Early infrared flame feature extraction and early warning method and system for new energy vehicle thermal runaway

Abstract

This invention relates to the field of fire monitoring, early warning, and image processing technology, specifically to a method and system for early warning of thermal runaway in new energy vehicles using infrared flame features. By directly reading the raw 16-bit temperature matrix from an infrared thermal imager, a dynamic range compression mapping method based on adaptive stretching from 1% to 99% quantiles is proposed, significantly improving the contrast between weak flames and the background. Furthermore, an optimized Mask R-CNN segmentation network is constructed to accurately extract the apparent area and surface equivalent temperature of the flame. Finally, the feature time-series data is used for fire source power inversion and multi-level early warning. The system of this invention has clearly defined modules, integrating data acquisition, preprocessing, segmentation, and feature calculation into a single unit, outputting structured feature data for easy integration with downstream fire intensity analysis modules. It fully preserves the dynamic temperature gradient from 20℃ to 950℃, exhibiting high sensitivity to small flames in the early stages of thermal runaway, providing a reliable early warning method for scenarios such as new energy vehicle charging stations and underground parking garages.

Description

Technical Field

[0001] This invention relates to the field of fire monitoring, early warning and image processing technology, specifically to a method and system for early warning of thermal runaway in new energy vehicles by extracting infrared flame features. Background Technology

[0002] With the rapid growth in the number of new energy vehicles, fires caused by thermal runaway of power batteries are occurring frequently, especially in confined spaces such as underground garages and charging stations. These fires are characterized by their suddenness, rapid spread, and highly toxic smoke. The early stage of thermal runaway often manifests as an abnormal increase in local temperature and the appearance of small flames. If the flames can be accurately identified and their intensity assessed at this stage, valuable time can be gained for personnel evacuation and automatic fire extinguishing.

[0003] Infrared thermal imaging technology has become an important means of fire monitoring due to its advantages such as non-contact operation and resistance to smoke interference. However, existing infrared monitoring systems generally have a fundamental defect: in order to be compatible with conventional display and storage, the infrared camera internally performs automatic gain control and dynamic range compression on the raw sensor data, converting the 16-bit raw radiation data into an 8-bit image. This process results in the loss of a large amount of temperature details, specifically: (1) the small flames in the early stage of thermal runaway (the temperature may only be tens of degrees higher than the background) have extremely low contrast in the 8-bit image and are difficult to detect; (2) the temperature information of high-temperature areas is truncated or saturated, and cannot reflect the actual fire temperature; (3) the flame boundaries are blurred, affecting the accuracy of subsequent feature extraction.

[0004] A few studies have attempted to use raw 16-bit data directly, but there is a lack of effective preprocessing methods for wide dynamic range images, and directly inputting them into deep learning models can lead to training instability. In addition, existing methods mostly focus on the final inversion of fire source power, while the research on flame feature extraction itself is not in-depth enough. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for early warning of early infrared flame features in thermal runaway of new energy vehicles, aiming to solve the problem that existing infrared thermal imaging technology loses temperature details and is difficult to capture early weak flames due to the use of compressed images.

[0006] To achieve the above objectives, this invention provides a method for early warning of thermal runaway in new energy vehicles by extracting infrared flame features, comprising the following steps:

[0007] Step 1: Obtain the raw 16-bit temperature matrix data output by the infrared thermal imager. The temperature matrix completely preserves the full dynamic temperature information from 20℃ to 950℃ captured by the infrared thermal imager sensor.

[0008] Step 2: Convert the temperature matrix into a 16-bit grayscale image, and use an adaptive stretching method based on the 1% quantile and 99% quantile to perform dynamic range compression on the image, suppress extreme background noise and enhance the contrast of weak flame areas, to obtain the preprocessed image;

[0009] Step 3: Construct an optimized Mask R-CNN instance segmentation model to perform pixel-level analysis of the flame region in the compressed image and output the flame mask;

[0010] Step 4: Based on the flame mask and the original 16-bit temperature matrix data, perform flame feature extraction and calculation, calculate the apparent area and surface equivalent temperature of the flame, and output them as flame features.

[0011] Step 5: Fire source power inversion, and trigger graded early warning according to preset thresholds.

[0012] Optionally, during the dynamic range compression of the image using the adaptive stretching method described in step 2, the 1st quantile P1 and the 99th quantile P2 of the pixel values ​​in each frame are statistically analyzed. 99 The pixel values ​​are linearly mapped to the [0,1] interval using the following formula:

[0013]

[0014] in, Let be the original pixel value of any pixel in the grayscale image. The clip function truncates values ​​outside the [0,1] range to 0 or 1, representing the normalized pixel value.

[0015] Optional, 1st quantile P1 and 99th quantile P 99 The selection criteria are as follows: the pixel proportion of the flame area in the entire image during the initial stage of thermal runaway is less than 0.5%, and low-temperature background noise is effectively removed with P1 as the lower bound. 99 To preserve the high-temperature details of the flame at the upper limit and suppress abnormal spikes from the sensor.

[0016] Optionally, in step 2, after obtaining the normalized image through the adaptive stretching method, the image is further scaled to 512×512 pixels using bilinear interpolation, and the single-channel image is converted to a three-channel pseudo-RGB format through channel duplication to adapt to the input requirements of the optimized Mask R-CNN instance segmentation model.

[0017] Optionally, the process of building the optimized Mask R-CNN instance segmentation model in step 3 includes the following steps:

[0018] Step 3.1: Construct a multi-scale feature extraction backbone network using ResNet-50 and a feature pyramid network;

[0019] Step 3.2: Set anchor frames with aspect ratios of 1:2, 1:1, and 2:1 to match common flame shapes;

[0020] Step 3.3: Replace RoIPooling with RoIAlign to eliminate spatial quantization error;

[0021] Step 3.4: Set the multi-task loss function as follows .

[0022] Optionally, the formula for calculating the apparent area of ​​the flame in step 4 is:

[0023] A j (t)=N pixel ×α

[0024] Among them, A j (t) represents the projected area of ​​the j-th flame instance at time t, N pixel α is the number of foreground pixels in the mask corresponding to the flame instance, and α is a pre-calibrated conversion factor for the actual pixel area.

[0025] The surface equivalent temperature is obtained by indexing the temperature values ​​of the corresponding pixels of the mask from the original temperature matrix and calculating the arithmetic mean.

[0026] Optionally, in step 4, when there are multiple flame instances in a single frame image, the apparent areas of each instance are summed, and the temperature is weighted by area to obtain the comprehensive flame features of the frame.

[0027] Optionally, the fire source power inversion process in step 5 adopts the simplified radiation expression HRR = k·A·T n Where HRR is the heat release rate obtained by inversion, k is the comprehensive calibration coefficient calibrated by experimental data, A is the apparent area of ​​the flame, T is the surface equivalent temperature (in Kelvin K), and n is the temperature index calibrated by experimental data.

[0028] This invention also proposes an early warning system for early thermal runaway infrared flame features in new energy vehicles, using the aforementioned method for early thermal runaway infrared flame feature extraction and warning, comprising:

[0029] The raw data acquisition module is used to read the 16-bit temperature matrix of the infrared thermal imager;

[0030] The image preprocessing module is used to perform adaptive stretching, scaling, and format conversion;

[0031] The flame analysis module deploys a pre-trained optimized Mask R-CNN instance segmentation model and outputs a flame mask.

[0032] The feature calculation module is used to calculate the apparent area of ​​the flame and the surface equivalent temperature.

[0033] The early warning decision module is used to trigger audible and visual alarms and linkage control based on flame characteristics and fire source power thresholds.

[0034] The flame segmentation module adopts an asynchronous pipeline architecture, and the model inference speed is matched with the image acquisition frame rate.

[0035] This invention provides a method and system for early warning of thermal runaway in new energy vehicles by extracting infrared flame features. By directly reading the raw 16-bit temperature matrix from an infrared thermal imager, a dynamic range compression mapping method based on adaptive stretching from 1% to 99% quantiles is proposed, significantly improving the contrast between weak flames and the background. Furthermore, an optimized Mask R-CNN segmentation network is constructed to accurately extract the apparent area and surface equivalent temperature of the flame. Finally, the feature time-series data is used for fire source power inversion and multi-level early warning. The system's modules have clearly defined functions, integrating data acquisition, preprocessing, segmentation, and feature calculation into a single unit. It outputs structured feature data, facilitating integration with downstream fire intensity analysis modules. It fully preserves the dynamic temperature gradient from 20℃ to 950℃, exhibiting high sensitivity to small flames in the early stages of thermal runaway, providing a reliable early warning method for scenarios such as new energy vehicle charging stations and underground parking garages. Attached Figure Description

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

[0037] Figure 1 This is a flowchart illustrating the steps of an early warning method for infrared flame feature extraction in the early stage of thermal runaway in new energy vehicles according to the present invention.

[0038] Figure 2 This is a schematic diagram of the structure of the optimized Mask R-CNN instance segmentation model of the present invention.

[0039] Figure 3 These are example diagrams illustrating flame segmentation results at different fire stages in specific embodiments of the present invention. Detailed Implementation

[0040] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0041] This invention provides a method and system for early warning of thermal runaway in new energy vehicles by extracting infrared flame features, comprising the following steps:

[0042] Step 1: Obtain the raw 16-bit temperature matrix data output by the infrared thermal imager. The temperature matrix completely preserves the full dynamic temperature information from 20℃ to 950℃ captured by the infrared thermal imager sensor.

[0043] Step 2: Convert the temperature matrix into a 16-bit grayscale image, and use an adaptive stretching method based on the 1% quantile and 99% quantile to perform dynamic range compression on the image, suppress extreme background noise and enhance the contrast of weak flame areas, to obtain the preprocessed image;

[0044] Step 3: Construct an optimized Mask R-CNN instance segmentation model to perform pixel-level analysis of the flame region in the compressed image and output the flame mask;

[0045] Step 4: Based on the flame mask and the original 16-bit temperature matrix data, perform flame feature extraction and calculation, calculate the apparent area and surface equivalent temperature of the flame, and output them as flame features.

[0046] Step 5: Fire source power inversion, and trigger graded early warning according to preset thresholds.

[0047] The specific methods and procedures are as follows: Figure 1 As shown, the following provides further explanation in conjunction with the execution steps:

[0048] Step 1 involves acquiring high dynamic range infrared data. An infrared thermal imager with raw data output capability is used in step 1 to directly output the temperature matrix captured by the sensor in CSV text format. Each element of this matrix represents the temperature value (unit: °C) of the corresponding point, with a dynamic range of 20 °C to 950 °C, fully preserving the temperature evolution details during thermal runaway. Data is continuously acquired at a frame rate of 1 Hz or higher.

[0049] Step 2 involves adaptive stretching dynamic range compression based on quantile statistics, as well as image size standardization and format adaptation. Specifically, the 16-bit grayscale image obtained from the temperature matrix conversion has an extremely wide pixel value range (20~950℃), and the temperature distribution varies significantly at different fire stages. Directly feeding it into a deep learning model can lead to gradient instability. Therefore, this invention designs an adaptive stretching method:

[0050] For each frame of the image, calculate its 1st quantile P1 and 99th quantile P2. 99 Selection criteria: In the early stages of thermal runaway, the flame region (high-temperature pixels) typically accounts for less than 0.5% of the total image pixels. Using P1 as the lower limit can eliminate extremely low-temperature background noise and sensor anomalies. 99 The upper limit can suppress abnormal spikes in the sensor and highlight the flame area, enhancing the contrast between the flame and the background.

[0051] The mapping formula is as follows:

[0052]

[0053] The `clip` function truncates values ​​outside the range [0,1]. This method calculates the stretching parameters independently for each frame, exhibiting strong adaptability and effectively enhancing the contrast between the flames and the background.

[0054] Furthermore, the normalized images were scaled to 512×512 pixels using bilinear interpolation to maintain spatial continuity. The single-channel images were then copied to a three-channel pseudo-RGB format to facilitate transfer learning using the weights of a model pre-trained on ImageNet.

[0055] Step 3 is the flame region parsing process of the improved optimized Mask R-CNN instance segmentation model.

[0056] To achieve accurate identification and localization of flame regions, this invention constructs an optimized Mask R-CNN instance segmentation model (detailed structure as follows). Figure 2 (As shown). Mask R-CNN is a classic two-stage instance segmentation framework that can simultaneously output the bounding box, class label, and pixel-level segmentation mask of the target. This invention optimizes the model for the characteristics of infrared flame detection tasks as follows:

[0057] (1) Multi-scale feature extraction: ResNet-50 is used as the backbone network and combined with Feature Pyramid Network (FPN) to construct multi-scale feature representation. The residual structure of ResNet-50 is conducive to training deep networks, while FPN can fuse semantic information at different levels and enhance the model's ability to detect flames with drastic scale changes (from small flames to large plumes of fire).

[0058] (2) Flame shape adaptation anchor frame: Based on the statistical data of flame shape, three width-to-height ratio anchor frames (1:2, 1:1, 2:1) are set to adapt to different flame shapes such as vertical plume, circular fire core and horizontal spread.

[0059] (3) Boundary accuracy optimization: RoIAlign is used instead of the original RoIPooling to avoid spatial quantization error of feature map, thereby improving the boundary accuracy of segmentation mask, which is crucial for accurate calculation of flame area.

[0060] (4) Multi-task joint training: The model outputs classification, bounding box regression and segmentation mask in parallel. The overall loss function is the sum of the region proposal network loss, the detector head loss and the mask loss.

[0061]

[0062] Through end-to-end joint optimization, the model achieves the best balance between detection and segmentation tasks.

[0063] The model was trained on a self-built, high-quality dataset containing over 2,000 16-bit infrared images extracted from full-scale car fire experiments, with pixel-level flame instance annotations by experts. The dataset was divided into training and validation sets in an 8:2 ratio. 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 (terminating training if there is no improvement in validation set performance for 15 consecutive epochs) to ensure efficient model convergence and prevent overfitting.

[0064] Flame feature extraction and calculation are performed in step 4.

[0065] 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 the key physical features of each flame instance:

[0066] (1) Flame Projection Area: The projected area of ​​the instance on the image plane is obtained by multiplying the number of foreground pixels in the mask by the pre-calibrated actual area of ​​a single pixel. The calculation formula is:

[0067] Aj(t) = Npixel × α

[0068] Where Aj(t) is the projected area of ​​the j-th flame instance at time t (unit: m²), Npixel is the number of foreground pixels in the mask corresponding to the flame instance, and α is the pre-calibrated pixel actual area conversion coefficient (obtained by calibration using a reference object of known size).

[0069] (2) Average flame radiation temperature: Using the original 16-bit temperature matrix, index all pixels covered by the mask and calculate the arithmetic mean (unit: K) of the temperature values ​​of these pixels. This value reflects the average temperature level of the flame radiation surface.

[0070] When multiple flame instances exist in a single frame (such as jet fire and chassis fire coexisting), the system sums the projected areas of all instances and averages the temperatures by area to obtain the overall feature value of the frame:

[0071]

[0072]

[0073] The features of all flame instances and their corresponding timestamps are summarized and output as a structured CSV file, with each row recording the area and temperature values ​​of an instance at a specific moment. This feature sequence serves as the standard data interface connecting this method with the subsequent fire intensity analysis module.

[0074] Step 5: Perform fire source power inversion and flame characteristic application.

[0075] The extracted flame projection area and mean radiation temperature can be applied to various fire intensity analysis scenarios. A typical application is the construction of an analytical expression for the heat release rate based on a solid flame model.

[0076]

[0077] Where HRR is the heat release rate obtained by inversion, k is the comprehensive calibration coefficient calibrated by experimental data, A is the apparent area of ​​the flame, T is the surface equivalent temperature (in Kelvin K), and n is the temperature index calibrated by experimental data.

[0078] Furthermore, the present invention provides specific embodiments for auxiliary illustration, and the embodiments are analyzed and explained below in conjunction with the specific execution process:

[0079] 1. Infrared data acquisition and preprocessing

[0080] In a full-scale underground parking garage fire experiment, an ImageIR 8355 infrared thermal imager was used to acquire data. The imager was deployed above the parking spaces and continuously output temperature matrices in CSV format at a frame rate of 1 Hz. The experiment acquired approximately 10,000 images, covering the complete process of a fire from its initiation to its decay.

[0081] In the original temperature matrix, the temperature in the non-flame area is approximately 20℃~40℃, while the temperature in the flame area can reach 800℃~950℃. Using the adaptive stretching method in step 2 of this invention, the 1% quantile P1 and 99% quantile P2 are independently calculated for each frame of the image. 99 Statistics show that P1 is typically located in the range of 20℃ to 30℃. 99 It is located in the range of 600℃ to 950℃.

[0082] Experimental results show that adaptive stretching significantly improves the visual contrast of the flame region. Before stretching, the flame region has low contrast in the 16-bit grayscale image, with little difference in grayscale from the background, and the flame boundary is blurred. After stretching, the flame region is effectively enhanced, forming a clear contrast with the background, and the flame boundary is clearly distinguishable, providing high-quality input for subsequent flame segmentation. This method, through independent adaptive stretching for each frame, can adapt to the temperature distribution changes at different fire stages (initiation, development, and decay), maintaining a stable flame enhancement effect.

[0083] 2. Training and Evaluation of Flame Instance Segmentation Model

[0084] The model was trained on an NVIDIA GeForce RTX 2070 GPU using the PyTorch 1.12.1 framework. Input images were adaptively stretched to 512×512 pixels and converted to a pseudo-RGB three-channel format. Training parameters included the AdamW optimizer, initial learning rate of 0.001, batch size of 4, maximum training epochs of 50, and an early stopping mechanism (termination if there is no improvement on the validation set after 15 consecutive epochs).

[0085] The training set contains 1,969 labeled images, and the validation set contains 493 frames. The model's evaluation results on the validation set are shown in Table 1.

[0086]

[0087] A recall rate of 91.35% indicates that the model has a low risk of missing detection for flame targets and can effectively detect early and weak flames; an AP50 of 87.36% indicates that it has reliable detection performance at the 0.5 IoU threshold commonly used in industry and meets the requirements of actual deployment.

[0088] Figure 3 The model's segmentation results at different fire stages are shown: (a) In the early stage of the fire, the flame area is small and the temperature is low, and the model can still segment accurately; (b) In the middle stage of the fire, the flame develops into a large-area plume, and the model can completely segment the flame area; (c) In the later stage of the fire, the flame weakens and becomes irregular in shape, but the model still maintains stable segmentation ability.

[0089] 3. Flame Feature Extraction and Verification

[0090] Based on the segmentation mask output by the model, calculate the projected area and average radiation temperature of the flame in step 5. For Figure 3 (a) shows an initial fire image with a flame projection area of ​​approximately 1.22 m² and an average radiant temperature of approximately 890 K; for Figure 3 (b) shows a mid-fire image with a flame projection area of ​​approximately 4.85 m² and an average radiant temperature of approximately 1,120 K; for Figure 3(c) shows a post-fire image with a flame projection area of ​​approximately 2.35 m² and an average radiant temperature of approximately 950 K.

[0091] The extracted feature sequences are consistent with the trends of fire development stages: initially, the area is small and the temperature is moderate; in the middle stage, the area reaches its peak and the temperature is highest; in the later stage, the area decreases and the temperature drops. This verifies that the flame features extracted by this invention can accurately reflect the fire evolution process.

[0092] 4. Key Module Contribution Analysis

[0093]

[0094] To quantify the contribution of each technical module, four sets of comparative experiments were designed:

[0095] Experimental results show that: (1) 16-bit high dynamic range input contributes the most to the flame segmentation accuracy. After removal, mAP decreases by 3.13 percentage points and recall decreases by 4.03 percentage points, which verifies the key role of the original high dynamic range temperature information in flame boundary detail extraction and weak flame recall; (2) RoIAlign operation significantly improves the segmentation accuracy under high IoU threshold; (3) Feature pyramid network makes an important contribution to large-size flame detection.

[0096] 5. Application

[0097] The flame features extracted in this invention are applied to the estimation of heat release rate, and an analytical model HRR = k·A·T is constructed. n Through parameter fitting, n≈3.45 was obtained, which is close to the theoretical value of 4 for radiation dominance, verifying the accuracy of feature extraction. On the independent validation set, the model's predicted HRR showed a high degree of consistency with the baseline value, demonstrating that the flame feature extraction method of this invention can provide a reliable data foundation for subsequent fire intensity analysis.

[0098] In summary, compared with the prior art, the present invention has the following beneficial effects:

[0099] 1) High dynamic range information is fully preserved: Directly process 16-bit raw infrared data and fully preserve temperature information from 20℃ to 950℃, avoiding the loss of high-temperature details and the failure to detect weak flames caused by traditional 8-bit image compression.

[0100] 2) Adaptive contrast enhancement: Based on the adaptive stretching method of 1%-99th percentile, the stretching parameters can be independently adjusted for the dynamic range of each frame of the image, effectively suppressing extreme outlier interference, while enhancing the contrast of the flame area.

[0101] 3) High flame segmentation accuracy: The improved Mask R-CNN model, combined with FPN multi-scale features, flame morphology-adaptive anchor boxes, RoIAlign boundary optimization and other technologies, achieves a recall rate of 91.35% and an mAP of 71.16% on the test set, meeting the deployment requirements of actual fire monitoring scenarios.

[0102] 4) High reliability of feature extraction: The projected area and average radiation temperature of the flame are extracted based on pixel-level segmentation mask and original temperature matrix, avoiding the threshold dependence and noise sensitivity problems of traditional image processing methods.

[0103] 5) High system integration: It integrates data acquisition, preprocessing, segmentation, and feature calculation, and outputs structured feature data, which is convenient for interfacing with downstream fire intensity analysis modules.

[0104] The above description discloses only one preferred embodiment of the present invention, and should not be construed as limiting the scope of the present invention. Those skilled in the art will understand that all or part of the processes of the above embodiments can be implemented, and equivalent changes made in accordance with the claims of the present invention are still within the scope of the invention.

Claims

1. A method for early warning of thermal runaway in new energy vehicles by extracting infrared flame features, characterized in that, Includes the following steps: Step 1: Obtain the raw 16-bit temperature matrix data output by the infrared thermal imager. The temperature matrix completely preserves the full dynamic temperature information from 20℃ to 950℃ captured by the infrared thermal imager sensor. Step 2: Convert the temperature matrix into a 16-bit grayscale image, and use an adaptive stretching method based on the 1% quantile and 99% quantile to perform dynamic range compression on the image, suppress extreme background noise and enhance the contrast of weak flame areas, to obtain the preprocessed image; Step 3: Construct an optimized Mask R-CNN instance segmentation model to perform pixel-level analysis of the flame region in the preprocessed image and output the flame mask; Step 4: Based on the flame mask and the original 16-bit temperature matrix data, perform flame feature extraction and calculation, calculate the apparent area and surface equivalent temperature of the flame, and output them as flame features. Step 5: Fire source power inversion, and trigger graded early warning according to preset thresholds.

2. The method for early infrared flame feature extraction and warning of thermal runaway in new energy vehicles as described in claim 1, characterized in that, In step 2, during the dynamic range compression of the image using the adaptive stretching method, the 1st quantile P1 and the 99th quantile P2 of the pixel values ​​in each frame are statistically analyzed. 99 The pixel values ​​are linearly mapped to the [0,1] interval using the following formula: in, Let be the original pixel value of any pixel in the grayscale image. The clip function truncates values ​​outside the [0,1] range to 0 or 1, representing the normalized pixel value.

3. The method for early infrared flame feature extraction and warning of thermal runaway in new energy vehicles as described in claim 2, characterized in that, 1% quantile P1 and 99% quantile P 99 The selection criteria are as follows: the pixel proportion of the flame area in the entire image during the initial stage of thermal runaway is less than 0.5%, and low-temperature background noise is effectively removed with P1 as the lower bound. 99 To preserve the high-temperature details of the flame at the upper limit and suppress abnormal spikes from the sensor.

4. The method for early infrared flame feature extraction and warning of thermal runaway in new energy vehicles as described in claim 1, characterized in that, In step 2, after obtaining the normalized image through the adaptive stretching method, the image is further scaled to 512×512 pixels using bilinear interpolation, and the single-channel image is converted to a three-channel pseudo-RGB format through channel duplication to meet the input requirements of the optimized Mask R-CNN instance segmentation model.

5. The method for early infrared flame feature extraction and warning of thermal runaway in new energy vehicles as described in claim 1, characterized in that, Step 3, which involves building an optimized Mask R-CNN instance segmentation model, includes the following steps: Step 3.1: Construct a multi-scale feature extraction backbone network using ResNet-50 and a feature pyramid network; Step 3.2: Set anchor frames with aspect ratios of 1:2, 1:1, and 2:1 to match common flame shapes; Step 3.3: Replace RoIPooling with RoIAlign to eliminate spatial quantization error; Step 3.4: Set the multi-task loss function as follows .

6. The method for early infrared flame feature extraction and warning of thermal runaway in new energy vehicles as described in claim 1, characterized in that, The formula for calculating the apparent area of ​​the flame in step 4 is: A j (t)=N pixel ×α Among them, A j (t) represents the projected area of ​​the j-th flame instance at time t, N pixel α is the number of foreground pixels in the mask corresponding to the flame instance, and α is a pre-calibrated conversion factor for the actual pixel area. The surface equivalent temperature is obtained by indexing the temperature values ​​of the corresponding pixels of the mask from the original temperature matrix and calculating the arithmetic mean.

7. The method for early infrared flame feature extraction and warning of thermal runaway in new energy vehicles as described in claim 1, characterized in that, In step 4, when there are multiple flame instances in a single frame image, the apparent areas of each instance are summed, and the temperature is weighted by area to obtain the comprehensive flame features of the frame.

8. The method for early infrared flame feature extraction and warning of thermal runaway in new energy vehicles as described in claim 1, characterized in that, In step 5, the fire source power inversion process uses the simplified radiation expression HRR = k·A·T n Where HRR is the heat release rate obtained by inversion, k is the comprehensive calibration coefficient calibrated by experimental data, A is the apparent area of ​​the flame, T is the surface equivalent temperature in Kelvin K, and n is the temperature index calibrated by experimental data.

9. A new energy vehicle early thermal runaway infrared flame feature extraction and early warning system, using the new energy vehicle early thermal runaway infrared flame feature extraction and early warning method as described in any one of claims 1 to 8, characterized in that, include: The raw data acquisition module is used to read the 16-bit temperature matrix of the infrared thermal imager; The image preprocessing module is used to perform adaptive stretching, scaling, and format conversion; The flame analysis module deploys a pre-trained optimized Mask R-CNN instance segmentation model and outputs a flame mask. The feature calculation module is used to calculate the apparent area of ​​the flame and the surface equivalent temperature. The early warning decision module is used to trigger audible and visual alarms and linkage control based on flame characteristics and fire source power thresholds.

10. The early infrared flame feature extraction and early warning system for thermal runaway in new energy vehicles as described in claim 9, characterized in that, The flame segmentation module adopts an asynchronous pipeline architecture, and the model inference speed is matched with the image acquisition frame rate.

Images