A method for detecting a combustion plume of a liquid rocket engine

Abstract

The application provides a liquid rocket engine combustion tail flame detection method, and belongs to the technical field of tail flame detection. The method first builds a collaborative detection environment for fusion detection and tracking, and integrates bidirectional interaction, dynamic weight distribution, lightweight adaptation and multi-modal noise suppression modules. Then, a full-scene test video is collected to construct a tail flame dataset. A lightweight detection model is obtained through lightweight modification, and is jointly trained with a SiamMask detection model to obtain an integrated detection model. During detection, the starting frame and initial target box are determined through a detection branch, and the tracking template is initialized. Then, tail flame detection and tracking are performed synchronously. The detection and tracking results of subsequent frames are fused through a dynamic weight distribution module, the tail flame area is denoised, key parameters are extracted, and a detection report is generated. The application improves the detection stability under complex working conditions through bidirectional interaction and dynamic weight distribution mechanism, ensures the recognition accuracy in harsh environments, and reduces the false detection and missed detection rate.

Description

Technical Field

[0001] This invention relates to the field of exhaust flame detection technology, and in particular to a method for detecting the exhaust flame of a liquid rocket engine combustion. Background Technology

[0002] The detection and identification of exhaust plumes from liquid rocket engines is essentially an application of image-based flame target detection. With the development of digital image processing and computer vision technologies, related exhaust plume detection technologies have received widespread attention.

[0003] Existing flame detection methods based on manual feature extraction suffer from problems such as complex modeling, low detection accuracy, high false detection rate, and poor transferability, making them difficult to adapt to the detection needs of liquid rocket engine exhaust plumes with different thrust models and shooting angles. To address the shortcomings of manual feature extraction-based flame detection methods, current approaches typically use a target detection model as the detection backbone, combined with a corresponding target tracking model, and automatically learn the essential features of the flame using a data-driven approach. This method improves detection accuracy, stability, and transferability to some extent, solving some of the shortcomings of traditional manual feature extraction methods. However, it lacks bidirectional interaction and collaborative optimization between the target detection model and the target tracking model. The tracking process relies on the initial detection results of the target detection model; if the initial frame fails to detect due to weak exhaust plumes, occlusion, or background interference, subsequent tracking will directly fail. Furthermore, no dynamic fusion strategy for detection and tracking results is designed. When the exhaust plume morphology changes drastically due to operating conditions, the detection and tracking results are prone to conflict, leading to decreased detection stability and unsatisfactory detection results for engine exhaust plumes. Therefore, further improvements to the exhaust plume detection method are needed. Summary of the Invention

[0004] In view of this, the present invention provides a method for detecting the combustion exhaust plume of a liquid rocket engine, in order to improve the stability of exhaust plume detection under complex operating conditions.

[0005] The technical solution adopted in this invention is: This invention provides a method for detecting the combustion exhaust plume of a liquid rocket engine, comprising: We construct a collaborative operating environment for the YOLOv5s detection model and the SiamMask tracking model, and integrate a two-way interaction module for detection and tracking, a dynamic weight allocation module, a lightweight model adaptation component, and a multimodal noise suppression component. Test videos of liquid rocket engines under various operating conditions were collected, and images were extracted and the exhaust plume region was labeled to construct a full-scene exhaust plume dataset; A lightweight detection model is obtained by modifying the YOLOv5s basic model based on a lightweight model adaptation component. Using a full-scene exhaust plume dataset, the lightweight detection model and the SiamMask tracking model are jointly trained by calling the detection and tracking bidirectional interaction module to obtain an integrated detection model. The integrated detection model includes a detection branch and a tracking branch. The test video is preprocessed using a multimodal noise suppression component, and the preprocessed image is input into the detection branch of the integrated detection model for initial frame processing to determine the starting frame and initial target box position of the tail flame and initialize the tracking matching template. Based on the starting frame and tracking matching template of the exhaust flame, the preprocessed image is detected and tracked synchronously by the detection branch and the tracking branch. The detection results and tracking results of subsequent frames are dynamically fused through the dynamic weight allocation module to output the fused exhaust flame target box. The exhaust flame region of the fused exhaust flame target box is denoised and key parameters are extracted to generate a detection report.

[0006] Furthermore, the construction of a collaborative operating environment for the YOLOv5s detection model and the SiamMask tracking model, integrating a bidirectional interaction module for detection and tracking, a dynamic weight allocation module, a lightweight model adaptation component, and a multimodal noise suppression component, includes: Set up a collaborative runtime environment for the YOLOv5s detection model and the SiamMask tracking model, and install the dependency packages required for model operation; By sharing the feature extraction layer of the backbone network, a bidirectional feature transmission channel is established between YOLOv5s and SiamMask, so that the spatial features of the exhaust flame extracted by YOLOv5s can be transmitted to SiamMask in real time, while the temporal motion features of the exhaust flame extracted by SiamMask are fed back to YOLOv5s, thus obtaining a bidirectional interactive module for detection and tracking. The logic for judging the change of tail flame shape is set based on the overlap of the target boxes of the tail flame in adjacent frames and the rate of change of tail flame area. A dynamic weight allocation module is constructed based on the logic for judging the change of tail flame shape, and a lightweight model adaptation component is constructed based on the logic for lightweight transformation of the model. Based on different types of image interference, we select appropriate halo suppression algorithms, smoke occlusion penetration algorithms, and strong light reflection filtering algorithms to integrate and construct a multimodal noise suppression component.

[0007] Furthermore, the process involves collecting test videos of liquid rocket engines under various operating conditions, extracting images, labeling the exhaust plume region, and constructing a full-scene exhaust plume dataset, including: Liquid rocket engines with different thrust models were selected and tested from various shooting angles to record steady-state exhaust flame videos under rated operating conditions. Simulate engine start-up, shutdown, and combustion instability fault scenarios, and record unsteady exhaust flame videos; create high-altitude, low-temperature, low-pressure, rain, snow, fog, dust, strong light reflection, exhaust flame halo, and smoke obscuring environments using environmental simulation equipment, and record exhaust flame test videos for the corresponding environments. All recorded exhaust flame experimental videos were processed by frame extraction to extract image frames that meet the preset clarity requirements, and the image frames were divided into training set and validation set according to the preset ratio. Using the annotation tool, establish a main file structure that includes an image storage folder and a label storage folder, and store the images to be annotated in the training set and validation set into the image storage folder; Launch the annotation tool, open the image to be annotated in the image storage folder, and annotate the exhaust flame region, object type, and name in the image; after annotation, generate the corresponding label file and save it to the label storage folder; Create a dataset configuration file and store the file storage paths for the training and validation sets, as well as the types and names of the objects in the dataset configuration file; By combining the image storage folder, the label storage folder, and the dataset configuration file, a full-scene exhaust flame dataset is obtained.

[0008] Furthermore, the lightweight detection model is obtained by modifying the YOLOv5s basic model based on the lightweight model adaptation component. The lightweight detection model and the SiamMask tracking model are then jointly trained using a full-scene exhaust plume dataset to obtain an integrated detection model, including: YOLOv5s is used as the base model. The architecture of the base model includes an input layer, a focus layer, a CSPDarknet53 backbone layer, an FPN+PAN neck layer, and a detection head output layer. A lightweight detection model is obtained by using a lightweight model adaptation component to lightweightly modify the input end, focus layer, FPN+PAN neck layer, and detection head output layer of the basic model. The lightweight detection model includes a lightweight input end, a lightweight focus layer, a CSPDarknet53 backbone layer, a lightweight FPN+PAN neck layer, and a lightweight detection head layer, with each layer connected sequentially. The lightweight modification includes channel pruning, quantization compression, and network layer merging. The full-scene exhaust plume dataset is input into the lightweight detection model for model training. During the training process, the detection performance of the model is evaluated periodically using the validation set. The hyperparameters of the lightweight model are adjusted based on the detection performance evaluation results to obtain the basic detection model with the best ability to recognize exhaust plume features. The detection and tracking bidirectional interaction module connects the core functional layers of SiamMask (feature extraction layer, cross-correlation layer, mask prediction layer, and target box regression layer) with the basic detection model. The module achieves the connection between the two by sharing the backbone feature extraction layer, building a bidirectional feature transmission channel, and having a built-in feature dimension transformation sub-layer. During joint training, the multi-task loss values ​​are backpropagated bidirectionally to each layer to achieve collaborative optimization. The basic detection model is used as the detection branch, and the SiamMask target tracking model is used as the tracking branch to build an integrated fusion model framework. A multi-task loss function, including detection loss and tracking loss, is adopted. The integrated fusion model is jointly trained using the training set and the validation set. The multi-task loss values ​​during joint training are backpropagated bidirectionally to each layer for collaborative optimization. After training, the integrated detection model is obtained.

[0009] Further, the preprocessing of the test video using a multimodal noise suppression component, and the input of the preprocessed image into the detection branch of the integrated detection model for initial frame processing to determine the starting frame and initial target box position of the exhaust flame and initialize the tracking matching template, includes: All recorded exhaust flame experimental videos are parsed into image sequences frame by frame. Each frame of the image sequence is fed into a multimodal noise suppression component, and image interference is removed sequentially through a halo suppression algorithm, a smoke occlusion penetration algorithm, and a strong light reflection filtering algorithm. The preprocessed image is then output. The preprocessed image is input into the detection branch of the integrated detection model. The detection branch identifies the tail flame and determines the starting frame of the tail flame combustion, while locating the initial target box position of the tail flame. The initial target box position is then passed to the tracking branch of the integrated model to complete the initialization of the tracking matching template.

[0010] Furthermore, the starting frame and tracking matching template based on the exhaust flame are used to simultaneously detect and track the preprocessed image using detection and tracking branches. The detection and tracking results of subsequent frames are dynamically fused through a dynamic weight allocation module to output a fused exhaust flame target bounding box, including: The detection branch detects the exhaust flame based on the starting frame of the exhaust flame and performs exhaust flame detection on subsequent image frames in the preprocessed image. Meanwhile, the tracking branch tracks the exhaust flame based on the tracking matching template and performs exhaust flame tracking on subsequent image frames in the preprocessed image, respectively obtaining the exhaust flame detection result and the tracking result. The dynamic weight allocation module is invoked to determine the degree of change in the tail flame morphology based on the overlap of the tail flame target boxes in adjacent frames and the rate of change of the tail flame area. Weights are then allocated to the detection and tracking results based on this degree of change. A weighted fusion method is used to sum the coordinate parameters of the detected and tracked target boxes to obtain the fused tail flame target box for the current image frame. The dynamic weight allocation module dynamically adjusts the interaction weights of the detection and tracking features based on the degree of change in the tail flame morphology, enabling dynamic collaboration between detection and tracking.

[0011] Furthermore, the method also includes: calculating the center distance and area deviation between the tracking target box of the tracking branch and the detection target box of the detection branch in real time; when the center distance and area deviation exceed the preset drift range, it is determined to be a tracking drift, triggering the detection branch to re-detect the current frame and adjacent frames where the drift occurred; during detection, the same detection logic as the initial frame is executed, and the tracking matching template of the tracking branch is updated with the new target box detection results.

[0012] Furthermore, the step of denoising the exhaust region of the fused exhaust target box and extracting key parameters to generate a detection report includes: Morphological filtering was used to denoise the exhaust region of the fused exhaust target box to obtain a denoised image. Key parameters of the exhaust flame are extracted from the denoised image, and the extracted key parameters are integrated with the frame number, exhaust flame parameters and detection confidence level to generate a detection report according to a preset text format.

[0013] In summary, the beneficial effects of the present invention are as follows: This invention provides a method for detecting the combustion plume of liquid rocket engines. By constructing a collaborative detection environment using YOLOv5s and SiamMask, and integrating a two-way interactive module for detection and tracking with a dynamic weight allocation module, it achieves deep collaboration and dynamic fusion of detection and tracking results, eliminating the unidirectional dependence on the initial detection result. Simultaneously, by constructing a full-scene plume dataset covering steady-state, non-steady-state, and various complex environmental disturbances, it overcomes the limitation of single-scene coverage in existing technologies, enabling the trained integrated model to adapt to different application scenarios and enhancing its generalization ability. During detection, a multimodal noise suppression component is used for video preprocessing to specifically eliminate various image interferences. The detection branch determines the starting frame and initial target box and initializes the tracking template. The dynamic weight allocation module fuses the detection and tracking results of subsequent frames, denoises the plume region, extracts key parameters, and generates a detection report. This provides reliable data support for engine condition monitoring and fault early warning, ensuring the stability of plume detection. Attached Figure Description

[0014] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments of the present invention will be briefly introduced below. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, and these are all within the protection scope of the present invention.

[0015] Figure 1 This is a flowchart of a liquid rocket engine combustion exhaust flame detection method according to the present invention. Detailed Implementation

[0016] 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. It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Unless otherwise specified, the present invention and the various features in the embodiments can be combined with each other, all of which are within the protection scope of the present invention.

[0017] Example 1: Refer to Figure 1 As shown in the figure, an embodiment of the present invention provides a method for detecting the combustion exhaust plume of a liquid rocket engine, the method mainly including: S1: Construct a collaborative operating environment for the YOLOv5s detection model and the SiamMask tracking model, and integrate a two-way interaction module for detection and tracking, a dynamic weight allocation module, a lightweight model adaptation component, and a multimodal noise suppression component; S2: Collect test videos of liquid rocket engines under various operating conditions, extract images and label the exhaust plume area to construct a full-scene exhaust plume dataset; S3: Based on the lightweight model adaptation component, the YOLOv5s basic model is modified to obtain a lightweight detection model. Using the full-scene exhaust flame dataset, the detection and tracking bidirectional interaction module is called to jointly train the lightweight detection model and the SiamMask tracking model to obtain an integrated detection model; the integrated detection model includes a detection branch and a tracking branch. S4: The test video is preprocessed using a multimodal noise suppression component, and the preprocessed image is input into the detection branch of the integrated detection model for initial frame processing to determine the starting frame and initial target box position of the tail flame and initialize the tracking matching template. S5: Based on the starting frame and tracking matching template of the tail flame, the detection branch and the tracking branch are used to detect and track the preprocessed image simultaneously. The detection results and tracking results of subsequent frames are dynamically fused through the dynamic weight allocation module to output the fused tail flame target box. The tail flame region of the fused tail flame target box is denoised and key parameters are extracted to generate a detection report.

[0018] In this embodiment, S1 constructs a collaborative operating environment for the YOLOv5s detection model and the SiamMask tracking model, and integrates a bidirectional interaction module for detection and tracking, a dynamic weight allocation module, a lightweight model adaptation component, and a multimodal noise suppression component, specifically including: Set up a collaborative runtime environment for the YOLOv5s detection model and the SiamMask tracking model, and install the necessary dependencies for model operation. These dependencies include basic programming environment, deep learning framework, and image processing libraries, ensuring the two models can work together correctly.

[0019] By sharing the feature extraction layer of the backbone network, a bidirectional feature transmission channel is established between YOLOv5s and SiamMask, enabling the spatial features of the exhaust plume extracted by YOLOv5s to be transmitted to SiamMask in real time, while the temporal motion features of the exhaust plume extracted by SiamMask are fed back to YOLOv5s, thus realizing feature exchange between the two and obtaining a bidirectional interactive module for detection and tracking.

[0020] The logic for judging tail flame morphology changes is set based on the overlap of target boxes in adjacent frames and the rate of change of tail flame area. A dynamic weight allocation module is constructed based on this logic, and a lightweight model adaptation component is built based on the model's lightweight modification logic. Specifically, the tail flame morphology change judgment logic is used to flexibly adjust the weight ratio of detection and tracking results according to the stability of the tail flame morphology. The lightweight modification logic involves removing redundant feature channels through channel pruning, simplifying model parameter types through quantization compression, and merging functionally related network layers using layer fusion technology.

[0021] Based on different types of image interference, suitable algorithms for halo suppression, smoke occlusion penetration, and strong light reflection filtering are selected and integrated to construct a multimodal noise suppression component. This multimodal noise suppression component can be used to specifically address different types of image interference.

[0022] In this embodiment, test videos of liquid rocket engines under various operating conditions are collected in S2, and images are extracted and the exhaust plume region is labeled to construct a full-scene exhaust plume dataset, including: Liquid rocket engines with different thrust models were selected and tested from various shooting angles to record steady-state exhaust flame videos under rated operating conditions. Simulate engine start-up, shutdown, and combustion instability fault scenarios, and record unsteady exhaust flame videos; create high-altitude, low-temperature, low-pressure, rain, snow, fog, dust, strong light reflection, exhaust flame halo, and smoke obscuring environments using environmental simulation equipment, and record exhaust flame test videos for the corresponding environments. All recorded exhaust plume experimental videos were processed by frame extraction to extract image frames that met the preset clarity requirements. The image frames were then divided into training and validation sets according to a preset ratio for model training and performance evaluation.

[0023] Using the annotation tool, establish a main file structure that includes an image storage folder and a label storage folder, and store the images to be annotated in the training set and validation set into the image storage folder; Launch the annotation tool, open the image to be annotated in the image storage folder, and annotate the exhaust flame region, object type, and name in the image. After annotation, generate the corresponding label file and save it to the label storage folder. Specifically, the labelimg annotation tool is used. When annotating, launch the labelimg annotation tool, switch to the corresponding annotation mode, open the image to be annotated, and accurately annotate the exhaust flame region in the image by dragging and selecting with the mouse.

[0024] Create a dataset configuration file and store the file storage paths for the training and validation sets, as well as the types and names of the objects in the dataset configuration file; By combining the image storage folder, the label storage folder, and the dataset configuration file, a full-scene exhaust flame dataset is obtained.

[0025] In this embodiment, in S3, the YOLOv5s basic model is modified based on the lightweight model adaptation component to obtain a lightweight detection model. The lightweight detection model and the SiamMask tracking model are jointly trained using a full-scene exhaust plume dataset to obtain an integrated detection model, including: YOLOv5s is used as the base model. The architecture of the base model includes an input layer, a focus layer, a CSPDarknet53 backbone layer, an FPN+PAN neck layer, and a detection head output layer. A lightweight detection model was obtained by using a lightweight model adaptation component to lightweighten the input, focus layer, FPN+PAN neck layer, and detection head output layer of the base model. The modified lightweight detection model retains the original four-layer architecture logic, including a lightweight input, a lightweight focus layer, a CSPDarknet53 backbone layer, a lightweight FPN+PAN neck layer, and a lightweight detection head layer. These layers are sequentially connected, with the neck layer and backbone layer performing multi-scale feature fusion across layers. The backbone layer also shares features with the SiamMask feature extraction layer to achieve feature reuse. The lightweight modification includes channel pruning, quantization compression, and network layer merging.

[0026] A full-scene exhaust plume dataset was input into a lightweight detection model for training. During training, the model's detection performance was periodically evaluated using a validation set. Based on the performance evaluation results, the hyperparameters of the lightweight model were adjusted to obtain a basic detection model with optimal exhaust plume feature recognition capabilities. Hyperparameters included learning rate and anchor box size.

[0027] The detection and tracking bidirectional interaction module connects the SiamMask core functional layer with the basic detection model. The module completes the connection by sharing a backbone feature extraction layer, building a bidirectional feature transmission channel, and having a built-in feature dimension transformation sub-layer. It uses the basic detection model as the detection branch and the SiamMask target tracking model as the tracking branch to build an integrated fusion model framework.

[0028] The core functional layers of the SiamMask target tracking model are the feature extraction layer, the cross-correlation layer between the template branch and the search branch, the mask prediction layer, and the target bounding box regression layer. The functions of each core layer are as follows: the feature extraction layer extracts the deep semantic and shallow texture features of the exhaust flame image, providing basic feature support for tracking; the cross-correlation layer completes the feature matching between the template branch and the search branch to locate the approximate area of ​​the exhaust flame; the mask prediction layer generates a pixel-level mask to accurately segment the exhaust flame from the background; and the target bounding box regression layer makes fine adjustments to the matching area and outputs an accurate exhaust flame target bounding box.

[0029] The specific integration process of connecting the SiamMask core functional layer with the basic detection model through the detection and tracking bidirectional interaction module is as follows: First, a shared feature extraction layer is established, where the Focus layer + CSPDarknet53 layer of the basic detection model is shared with the SiamMask feature extraction layer. This enables one-time extraction and bidirectional reuse of exhaust plume image features, outputting the spatial features required for detection and the temporal motion features required for tracking. Second, a bidirectional feature transmission channel is established, transferring the spatial features of the basic detection model to the SiamMask cross-correlation layer to improve tracking matching accuracy. Simultaneously, the segmentation features of the SiamMask mask prediction layer and the motion features of the target bounding box regression layer are also transferred. The first step involves feature feedback to the detection head layer of the basic detection model to correct detection box bias. The second step is feature dimension adaptation, which uses a built-in feature dimension transformation sublayer to reduce the dimensionality of the detection features input to SiamMask to match its feature dimension, and to increase the dimensionality and perform feature mapping on the tracking features fed back to the basic detection model to adapt to the input requirements of the detection head. The third step is joint training loss propagation, which propagates the tracking loss (mask loss + box regression loss) in the multi-task loss function back to the mask prediction layer, target box regression layer, and cross-correlation layer of SiamMask, and propagates the detection loss combined with the tracking loss feedback value back to each layer of the basic detection model, thereby achieving collaborative optimization training of the two models.

[0030] A multi-task loss function, including detection loss and tracking loss, is adopted. The integrated fusion model is jointly trained using the training set and the validation set. The multi-task loss values ​​during joint training are backpropagated bidirectionally to each layer for collaborative optimization. After training, the integrated detection model is obtained.

[0031] Among them, bidirectional backpropagation involves backpropagating the tracking loss to each core layer of SiamMask, and backpropagating the detection loss combined with the tracking loss feedback value to each layer of the basic detection model, thereby achieving collaborative optimization between the two.

[0032] In this embodiment, S4 preprocesses the test video using a multimodal noise suppression component, and inputs the preprocessed image into the detection branch of the integrated detection model for initial frame processing. This process determines the starting frame and initial target box position of the exhaust flame and initializes the tracking matching template, including: All recorded exhaust flame experimental videos are parsed into image sequences frame by frame. Each frame of the image sequence is fed into a multimodal noise suppression component, and image interference is removed sequentially through a halo suppression algorithm, a smoke occlusion penetration algorithm, and a strong light reflection filtering algorithm. The preprocessed image is then output. The preprocessed image is input into the detection branch of the integrated detection model. The detection branch identifies the tail flame and determines the starting frame of the tail flame combustion, while locating the initial target box position of the tail flame. The initial target box position is then passed to the tracking branch of the integrated model to complete the initialization of the tracking matching template.

[0033] In this embodiment, S5 uses the starting frame of the exhaust plume and the tracking matching template to simultaneously detect and track the preprocessed image using the detection branch and the tracking branch. The detection and tracking results of subsequent frames are dynamically fused through the dynamic weight allocation module to output the fused exhaust plume target bounding box, including: The detection branch detects the exhaust flame based on the starting frame of the exhaust flame and performs exhaust flame detection on subsequent image frames in the preprocessed image. Meanwhile, the tracking branch tracks the exhaust flame based on the tracking matching template and performs exhaust flame tracking on subsequent image frames in the preprocessed image, respectively obtaining the exhaust flame detection result and the tracking result. The dynamic weight allocation module is invoked to determine the degree of change in the tail flame shape based on the overlap of the tail flame target boxes in adjacent frames and the rate of change of the tail flame area. The weights of the detection results and tracking results are allocated according to the degree of change in the tail flame shape. A weighted fusion method is adopted to sum the x / y center coordinates and width / height parameters of the detection and tracking target boxes respectively to obtain the precise coordinates of the fused tail flame target box. The detection and tracking bidirectional interaction module dynamically adjusts the feature interaction weights according to the degree of change in the tail flame shape. When the shape is stable, the tracking feature weight is increased, and when the shape changes drastically, the detection feature weight is increased, realizing the dynamic collaborative adaptation of detection and tracking, and fusing to obtain the fused tail flame target box of the current image frame.

[0034] In this embodiment, S5 further includes: calculating the center distance and area deviation between the tracking target box of the tracking branch and the detection target box of the detection branch in real time; when the center distance and area deviation exceed the preset drift range, it is determined to be tracking drift, triggering the detection branch to re-detect the current frame and adjacent frames where the drift occurred (not returning to the original tail flame combustion start frame of the video); after detection, the tracking matching template of the tracking branch is updated with the new target box detection result, allowing tracking to start again from the drift frame, and the tracking matching template of the tracking branch is updated with the new start frame detection result to ensure tracking stability.

[0035] In this embodiment, step S5 involves denoising the exhaust region of the fused exhaust target box and extracting key parameters to generate a detection report, including: Morphological filtering is used to denoise the tail flame region of the fused tail flame target box, removing isolated noise points in the image to obtain a denoised image. Key parameters of the exhaust plume are extracted from the denoised image. These extracted parameters are then integrated with the frame number, exhaust plume parameters, and detection confidence level to generate a detection report in a preset text format, providing data support for engine condition monitoring and fault early warning. The key parameters include crucial feature parameters such as the area, length, and center coordinates of the exhaust plume.

[0036] This invention achieves deep collaboration between YOLOv5s and SiamMask through two-way interaction of detection and tracking, dynamic weight allocation and drift correction mechanism, which gets rid of the one-way dependence on the initial detection result in the citation technology. Even if there is a slight deviation in the initial detection, it can be corrected in time through tracking feedback, which helps to avoid conflict between detection and tracking results and improves the stability of exhaust flame detection under complex working conditions. A full-scenario dataset covering steady-state, non-steady-state, and complex environmental disturbances was constructed, which made up for the deficiency of the single scenario coverage of the citation technology. This enabled the model to adapt to different application scenarios such as ground test driving and high-altitude flight, as well as various working conditions such as startup, shutdown, and failure. The false detection rate and false negative rate were significantly reduced in complex environments, and the generalization ability was stronger. The detection model obtained through lightweight modification is adapted to low-computing embedded terminals, which solves the problem that large models in the citation technology are difficult to deploy on rocket-borne equipment. It can achieve real-time detection without relying on high-performance GPUs, which is more in line with the engineering application requirements of lightweight and low-computing aerospace equipment, and its applicability is improved. The multimodal noise suppression component specifically addresses various image interferences such as halos, smoke, and strong light, effectively restoring the true characteristics of the tail flame. It solves the problems of citation technology lacking a dedicated anti-interference mechanism and insufficient detection accuracy in complex environments, ensuring stable identification of the tail flame even in harsh environments.

[0037] 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 the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting the combustion exhaust plume of a liquid rocket engine, characterized in that, include: We construct a collaborative operating environment for the YOLOv5s detection model and the SiamMask tracking model, and integrate a two-way interaction module for detection and tracking, a dynamic weight allocation module, a lightweight model adaptation component, and a multimodal noise suppression component. Test videos of liquid rocket engines under various operating conditions were collected, and images were extracted and the exhaust plume region was labeled to construct a full-scene exhaust plume dataset; A lightweight detection model is obtained by modifying the YOLOv5s basic model based on a lightweight model adaptation component. The lightweight detection model and the SiamMask tracking model are jointly trained using a full-scene exhaust flame dataset to obtain an integrated detection model. The integrated detection model includes a detection branch and a tracking branch. The test video is preprocessed using a multimodal noise suppression component, and the preprocessed image is input into the detection branch of the integrated detection model for initial frame processing to determine the starting frame and initial target box position of the tail flame and initialize the tracking matching template. Based on the starting frame and tracking matching template of the exhaust flame, the preprocessed image is detected and tracked synchronously by the detection branch and the tracking branch. The detection results and tracking results of subsequent frames are dynamically fused through the dynamic weight allocation module to output the fused exhaust flame target box. The exhaust flame region of the fused exhaust flame target box is denoised and key parameters are extracted to generate a detection report.

2. The method for detecting the combustion exhaust plume of a liquid rocket engine according to claim 1, characterized in that, The constructed collaborative operating environment for the YOLOv5s detection model and the SiamMask tracking model, integrating a bidirectional interaction module for detection and tracking, a dynamic weight allocation module, a lightweight model adaptation component, and a multimodal noise suppression component, includes: Set up a collaborative runtime environment for the YOLOv5s detection model and the SiamMask tracking model, and install the dependency packages required for model operation; By sharing the feature extraction layer of the backbone network, a bidirectional feature transmission channel is established between YOLOv5s and SiamMask, so that the spatial features of the exhaust flame extracted by YOLOv5s can be transmitted to SiamMask in real time, while the temporal motion features of the exhaust flame extracted by SiamMask are fed back to YOLOv5s, thus obtaining a bidirectional interactive module for detection and tracking. The logic for judging the change of tail flame shape is set based on the overlap of the target boxes of the tail flame in adjacent frames and the rate of change of tail flame area. A dynamic weight allocation module is constructed based on the logic for judging the change of tail flame shape, and a lightweight model adaptation component is constructed based on the logic for lightweight transformation of the model. Based on different types of image interference, we select appropriate halo suppression algorithms, smoke occlusion penetration algorithms, and strong light reflection filtering algorithms to integrate and construct a multimodal noise suppression component.

3. The method for detecting the combustion exhaust plume of a liquid rocket engine according to claim 1, characterized in that, The process involves collecting test videos of liquid rocket engines under various operating conditions, extracting images, labeling the exhaust plume region, and constructing a full-scene exhaust plume dataset, including: Liquid rocket engines with different thrust models were selected and tested from various shooting angles to record steady-state exhaust flame videos under rated operating conditions. Simulate engine start-up, shutdown, and combustion instability fault scenarios, and record unsteady exhaust flame videos; create high-altitude, low-temperature, low-pressure, rain, snow, fog, dust, strong light reflection, exhaust flame halo, and smoke obscuring environments using environmental simulation equipment, and record exhaust flame test videos for the corresponding environments. All recorded exhaust flame experimental videos were processed by frame extraction to extract image frames that meet the preset clarity requirements, and the image frames were divided into training set and validation set according to the preset ratio. Using the annotation tool, establish a main file structure that includes an image storage folder and a label storage folder, and store the images to be annotated in the training set and validation set into the image storage folder; Launch the annotation tool, open the image to be annotated in the image storage folder, and annotate the exhaust flame region, object type, and name in the image; after annotation, generate the corresponding label file and save it to the label storage folder; Create a dataset configuration file and store the file storage paths for the training and validation sets, as well as the types and names of the objects in the dataset configuration file; By combining the image storage folder, the label storage folder, and the dataset configuration file, a full-scene exhaust flame dataset is obtained.

4. The method for detecting the combustion exhaust plume of a liquid rocket engine according to claim 1, characterized in that, The lightweight detection model is obtained by modifying the YOLOv5s basic model based on the lightweight model adaptation component. The lightweight detection model and the SiamMask tracking model are then jointly trained using a full-scene exhaust plume dataset to obtain an integrated detection model, including: YOLOv5s is used as the base model. The architecture of the base model includes an input layer, a focus layer, a CSPDarknet53 backbone layer, an FPN+PAN neck layer, and a detection head output layer. A lightweight detection model is obtained by using a lightweight model adaptation component to lightweightly modify the input end, focus layer, FPN+PAN neck layer, and detection head output layer of the basic model. The lightweight detection model includes a lightweight input end, a lightweight focus layer, a CSPDarknet53 backbone layer, a lightweight FPN+PAN neck layer, and a lightweight detection head layer, with each layer connected sequentially. The lightweight modification includes channel pruning, quantization compression, and network layer merging. The full-scene exhaust plume dataset is input into the lightweight detection model for model training. During the training process, the detection performance of the model is evaluated periodically using the validation set. The hyperparameters of the lightweight model are adjusted based on the detection performance evaluation results to obtain the basic detection model with the best ability to recognize exhaust plume features. The detection and tracking bidirectional interaction module connects the core functional layer of SiamMask with the basic detection model, and uses the basic detection model as the detection branch and the SiamMask target tracking model as the tracking branch to build an integrated fusion model framework; the core functional layer includes a feature extraction layer, a cross-correlation layer, a mask prediction layer and a target box regression layer. A multi-task loss function, including detection loss and tracking loss, is adopted. The integrated fusion model is jointly trained using the training set and the validation set. The multi-task loss values ​​during joint training are backpropagated bidirectionally to each layer for collaborative optimization. After training, the integrated detection model is obtained.

5. The method for detecting the combustion exhaust plume of a liquid rocket engine according to claim 1, characterized in that, The process involves preprocessing the test video using a multimodal noise suppression component, inputting the preprocessed image into the detection branch of the integrated detection model for initial frame processing, determining the starting frame and initial target box position of the exhaust flame, and initializing the tracking matching template, including: All recorded exhaust flame experimental videos are parsed into image sequences frame by frame. Each frame of the image sequence is fed into a multimodal noise suppression component, and image interference is removed sequentially through a halo suppression algorithm, a smoke occlusion penetration algorithm, and a strong light reflection filtering algorithm. The preprocessed image is then output. The preprocessed image is input into the detection branch of the integrated detection model. The detection branch identifies the tail flame and determines the starting frame of the tail flame combustion, while locating the initial target box position of the tail flame. The initial target box position is then passed to the tracking branch of the integrated model to complete the initialization of the tracking matching template.

6. The method for detecting the combustion exhaust plume of a liquid rocket engine according to claim 1, characterized in that, The starting frame and tracking matching template based on the exhaust plume are used to simultaneously detect and track the preprocessed image using detection and tracking branches. The detection and tracking results of subsequent frames are dynamically fused through a dynamic weight allocation module to output a fused exhaust plume target bounding box, including: The detection branch detects the exhaust flame based on the starting frame of the exhaust flame and performs exhaust flame detection on subsequent image frames in the preprocessed image. Meanwhile, the tracking branch tracks the exhaust flame based on the tracking matching template and performs exhaust flame tracking on subsequent image frames in the preprocessed image, respectively obtaining the exhaust flame detection result and the tracking result. The dynamic weight allocation module is invoked to determine the degree of change in the tail flame shape based on the overlap of the tail flame target boxes in adjacent frames and the rate of change of the tail flame area. The weights of the detection results and the tracking results are allocated according to the degree of change in the tail flame shape. The coordinate parameters of the detected target box and the tracked target box are weighted and summed using a weighted fusion method to obtain the fused tail flame target box of the current image frame.

7. The method for detecting the combustion exhaust plume of a liquid rocket engine according to claim 6, characterized in that, Also includes: The center distance and area deviation between the tracking target box of the tracking branch and the detection target box of the detection branch are calculated in real time. When the center distance and area deviation exceed the preset drift range, it is determined to be tracking drift, and the detection branch is triggered to re-detect the current frame and adjacent frames where the drift occurred. During detection, the same detection logic as the initial frame is executed, and the tracking matching template of the tracking branch is updated with the new target box detection results.

8. The method for detecting the combustion exhaust plume of a liquid rocket engine according to claim 1, characterized in that, The step of denoising the tail flame region of the fused tail flame target box and extracting key parameters to generate a detection report includes: using morphological filtering to denoise the tail flame region of the fused tail flame target box to obtain a denoised image; extracting key tail flame parameters from the denoised image, and integrating the extracted key parameters with the frame number, tail flame parameters and detection confidence score to generate a detection report according to a preset text format.

Images