Combustion flame propagation velocity identification method and system based on deep learning

Abstract

The invention discloses a combustion flame propagation velocity recognition method based on deep learning. The method comprises the following steps: step 1, acquiring continuous video frames of combustion flame; 2, performing flame region detection on each frame of image by adopting a YOLOv5s model, and outputting a flame bounding box; 3, inputting the flame bounding box image into a DINOv2 model to extract a feature vector, and judging the validity of the flame image through an MLP classifier; 4, a combustion flame image is extracted from the effective flame image, Canny edge detection is conducted on the combustion flame image, and the flame surface area is calculated in combination with tail flame removal and piecewise function fitting; and step 5, tracking flame front displacement in continuous frames through an ORB feature point matching algorithm, and calculating flame propagation speed in combination with a time interval. The invention further discloses a combustion flame propagation velocity recognition system based on deep learning.

Description

Technical Field

[0001] The present invention belongs to the technical field of flame propagation speed identification, and specifically provides a combustion flame propagation speed identification method and system based on deep learning. Background Art

[0002] Accurate identification of flame propagation velocity is of far-reaching significance for preventing and controlling fires, explosions, and the catastrophic consequences they bring. Therefore, studying the structural characteristics of early flames will help to better prevent and suppress the occurrence of gas explosion disasters. In the development of new clean fuels, such as ammonia-hydrogen mixtures, accurate measurement of flame propagation velocity is an important indicator for evaluating their combustion efficiency and stability. Accurate measurement of flame propagation velocity is crucial for combustion chamber stability design and is an important bridge connecting basic combustion research and engineering applications. In the existing technology, there are mainly the following methods for identifying flame propagation velocity.

[0003] The first approach combines an improved YOLOv7 with a DenseNet121 detection model. YOLOv7 is one of the newest and most advanced models in the YOLO family, boasting high detection speed and accuracy. Its improvements incorporate various mechanisms, such as attention mechanisms (e.g., the SE module and SimAM module) and lightweight architectures (e.g., GhostNet and ConvNeXt), to enhance the model's detection capabilities for small objects and reduce computational overhead. For example, some existing techniques suggest that the SE attention mechanism allows the model to better focus on channel features, thereby improving detection accuracy. DenseNet121 is a convolutional neural network with a densely connected structure that effectively mitigates the vanishing gradient problem and enhances feature reuse. Combining DenseNet121 as a feature extraction network with the improved YOLOv7 improves the model's feature extraction capabilities in complex backgrounds. Despite constructing and expanding a self-built fire smoke dataset, some scenarios are still not included, such as nighttime fire smoke and smoke generated by special occasions. These uncovered scenarios may result in poor performance of the model in real-world applications. Although the model's detection capabilities in complex backgrounds have been improved, in some extremely complex environments, the model may still be interfered with, resulting in false detection or missed detection.

[0004] The second method is flame edge detection and propagation velocity estimation based on the Canny operator. The Canny operator is used to extract the flame edge and, combined with image processing techniques, calculate the flame radius, thereby estimating the flame propagation velocity. For example, in an existing research example, the Canny algorithm was used to detect edges in 16ms and 17ms binary images to obtain the flame edge contour. The 17ms flame edge contour was then obtained by inverting and combining them. Subsequently, the distance between the flame contours of adjacent frames was calculated to derive the flame propagation velocity. However, this method has significant drawbacks. First, the Canny operator is prone to false detections or missed detections when there are numerous flame projections, resulting in significant deviations in the calculated quantitative metrics. Furthermore, the Canny operator requires manual setting of three parameters (high threshold, low threshold, and σ value), is sensitive to noise, and may not accurately capture all key geometric features when processing complex and highly wrinkled flame fronts. Finally, while the Canny operator performs well in most situations, it may underperform other improved algorithms under certain conditions, such as hydrogen-rich flames.

[0005] The third method involves developing a processing program based on the MATLAB platform to automatically analyze gas explosion flame images. They employed a flame brightness threshold method to extract the flame boundary and combined it with an inter-frame difference method to track the displacement of the flame front. This method automatically calculated the flame propagation velocity, significantly improving processing efficiency. While this method offers advantages in improving the efficiency of flame propagation velocity calculation, its stability under complex backgrounds or with non-uniform flame distributions remains to be improved. This method is sensitive to the choice of threshold, which limits its applicability in complex environments.

[0006] The fourth method uses the Laplacian of Gaussian (LoG) operator to detect edges in flame images and fit the extracted edge contours to a function expression to calculate the flame's outer boundary area. Specifically, in one example, the image is first optimized to reduce noise and improve image quality. Next, the LoG operator is used to detect edge information in the image and smoothed to obtain a clearer edge contour. These discrete edge points are then fitted to a continuous function expression, and the error in the fitting curve is corrected using polynomial approximation to obtain a more accurate edge contour. Finally, the fitted function is converted to real coordinates, and the flame's outer boundary area is calculated through area integration. This method relies on image processing optimization steps, such as noise removal and smoothing, which can be affected by image quality in practical applications, resulting in suboptimal edge detection results. Furthermore, while the LoG operator has good robustness, its isotropic nature may not be applicable in certain situations with directional differences, thus affecting edge detection accuracy.

[0007] Therefore, although existing research has made some progress in the imaging measurement of flame propagation velocity, many deficiencies still exist. Therefore, a flame propagation velocity measurement method with high precision, high robustness, and high automation capabilities is urgently needed to make up for the shortcomings of current technology in terms of real-time performance, stability, and adaptability to diverse scenarios. Summary of the Invention

[0008] In view of this, the purpose of the present invention is to provide a combustion flame propagation velocity identification method and system based on deep learning to make up for the shortcomings of current technology in real-time, stability and adaptability to diverse scenarios, and to have the technical advantages of high precision, high robustness and high degree of automation.

[0009] In order to achieve the above object, the present invention provides the following technical solutions: The present invention first proposes a combustion flame propagation velocity identification method based on deep learning, which includes the following steps: Step 1: Obtain continuous video frames of burning flames; Step 2: Use the YOLOv5s model to detect the flame area of ​​each frame image and output the flame bounding box; Step 3: Input the flame bounding box image into the DINOv2 model to extract feature vectors, and use the MLP classifier to determine the validity of the flame image; Step 4: Extract the combustion flame image from the effective flame image, perform Canny edge detection on the combustion flame image, and calculate the flame surface area by combining tail flame removal and piecewise function fitting; Step 5: Track the flame front displacement in consecutive frames using the ORB feature point matching algorithm, and calculate the flame propagation speed based on the time interval.

[0010] Furthermore, in step 2, the target confidence of the YOLOv5s model is expressed as: in: Indicates target confidence; is the Sigmoid function, Pc is the category score output by the network; The loss function of the YOLOv5s model is: in: The CIoU loss of the bounding box is used to measure the distance, overlap and aspect ratio between the predicted box and the true box; Represents the target confidence loss, using the Binary Cross Entropy loss function; Represents the category prediction loss, and performs multi-classification cross entropy calculation on the category predicted by each anchor box; 、 and are the weights of each loss item respectively.

[0011] Furthermore, the MLP classifier includes three layers of fully connected neural networks: The first layer of fully connected neural network maps 768-dimensional features to 512 dimensions, normalized by LayerNorm and activated by ReLU; The second layer of fully connected neural network maps the 512-dimensional features to 256 dimensions, normalized by LayerNorm and activated by ReLU; The third layer of fully connected neural network maps the 256-dimensional features into binary classification probability outputs.

[0012] Furthermore, in step 4, the method for extracting the combustion flame image is: The global brightness calculation method is used to capture the dynamic characteristics of the flame and obtain the burning flame image; The automatic anchoring target algorithm is used to select the first n burning flame images with the highest brightness, and the selected burning flame images are processed including removing noise, small holes and burrs.

[0013] Furthermore, in step 4, removing the tail flame includes: The flame area is identified using the 8-connected region labeling algorithm; The connected area with the largest number of pixels is retained as the valid flame area.

[0014] Furthermore, in step 4, the piecewise function fitting includes: Convert the discrete points of the flame edge into a scatter plot, retaining the point with the largest y value under the same x coordinate and the first and last data; Use spline interpolation to interpolate scattered points; The flame profile curve is fitted using a piecewise polynomial function.

[0015] Furthermore, in step 5, the ORB feature point matching includes: Perform pseudo color enhancement on two flame images; Extract rotation-invariant rBRIEF descriptor; Filter mismatched point pairs through brute force matching and y-direction difference; The first ten groups of matching points with the smallest distance are selected to calculate the average displacement.

[0016] Furthermore, in step 5, the flame propagation speed is: in: is the actual distance of the feature point displacement converted by the image scale; is the time interval between two frames of flame images; The flame burning speed is: in: is the flame burning speed; is the flame propagation speed; is the cross-sectional area of ​​the flame base; is the flame surface area.

[0017] The present invention also proposes a combustion flame propagation speed identification system based on deep learning, comprising: The data acquisition layer is used to input the burning flame video collected by high-speed camera equipment and perform frame processing on the burning flame video; The intelligent analysis layer includes a processing module and a computing module. The processing module integrates a YOLOv5s target detection unit, a DINOv2 feature extraction unit, and an MLP classification unit. The computing module performs Canny edge detection, ORB feature point matching, and speed calculation algorithms. The visual interaction layer includes a visual interface for real-time display of flame enhancement images, edge images, feature matching images and speed parameters.

[0018] Furthermore, the visual interface includes: Four-image synchronous display area, showing flame enhancement image, binarization image, edge detection image and feature point matching image in parallel; Parameter output area, real-time display of flame propagation speed, combustion speed, flame surface area and bottom cross-sectional area; The data fitting area provides a quadratic polynomial fitting curve of equivalence ratio and combustion speed.

[0019] The beneficial effects of the present invention are: The combustion flame propagation velocity identification method based on deep learning of the present invention has the following technical effects: (1) Compared with the false detection problem of traditional methods in complex backgrounds, this paper combines a lightweight object detection model (YOLOv5s model) and a self-supervised visual Transformer model (DINOv2-Small) to achieve rapid positioning and high-precision classification of flame areas; YOLOv5s is responsible for real-time detection of flame positions, and DINOv2 improves fine-grained classification capabilities through deep semantic feature extraction; (2) Compared with the traditional threshold segmentation method, this paper designs a multi-layer perceptron (MLP) classification head based on the particularity of flame images, mapping the 768-dimensional features output by DINOv2 into a binary classification task (valid / invalid flame), effectively improving the recognition accuracy; (3) Compared with the error accumulation problem of the traditional manual calibration method, the present invention uses the Canny operator to extract the flame edge, combines piecewise function fitting (such as polynomial or spline curve) to calculate the flame surface area, and uses the ORB (OrientedFAST and Rotated BRIEF) algorithm to extract the flame feature points in continuous frames, calculates the ratio of the displacement difference to the time interval, and realizes dynamic tracking of the propagation speed. The robustness is stronger than the traditional optical flow method.

[0020] (4) Compared with the scarcity of professional data in the combustion field, this paper reduces the dependence on labeled data through cross-domain integration and utilizes the self-supervised pre-training capability of DINOv2. It also integrates technologies from multiple fields such as combustion science (ISO 00817-2014 standard), computer vision (YOLO / ViT), and image processing (ORB / Canny), providing a model for interdisciplinary research.

[0021] In summary, the combustion flame propagation speed identification method based on deep learning in the present invention can meet the requirements of combustion flame propagation speed identification for real-time, stability and adaptability to diverse scenarios, and has the technical advantages of high precision, high robustness and high degree of automation. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to make the purpose, technical solutions and beneficial effects of the present invention more clear, the present invention provides the following drawings for illustration: Figure 1 This is a flow chart of the combustion flame propagation velocity identification method based on deep learning of the present invention; Figure 2 This is the structural diagram of the YOLOv5s network; Figure 3 Schematic diagram of the global brightness calculation method for capturing the dynamic characteristics of flames; Figure 4 This is the tail flame treatment effect diagram; Figure 5 This is the effect of using the expansion algorithm to remove burrs; Figure 6 Schematic diagram for obtaining computable edge information; Figure 7 Fitting marginal plots to functions; Figure 8 It is the effect diagram of segmented fitting; Figure 9 Calculate the visualization graph for the standard; Figure 10 This is a schematic diagram of the flame surface area calculation principle; Figure 11 Schematic diagram of the propagation of combustion flame in a vertical cylindrical tube; Figure 12This is a framework diagram of the combustion flame propagation speed identification system based on deep learning of the present invention; Figure 13 It is a visual interface; Figure 14 It is a visual interface during the analysis process; Figure 15 It is the parameter output area; Figure 16 is the data fitting area; Figure 17-18 For the experimental device; Figure 19-20 The collected flame image. DETAILED DESCRIPTION

[0023] The present invention will be further described below with reference to the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it. However, the embodiments are not intended to limit the present invention.

[0024] like Figure 1 As shown, the combustion flame propagation speed identification method based on deep learning in this embodiment includes the following steps.

[0025] Step 1: Obtain continuous video frames of burning flames.

[0026] Specifically, this embodiment uses a high-speed camera to capture the dynamic propagation process of the combustion flame and obtain continuous video frames of the combustion flame.

[0027] Step 2: Use the YOLOv5s model to detect the flame area of ​​each frame image and output the flame bounding box.

[0028] (1) YOLOv5s model principle The YOLOv5s network series covers multiple versions such as YOLOv5s, YOLOv5m, YOLOv5l and YOLOv5x. Specifically, Figure 2 As shown in the figure, the Backbone network is the core feature extraction module of YOLOv5s. It uses CSPDarknet53 as the backbone network. Its main components include the Focus layer, CSP (Cross Stage Partial) structure, and SiLU (Swish) activation function. The Focus layer converts the spatial dimensions of the input image into channel dimensions through a slicing operation, reducing computation and improving inference efficiency. The CSP structure divides the feature map into two parts, one of which is convolved and then cross-stage fused with the other. The smooth nonlinear characteristics of the SiLU activation function help improve the stability of gradient propagation, thereby enhancing model training results.

[0029] Specifically, YOLOv5s uses an anchor-based mechanism to predict the target category, bounding box coordinates, and target confidence at each grid point. Specifically, the target confidence of the YOLOv5s model is expressed as: in: Indicates target confidence; is the Sigmoid function, and Pc is the category score output by the network.

[0030] Through this mechanism, YOLOv5s can efficiently complete target detection tasks while maintaining high detection accuracy and real-time performance.

[0031] (2) YOLOv5s model training settings and model evaluation indicators Model training is based on the PyTorch deep learning framework and is implemented using the official open-source training code for YOLOv5. The training data is the aforementioned custom-annotated VOC-format flame image dataset, which is converted to the YOLO format and then fed into the model. The SGD (Stochastic Gradient Descent) optimizer is used during training, with an initial learning rate of 0.01, a batch size of 16, and a total of 300 training rounds. The loss function of the YOLOv5 model consists of three parts, the weighted sum of which constitutes the final total loss function. Specifically, the loss function of the YOLOv5s model is: in: The CIoU loss of the bounding box is used to measure the distance, overlap and aspect ratio between the predicted box and the true box; Represents the target confidence loss, using the Binary Cross Entropy loss function; Represents the category prediction loss, and performs multi-classification cross entropy calculation on the category predicted by each anchor box; 、 and are the weights of each loss item. In this embodiment, they are [0.05, 1.0, 0.5] respectively.

[0032] This example selects the following target detection evaluation indicators to evaluate the performance of the YOLOv5 model.

[0033] (1) Precision: This refers to the proportion of samples predicted by the model to be positive that are actually positive. (2) Recall rate: It indicates the proportion of all true positive samples that are successfully detected. (3) Average precision (AP): Integrate the Precision-Recall curve under different confidence thresholds to obtain the average precision of a single class of targets.

[0034] (4) mAP (mean Average Precision) is the average value of AP under multiple target categories, which measures the overall detection performance. This example uses an IoU threshold of 0.5: Where TP is the number of correctly detected positive samples, FP is the number of falsely detected negative samples, FN is the number of missed positive samples, and N is the total number of target categories. Since the flame detection task in this example is a single target (flame), the mAP metric can be directly used for overall performance evaluation.

[0035] Once the YOLOv5 model successfully frames a flame, it can not only identify its specific location within the image but also calculate its propagation speed by combining the positional information between frames with the time difference. Furthermore, these framed image regions can be used as input for subsequent classification tasks to identify the specific flame type.

[0036] Step 3: Input the flame bounding box image into the DINOv2 model to extract the feature vector, and use the MLP classifier to determine the validity of the flame image.

[0037] This embodiment designs an MLP classification head consisting of a three-layer fully connected neural network. This classification head can perform image validity judgment on the 768-dimensional image feature vector output by DINOv2. In other words, in this embodiment, the MLP classifier includes a three-layer fully connected neural network. The specific principle is as follows.

[0038] The first layer of the fully connected neural network maps the 768-dimensional features to 512 dimensions, normalized by LayerNorm and activated by ReLU, specifically including: Receive the output feature vector of Transformer Perform the first layer of linear transformation: in: is the output vector of the first layer linear transformation; 768-dimensional feature vector output by the DINOv2 model; is the first layer weight matrix (512 rows × 768 columns); is the first layer bias vector.

[0039] Perform Layer Normalization (parameters are , ): in: is the normalized eigenvector; for The mean of for The standard deviation of is a learnable scaling parameter; is a learnable offset parameter.

[0040] ReLU function activation: in: is the output after activation of the first layer.

[0041] The second layer of fully connected neural network maps the 512-dimensional features to 256 dimensions, normalized by LayerNorm and activated by ReLU, specifically including: The first layer output Input to the second layer: in: is the second layer linear output; is the second layer weight matrix; is the second layer bias vector.

[0042] Perform Layer Normalization (parameters are , ): in: is the output vector after layer normalization; is the mean of the input data; is the standard deviation of the input data; is the scaling parameter in layer normalization; is the offset parameter in layer normalization.

[0043] The third layer of the fully connected neural network maps the 256-dimensional features into binary classification probability outputs, specifically including: Will Mapped into two categories of logits: in: is the binary logits vector, the output vector of the third fully connected layer; is the weight matrix of the third fully connected layer; is the bias vector of the third fully connected layer.

[0044] Use the Softmax function to normalize the probabilities: in: are probability vectors, representing the probabilities of the image being a “valid” and “invalid” flame image respectively; is the sum of the exponentials of two logit values, used to normalize the logits into a probability distribution.

[0045] The final category is determined by the item with the highest probability: if , the image is judged as a valid image and enters the subsequent flame edge extraction and combustion speed calculation process; otherwise it is discarded.

[0046] In the model application process, YOLOv5 is first used to detect the flame region, outputting a flame frame as the input image to the DINOv2 model. DINOv2 extracts the image's deep features, then passes them to an MLP classifier for classification and identification, determining whether the image represents a valid flame. This example utilizes a large amount of flame image data annotated with Labelme, and uses supervised learning to fine-tune the parameters of the final few layers of DINOv2 to adapt them to the characteristic distribution of flame images.

[0047] Step 4: Extract the combustion flame image from the effective flame image, perform Canny edge detection on the combustion flame image, and calculate the flame surface area by combining tail flame removal and piecewise function fitting.

[0048] (1) Extracting combustion flame images The method for extracting combustion flame images is as follows: a global brightness calculation method is used to capture the dynamic characteristics of the flame to obtain a combustion flame image; an automatic anchoring target algorithm is used to select the first N combustion flame images with the highest brightness, and the selected combustion flame images are processed including removing noise, small voids and burrs.

[0049] Specifically, after obtaining the image from the video frame, it is also necessary to extract the flame image from the image. The global brightness calculation method can well capture the dynamic characteristics of the flame. Specifically, first calculate the overall brightness of the first frame in the video, and then select the next frame image for extraction based on this brightness value. This process can make the extracted image have more prominent flame brightness characteristics, which is convenient for subsequent processing. In addition, the brightness space mapping technology is used to map each frame image to the color space, thereby enhancing the color channel in the image. This process is as follows Figure 3 As shown in the figure, there are the original image, the image with enhanced color space, and the enhanced green channel image.

[0050] After extracting the flame image, the automatic anchoring algorithm sets an all-ones matrix of size (n,m). This matrix is ​​then treated as a mask, first horizontally and then vertically translated by one pixel. The mask is multiplied by the anchor region and then summed to obtain the total brightness. The first n (default = 600) brightest images are selected to remove background noise and improve the accuracy of flame region extraction. For the first image obtained, the object in this image is selected near the center, and a Gaussian weighted kernel (the currently used ellipsoidal Gaussian weighted kernel) is applied to obtain an image relatively close to the center.

[0051] In this embodiment, the opening and closing operations of digital image processing are used to remove small holes and burrs, and then the clustering algorithm DBSCAN algorithm is used to remove irrelevant noise.

[0052] (2) Tail flame removal Tail flame removal includes: using 8-connected region labeling algorithm to identify the flame area; retaining the connected region with the largest number of pixels as the valid flame area.

[0053] Specifically, when processing the flame front and tail flame data, there is redundant tail flame data in the flame image that needs to be processed, such as Figure 4 As shown in Figure 2, we use the 8-connected labeling algorithm to mark each part of the image, retaining the largest number of connected regions as the valid parts for area calculation. This method can effectively remove redundant tail flame data and improve calculation accuracy.

[0054] (3) Piecewise function fitting The piecewise function fitting includes: converting the discrete points of the flame edge into a scatter plot, retaining the point with the largest y value under the same x coordinate and the first and last data; using the spline interpolation method to interpolate the scattered points; and using the piecewise polynomial function to fit the flame contour curve.

[0055] Specifically, before using the Canny operator to extract edge features, it is necessary to use the dilation algorithm to remove burrs. The effect after processing is as follows: Figure 5 As shown. At the same time, the expansion algorithm will further correct the previous error. Figure 6 As shown, after obtaining the image of the edge that can be calculated, the upper boundary is obtained and the optimal segmentation point is obtained through the algorithm, and the flame image is divided into the front segment and the back segment for calculation respectively.

[0056] Convert the image into a scatter plot for area calculation. For data with the same x, select the point with the largest y, and force the first and last data to be saved. Figure 7As shown in the figure, a scatter plot with a small theoretical error is obtained, and then the spline interpolation method is used to interpolate the scattered points of each segment to further improve the accuracy of the fitting. Finally, different polynomial functions are used to fit different parts of the segment. The segmented fitting diagram and the final function diagram are shown in the figure. Figure 8 and Figure 9 Finally, as Figure 10 As shown, the flame surface area can be calculated by substituting the following formula: in: is the overall surface area of ​​the flame; is the surface area of ​​a certain segment of the flame; and Respectively in and Flame diameter at (unit: mm); and are the horizontal coordinate differences, i.e. d = ; is the longitudinal coordinate difference, that is, the vertical height difference of the flame in the segment.

[0057] Step 5: Track the flame front displacement in consecutive frames using the ORB feature point matching algorithm, and calculate the flame propagation speed based on the time interval.

[0058] (1) ORB feature point matching In this embodiment, ORB feature point matching includes: performing pseudo-color enhancement on two flame image frames; extracting the rotation-invariant rBRIEF descriptor; filtering mismatched point pairs through brute force matching and y-direction difference; and selecting the top ten matching points with the smallest distance to calculate the average displacement.

[0059] Specifically, ORB uses rotation-invariant BRIEF (rBRIEF) to improve the rotation robustness of the descriptor. The specific calculation steps are as follows: (1) Image block sampling For each detected keypoint, a fixed-size image patch (31×31 pixels) is extracted centered at the keypoint. The BRIEF descriptor is calculated within this image patch.

[0060] (2) Binary comparison BRIEF generates a binary bit by randomly selecting a pair of pixels within the extracted image block and comparing their grayscale values. If the grayscale value of the first pixel is less than the grayscale value of the second pixel, the bit is 1, otherwise it is 0. By doing this multiple times (128 or 256 times), a binary string can be generated.

[0061] (3) Descriptor generation The results of all binary comparisons (0 or 1) form a binary string, which is the BRIEF descriptor for the keypoint. For example, if 128 comparisons are performed, the resulting descriptor is a 128-bit binary string. ORB optimizes the rotational invariance of the BRIEF descriptor by rotating the image patch and calculating the optimal matching direction, making feature matching more stable.

[0062] (2) Flame burning speed Flame burning speed test method Figure 11 The specific calculation is as follows: in: is the flame burning speed; is the flame propagation speed; is the cross-sectional area of ​​the flame base; is the flame surface area.

[0063] Specifically, this embodiment uses the ORB (Oriented FAST and Rotated BRIEF) feature detection and description algorithm to detect and match key points of two frames extracted from the experimental video (i.e., the start frame and the end frame), thereby estimating the spatial displacement of the flame front and calculating the propagation speed based on the time information.

[0064] First, two frames containing significant movement of the flame front were selected from the original video. The flame regions were cropped from each frame and pseudo-color enhancement was performed on the cropped images to improve the texture contrast in the key areas. The ORB algorithm was then used to extract feature points from the two frames and calculate their corresponding descriptors.

[0065] After feature extraction, the system pairs the descriptors using a brute-force matcher. It then filters out clearly unreasonable matching pairs using the y-direction difference to eliminate false matches. The matching results are then sorted in ascending order by distance, and the top ten feature point pairs are selected as valid references.

[0066] To calculate the spatial displacement of the flame front, the coordinates of the matching points in the cropped image are mapped back to the original image coordinate system, and the true position of the matching points in the full image is obtained using the following formula: in: are the key point coordinates in the cropped image, The coordinates of the upper left corner of the cropping box in the original image.

[0067] Then, the average displacement of all matching point pairs in the x direction is calculated: in: and For the The position of the matching point in the two frames of image; is the number of matching points.

[0068] The average displacement value is converted into the propagation distance by combining the set image scale (the actual length corresponding to the unit pixel) : in: is the image scale.

[0069] Finally, the flame propagation speed for: in: is the actual distance of the feature point displacement converted by the image scale; is the time interval between two flame images.

[0070] like Figure 12 As shown, this embodiment also proposes a combustion flame propagation speed identification system based on deep learning, including: The data acquisition layer is used to input the burning flame video collected by high-speed camera equipment and perform frame processing on the burning flame video; The intelligent analysis layer includes a processing module and a computing module. The processing module integrates a YOLOv5s target detection unit, a DINOv2 feature extraction unit, and an MLP classification unit. The computing module performs Canny edge detection, ORB feature point matching, and speed calculation algorithms. The visual interaction layer includes a visual interface for real-time display of flame enhancement images, edge images, feature matching images and speed parameters.

[0071] like Figure 13-14 As shown in the figure, the visualization interface includes: a four-image synchronous display area, which displays the flame enhancement image, binarization image, edge detection image and feature point matching image in parallel; a parameter output area, which displays the flame propagation speed, combustion speed, flame surface area and bottom cross-sectional area in real time; and a data fitting area, which provides a quadratic polynomial fitting curve of the equivalence ratio and combustion speed.

[0072] In this example, a visualization interface built on the Qt framework, combined with OpenCV and PyTorch backends, was used to process, identify, analyze, and predict combustion flame images. The interface design was completed using Qt Designer, the official Qt visualization tool. Qt Designer adheres to the MVC architectural concept, effectively decoupling view and logic. Interface controls are laid out using drag-and-drop. After design is complete, a .ui file is generated and automatically converted to a .py file using the pyuic5 tool. This file, combined with the Python main program, implements the system's functionality.

[0073] The visual interface is divided into a four-image simultaneous display area, a parameter output area, and a data fitting area. Each area is logically organized for intuitive operation and rapid access to test and calculation results. The system supports key tasks such as experimental video loading, frame image processing, image enhancement, flame detection, edge fitting, feature extraction, velocity calculation, and function prediction, and displays processing results in real time.

[0074] (1) Video stream input and frame image analysis: Users can import experimental videos through the button in the upper left corner of the interface. The system uses OpenCV to implement frame-by-frame video parsing and caching. Each frame of the image will be processed in the following order: Flame area enhancement: Improves image contrast and edge features to make the flame area clearer; Image binarization processing: extract the bright area of ​​the flame for contour extraction; Flame edge detection: Use the Canny algorithm to extract edges for easy fitting calculation; Feature point extraction and matching: ORB algorithm is used to track the movement of flame boundary for propagation speed calculation.

[0075] like Figure 13 As shown, the processed image is displayed synchronously in the four-image synchronous display area as four images, including: flame enhancement image, binary image, edge image and feature point matching image, which can help users intuitively observe the flame structure and propagation trend.

[0076] (2) Parameter calculation and speed identification: like Figure 15 As shown in the figure, after the image processing is completed, the system automatically calculates multiple key parameters and displays them in real time in the parameter output area on the right side of the interface, including: flame propagation speed (mm / s), flame bottom cross-sectional area (mm²), flame surface area (mm²), flame combustion speed (mm / s) and -fitted combustion speed (mm / s).

[0077] Combustion velocity calculations are based on a spherical flame propagation model, combining flame area and front displacement for high-precision analysis. Users can freely adjust parameters and repeat calculations, and the system supports automatic result refresh.

[0078] (3) Experimental data fitting and prediction function: In order to realize the flame propagation law modeling, the system provides a quadratic polynomial fitting function. The user inputs the equivalent ratio under different experimental conditions. Corresponding combustion speed , the system completes the fitting through the least squares method, and the function form is as follows: like Figure 16 As shown, after the fitting is completed, the system automatically draws the fitting curve and displays the fitting expression in real time in the data fitting area. The user can also enter any equivalent ratio The system predicts the combustion rate and fitting residuals to help judge the accuracy of the experimental law and the reliability of the fitting model.

[0079] The above embodiments are merely preferred embodiments for the purpose of fully illustrating the present invention, and the scope of protection of the present invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on the present invention are within the scope of protection of the present invention. The scope of protection of the present invention shall be subject to the claims.

Claims

1. A method for identifying combustion flame propagation velocity based on deep learning, characterized by: The steps include: Step 1: Obtain continuous video frames of burning flames; Step 2: Use the YOLOv5s model to detect the flame area of ​​each frame image and output the flame bounding box; Step 3: Input the flame bounding box image into the DINOv2 model to extract feature vectors, and use the MLP classifier to determine the validity of the flame image; Step 4: Extract the combustion flame image from the effective flame image, perform Canny edge detection on the combustion flame image, and calculate the flame surface area by combining tail flame removal and piecewise function fitting; Step 5: Track the flame front displacement in consecutive frames using the ORB feature point matching algorithm, and calculate the flame propagation speed based on the time interval.

2. The method for identifying combustion flame propagation velocity based on deep learning according to claim 1, characterized in that: In step 2, the target confidence of the YOLOv5s model is expressed as: in: Indicates target confidence; is the Sigmoid function, Pc is the category score output by the network; The loss function of the YOLOv5s model is: in: The CIoU loss of the bounding box is used to measure the distance, overlap and aspect ratio between the predicted box and the true box; Represents the target confidence loss, using the Binary Cross Entropy loss function; Represents the category prediction loss, and performs multi-classification cross entropy calculation on the category predicted by each anchor box; 、 and are the weights of each loss item respectively.

3. The method for identifying combustion flame propagation velocity based on deep learning according to claim 1, characterized in that: The MLP classifier includes three layers of fully connected neural networks: The first layer of fully connected neural network maps 768-dimensional features to 512 dimensions, normalized by LayerNorm and activated by ReLU; The second layer of fully connected neural network maps the 512-dimensional features to 256 dimensions, normalized by LayerNorm and activated by ReLU; The third layer of fully connected neural network maps the 256-dimensional features into binary classification probability outputs.

4. The method for identifying combustion flame propagation velocity based on deep learning according to claim 1, characterized in that: In the fourth step, the method for extracting the combustion flame image is: The global brightness calculation method is used to capture the dynamic characteristics of the flame and obtain the burning flame image; The automatic anchoring target algorithm is used to select the first n burning flame images with the highest brightness, and the selected burning flame images are processed including removing noise, small holes and burrs.

5. The method for identifying combustion flame propagation velocity based on deep learning according to claim 1, characterized in that: In the step 4, the tail flame removal includes: The flame area is identified using the 8-connected region labeling algorithm; The connected area with the largest number of pixels is retained as the valid flame area.

6. The method for identifying combustion flame propagation velocity based on deep learning according to claim 1, characterized in that: In the step 4, the piecewise function fitting includes: Convert the discrete points of the flame edge into a scatter plot, retaining the point with the largest y value under the same x coordinate and the first and last data; Use spline interpolation to interpolate scattered points; The flame profile curve is fitted using a piecewise polynomial function.

7. The method for identifying combustion flame propagation velocity based on deep learning according to claim 1, characterized in that: In the step 5, the ORB feature point matching includes: Perform pseudo color enhancement on two flame images; Extract rotation-invariant rBRIEF descriptor; Filter mismatched point pairs through brute force matching and y-direction difference; The first ten groups of matching points with the smallest distance are selected to calculate the average displacement.

8. The method for identifying combustion flame propagation speed based on deep learning according to claim 7, characterized in that: In step 5, the flame propagation speed is: in: is the actual distance of the feature point displacement converted by the image scale; is the time interval between two frames of flame images; The flame burning speed is: in: is the flame burning speed; is the flame propagation speed; is the cross-sectional area of ​​the flame base; is the flame surface area.

9. A combustion flame propagation velocity identification system based on deep learning, characterized by: include: The data acquisition layer is used to input the burning flame video collected by high-speed camera equipment and perform frame processing on the burning flame video; The intelligent analysis layer includes a processing module and a computing module. The processing module integrates a YOLOv5s target detection unit, a DINOv2 feature extraction unit, and an MLP classification unit. The computing module performs Canny edge detection, ORB feature point matching, and speed calculation algorithms. The visual interaction layer includes a visual interface for real-time display of flame enhancement images, edge images, feature matching images and speed parameters.

10. The combustion flame propagation velocity identification system based on deep learning according to claim 9, characterized in that: The visual interface includes: Four-image synchronous display area, showing flame enhancement image, binarization image, edge detection image and feature point matching image in parallel; Parameter output area, real-time display of flame propagation speed, combustion speed, flame surface area and bottom cross-sectional area; The data fitting area provides a quadratic polynomial fitting curve of equivalence ratio and combustion speed.

Images