temperature fields, which mainly includes the following modules: Infrared thermal imager module: A high-performance infrared thermal imager with raw 16-bit data output capability (such as the ImageIR 8355) is selected and deployed in appropriate locations in the fire monitoring scenario (such as garage corners or tunnel sidewalls) to ensure that the field of view covers the fire source area. The thermal imager is connected to the industrial control computer via Ethernet or Camera Link interface to continuously output temperature matrix data.

[0040] Image preprocessing module: A software module running on an industrial control computer that receives thermal imager data in real time, performs adaptive normalization, size scaling and pseudo-RGB conversion, and stores the processed image in a circular buffer for subsequent modules to read.

[0041] Flame segmentation module: Deploys a pre-trained improved Mask R-CNN model, utilizes GPU acceleration to perform inference on each pre-processed image frame, and outputs the mask, bounding box, and confidence score of the flame instance. The module adopts an asynchronous pipeline design to ensure that the processing speed matches the image acquisition frame rate (actually measured up to 11.5 FPS).

[0042] Feature extraction module: Receives the segmentation results and the original temperature matrix, calculates the projected area and average temperature of each flame instance based on the mask, and fuses them into a single frame overall feature value according to certain rules (such as area-weighted average) to generate a feature sequence.

[0043] HRR estimation module: Built-in fitted simplified model parameters The output of the feature extraction module and Substitute into the formula It calculates HRR values ​​in real time. The module also includes a data cache to store HRR sequences over a recent period for plotting curves and analyzing trends.

[0044] Data storage and display module: Developed based on a graphical user interface framework (such as PyQt or Web technology), it displays infrared images, segmentation results, feature values, and HRR curves in real time. Simultaneously, it stores raw data, intermediate results, and HRR records in a local database (such as SQLite or a time-series database), supporting historical playback and export.

[0045] Early warning and linkage module: Monitors the estimated HRR value, issues an audible and visual alarm when it exceeds a preset threshold, and sends linkage signals to external systems (such as fire alarm control panels or fire extinguishing controllers) via Modbus, relays, or network protocols. The module supports multi-level threshold customization.

[0046] The modules communicate with each other through loosely coupled interfaces, which can employ shared memory, message queues, or REST APIs, facilitating system expansion and maintenance. The entire system can be deployed on fixed industrial PCs or edge computing devices to meet the installation requirements of different scenarios.

[0047] One specific implementation method is as follows: 1. Experimental dataset construction: To verify the effectiveness and robustness of the method of this invention, a full-size experimental dataset of car fires in an underground parking garage was constructed. The experiment was conducted in an underground parking garage that complies with building codes, using real cars as the fire source, covering the complete process of a fire from initiation to development and decay.

[0048] Infrared data was acquired using an infrared thermal imager (such as the ImageIR 8355) with 16-bit raw data output capability, outputting a temperature matrix in CSV format with a dynamic temperature range of 20°C to 950°C. The heat release rate (HRR) baseline was derived from radiation pyroflumeter data through inversion using a solid flame model. All infrared images and HRR data were time-synchronized (1 Hz), forming a crucial dataset for model training and validation.

[0049] The flame instance segmentation model training set contains 2,462 labeled images, divided into a training set (1,969 frames) and a test set (493 frames) in an 8:2 ratio. The image resolution is 1024×768 pixels, with a total of 2,285 flame instances, and a positive sample (containing flames) ratio of 59.4%. The HRR fitting and validation dataset contains approximately 10,000 images, covering the entire fire initiation, growth, intense burning, and decay stages.

[0050] 2. Training and Evaluation of the Flame Instance Segmentation Model: The improved Mask R-CNN model was trained on an NVIDIA GeForce RTX 2070 GPU using the PyTorch 1.12.1 framework. Input images were adaptively percentile normalized (based on the 1st and 99th percentiles), scaled to 512×512 pixels, and converted to a pseudo-RGB three-channel format. Training employed the AdamW optimizer with an initial learning rate of 0.001, a batch size of 4, a maximum of 50 training epochs, and an early stopping mechanism (termination if there was no improvement in performance on the validation set after 15 consecutive epochs). Data augmentation included random horizontal flipping (probability 0.5) and random brightness-contrast adjustments. The model's evaluation results on the test set are shown in Table 1 below. Table 1. Evaluation results of the model on the test set.

[0051] Experimental results show that the model has a detection accuracy of 87.36% and a recall rate of 91.35% at an IoU threshold of 0.5, indicating that the model has a low risk of missing detection of flame targets and meets the deployment requirements of actual fire monitoring scenarios.

[0052] 3. Key Module Contribution Analysis: To quantify the contribution of each technical module to the system performance, four independent experiments were designed to evaluate the effects of 16-bit high dynamic range input, adaptive percentile normalization, RoIAlign operation, and Feature Pyramid Network (FPN). The experimental results are shown in Table 2 below: Table 2 Evaluation results of four independent experimental groups

[0053] Experimental results show that the 16-bit high dynamic range input contributes most to the flame segmentation accuracy, reducing mAP by 3.13 percentage points, AP75 by 4.23 percentage points, and recall by 4.03 percentage points after removal, verifying the crucial role of the original high dynamic range temperature information in flame boundary detail extraction and weak flame recall. The RoIAlign operation significantly improves segmentation accuracy under high IoU thresholds, reducing AP75 by 5.27 percentage points after removal. The feature pyramid network makes a significant contribution to large-size flame detection, reducing APL by 3.38 percentage points after removal. These results collectively validate the effectiveness of the modules proposed in this invention.

[0054] 4. HRR estimation model parameter fitting and validation: Based on the flame projection area extracted in the first stage With mean radiation temperature Temporal features, using a simplified model Perform parameter fitting. The optimization objective is to minimize the mean square error between the predicted HRR and the actual HRR calculated based on the radiative heat flux meter, with the following parameter constraints: , The L-BFGS-B algorithm is used to solve this problem.

[0055] On a full-scale car fire test dataset (training set), the optimal parameters were obtained after optimization convergence: k≈6.48×10⁻⁶. -9 n≈3.45. n≈3.45 is significantly greater than 1 and close to 4, indicating that radiative heat transfer dominates in automotive fires, consistent with classical fire dynamics theory. On the training set, the model's predicted HRR and the actual HRR show a high degree of consistency in the trends during the active flame phases (growth phase and intense combustion phase); on the validation set (another independent fire experiment dataset), good trend consistency is also observed, validating the model's cross-scenario generalization ability.

[0056] In particular, during the weak or brief extinguishing phase of combustion (such as when the thermal radiation heat flux meter data still maintains a certain value), the HRR estimated by the method of this invention rapidly approaches zero. This difference stems from the different mechanisms by which the two methods handle the thermal background: the thermal radiation heat flux meter method measures the total radiative heat flux, including thermal backgrounds such as high-temperature flue gas and heated vehicle bodies, resulting in a lag and overestimation; the method of this invention only responds to the open flame area segmented in the current frame, effectively removing thermal background interference and providing more timely "real-time open flame intensity" information. This characteristic has significant application value in dynamic fire early warning and intervention decision-making.

[0057] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for real-time estimation of heat release rate in automotive fires based on infrared temperature fields, characterized in that, Includes the following steps: Step S1: Infrared image acquisition and preprocessing: Infrared thermal imager with original 16-bit data output capability is used to continuously acquire temperature field data of fire scene, and the original temperature matrix is ​​preprocessed to obtain image sequence that can be input to deep learning model; Step S2: Flame instance segmentation and feature extraction: The improved Mask R-CNN instance segmentation model is used to infer the preprocessed image and output the pixel-level segmentation mask of the flame instance. Based on the segmentation mask and the original temperature matrix, the geometric and radiation features of the flame are extracted. Step S3: Construction of a physical heuristic heat release rate model: Based on the fundamental laws of heat transfer and combined with the characteristics of radiative heat transfer in automobile fires, a simplified two-parameter engineering model is constructed to establish the physical mapping relationship between flame characteristics and heat release rate. Step S4: Model parameter fitting and real-time estimation: The global parameters of the simplified model are calibrated using full-scale automobile fire test data. The calibrated parameters are then fixed in the system. Steps S1-S2 are repeated for the real-time acquired infrared images, and the real-time heat release rate value is calculated by substituting them into the model. Step S5: Visualization and Early Warning Linkage: Real-time display of fire monitoring data and heat release rate time-series curves, triggering alarms based on preset multi-level heat release rate thresholds, and linking with external fire protection systems.

2. The method for real-time estimation of heat release rate in automotive fires based on infrared temperature field according to claim 1, characterized in that: In step S1, the specific preprocessing steps include: Step S11: Convert the CSV format temperature matrix output by the infrared thermal imager into a 16-bit single-channel infrared image; Step S12: Process the image using a percentile-based adaptive normalization method, the formula is: ； in, These are the pixel values ​​of the original 16-bit image. The 1st percentile of pixel values ​​in a single frame of an image. The 99th percentile of pixel values ​​in a single frame of an image. For truncation functions, exceeding The values ​​in the interval are truncated to 0 and 1 respectively; Step S13: Scale the normalized image to 512×512 pixels using bilinear interpolation, and copy the single-channel image three times to construct a pseudo-RGB three-channel image.

3. The method for real-time estimation of heat release rate in automotive fires based on infrared temperature field according to claim 1, characterized in that: In step S2, the specific steps of the improved Mask R-CNN instance segmentation model are as follows: Step S211: Use ResNet-50 as the backbone network and combine it with the Feature Pyramid Network (FPN) to construct multi-scale feature representations; Step S212: Set up three anchor frames with aspect ratios of 1:2, 1:1, and 2:1 to adapt to different flame shapes; Step S213: Replace the original RoIPooling operation with RoIAlign to eliminate spatial quantization error of the feature map; Step S214: Employ a multi-task joint training paradigm, with the overall loss function as follows: ； in, The proposed network loss for the region includes foreground or background classification loss and bounding box regression loss. To detect head damage, This is the loss due to mask segmentation.

4. The method for real-time estimation of heat release rate in automotive fires based on infrared temperature field according to claim 3, characterized in that: In step S2, the specific steps for extracting the geometric and radiation features of the flame include: Step S221: Count the number of foreground pixels in the segmentation mask and calculate the projected area of ​​the flame. The formula is: ； in, For the first A flame instance in Projected area at time, This represents the number of foreground pixels in the mask. The actual area conversion factor corresponding to a pre-calibrated single pixel; Step S222: Index all pixels covered by the mask in the original 16-bit temperature matrix, calculate the arithmetic mean of the temperature values ​​of these pixels, and obtain the average absolute temperature of the flame surface. Step S223: If there are multiple flame instances in a single frame image, sum the projected areas of all instances, and average the temperatures by area to obtain the overall feature value of the frame.

5. The method for real-time estimation of heat release rate in automotive fires based on infrared temperature field according to claim 1, characterized in that: In step S3, the formula for the two-parameter simplified engineering model is as follows: ； in, for The predicted heat release rate at any time for The projected area of ​​the flame at any given moment. for Average absolute temperature of the flame surface at any given time To comprehensively measure the calibration coefficients, flame emissivity, Stefan-Boltzmann constant, convective heat transfer coefficient, ambient temperature, and system error are included. It is a dimensionless temperature index, with a theoretical range of values ​​of [value missing]. This corresponds to the heat transfer mechanism from pure convection to pure radiation.

6. The method for real-time estimation of heat release rate in automotive fires based on infrared temperature field according to claim 5, characterized in that: In step S4, model parameter fitting specifically includes: Step S41: Align the flame feature time series output in step S2 with the true heat release rate time series obtained by the radiation heat flux meter or oxygen consumption calorimeter, and use linear interpolation to interpolate the feature data to the sampling time of the true heat release rate. Step S42: Construct a constrained nonlinear least squares optimization problem with the objective of minimizing the mean square error between the model's predicted values ​​and the true values: ； in, The number of aligned sample points is given by the following constraint: and ; Step S43: Use the L-BFGS-B algorithm to solve for the globally optimal parameters. .

7. The method for real-time estimation of heat release rate in automotive fires based on infrared temperature field according to claim 1, characterized in that: In step S5, the linkage between visualization and early warning specifically includes: Real-time display of raw infrared or pseudo-color temperature images, flame segmentation results, flame projection area and average temperature value in the current frame, and heat release rate time-series curve; The system has three preset alarm thresholds for heat release rate: a warning threshold, an action threshold, and a danger threshold. When the real-time heat release rate exceeds any of these thresholds, an audible and visual alarm is triggered and the system displays a high-brightness indicator. Automatically capture image sequences and heat release rate data for preset durations before and after the alarm, and save them to a local database; The system sends linkage signals to the automatic fire alarm system and the fine water mist fire extinguishing system via Modbus, relay or network protocol.

8. A real-time estimation system for the heat release rate of an automotive fire based on an infrared temperature field, used to implement the real-time estimation method for the heat release rate of an automotive fire based on an infrared temperature field as described in any one of claims 1-7, characterized in that, include: Infrared thermal imager module: used to acquire raw 16-bit temperature matrix data; Image preprocessing module: used to perform adaptive normalization, resizing, and pseudo-RGB conversion; Flame segmentation module: Deploys a pre-trained improved Mask R-CNN model and outputs the mask, bounding box, and confidence score of flame instances; Feature extraction module: Calculates the projected area and average temperature of the flame based on the segmentation results and the original temperature matrix; HRR estimation module: Built-in calibrated model parameters, calculates heat release rate value in real time; Data storage and display module: used to display monitoring data and store raw data, intermediate results and heat release rate records; Early warning and linkage module: used to monitor the heat release rate value and trigger alarms and linkage with external systems.

9. The real-time estimation system for automotive fire heat release rate based on infrared temperature field according to claim 8, characterized in that: The modules communicate with each other through shared memory, message queues, or REST APIs.

10. The real-time estimation system for automotive fire heat release rate based on infrared temperature field according to claim 8, characterized in that: The flame segmentation module adopts an asynchronous pipeline design, with a processing speed matching the acquisition frame rate of the infrared thermal imager and an inference speed of no less than 11.5 FPS.

Images