Low-temperature propellant leakage detection positioning method and device based on TDGA-YOLO

Abstract

The application discloses a low-temperature propellant leakage detection positioning method and device based on TDGA-YOLO. The method is based on a backbone network, a neck and a detection head framework to construct a YOLOv7 model. A TDGA module is constructed based on a deep learning feature enhancement mechanism of a physical law. The TDGA module is integrated into the YOLOv7 model to construct a TDGA-YOLO model. The TDGA-YOLO model is trained through a low-temperature propellant leakage data set to obtain a trained TDGA-YOLO model. Target low-temperature propellant leakage data is input into the trained TDGA-YOLO model for processing, and a leakage detection result is output. The application solves the problems of micro-leakage detection and multi-point continuous leakage detection, and has the advantages of high accuracy and high precision.

Description

Technical Field

[0001] This invention relates to the field of aerospace propellant safety technology, specifically to a method and apparatus for detecting and locating cryogenic propellant leaks based on TDGA-YOLO. Background Technology

[0002] The flammability and explosiveness of propellants are a major reason for the high risks of space launches, making safe, efficient, and intelligent propellant loading a key task in the development of smart launch sites. According to Failure Mode and Effect Analysis (FMEA), leakage is the most common and dangerous failure in propellant loading systems. Therefore, it is necessary to develop efficient, high-precision, and intelligent methods for propellant leak detection and location.

[0003] Currently, propellant leak detection and location methods include acoustic signal detection, negative pressure wave detection, distributed fiber optic sensor detection, self-organizing wireless sensor model detection, and artificial intelligence detection. However, all of these methods have limitations and are not suitable for propellant leak detection and location. Acoustic signal-based leak detection methods are limited by the propagation distance of the acoustic signal and suffer from severe environmental interference and difficulty in location. Negative pressure wave detection methods can only detect leaks greater than 1% of the instantaneous output and cannot detect continuous small leaks. Distributed fiber optic sensors and self-organizing wireless sensors are not suitable for propellant leak detection at space launch sites. Model-based leak detection and location methods can detect small leaks and multi-point continuous leaks, but they rely on expert experience and are greatly affected by the accuracy of the mathematical model and sensors. Using artificial intelligence algorithms for leak detection and location has advantages such as strong robustness, independence from expert experience, and high detection accuracy, but it suffers from problems such as lack of interpretability and strong data dependence.

[0004] Therefore, there is a need to invent a leak detection and location method that is unaffected by propellant and pipeline type and can effectively detect minute leaks and multi-point continuous leaks, for use in the propellant loading system of aerospace launch sites for leak detection and location. Summary of the Invention

[0005] To address this issue, this invention provides a method and apparatus for detecting and locating cryogenic propellant leaks based on TDGA-YOLO. By introducing a thermally diffused attention mechanism, it solves the problem of the current YOLO algorithm's lack of interpretability. Using TDGA-YOLO to detect and locate cryogenic propellant leaks solves the problems of detecting minute leaks and multi-point continuous leaks, offering advantages such as high accuracy and precision.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a cryogenic propellant leak detection and location method based on TDGA-YOLO, comprising:

[0007] A YOLOv7 model is constructed based on the backbone network, neck and detection head framework; a TDGA module is constructed based on a deep learning feature enhancement mechanism based on physical laws.

[0008] The TDGA module is integrated into the YOLOv7 model to construct the TDGA-YOLO model.

[0009] The TDGA-YOLO model was trained using a cryogenic propellant leakage dataset to obtain a trained TDGA-YOLO model.

[0010] The target cryogenic propellant leakage data is input into the trained TDGA-YOLO model for processing, and the leakage detection results are output.

[0011] As a preferred embodiment of the TDGA-YOLO-based cryogenic propellant leak detection and localization method, the backbone network adopts an enhanced high-efficiency layer aggregation network; the neck adopts a feature pyramid network structure; and the detection head performs multi-scale prediction through several convolutional layers.

[0012] As a preferred embodiment of the TDGA-YOLO-based cryogenic propellant leak detection and location method, the processing steps of the TDGA-YOLO model are as follows:

[0013] The backbone network is used to extract features from the input image to obtain an image feature map;

[0014] Low-temperature regions are detected using a threshold segmentation strategy to generate a cold region mask. Based on the cold region mask, the heat diffusion process is simulated using Gaussian filtering to obtain a heat diffusion attention map.

[0015] Multiply the heat diffusion attention map with the image feature map to obtain a multi-scale feature map;

[0016] The multi-scale feature map is input into the neck area for fusion, and a neck feature map is output.

[0017] The detection head classifies and judges the candidate regions of the neck feature map to obtain the set anchor boxes, categories and confidence levels; non-maximum suppression processing is applied to the set anchor boxes, categories and confidence levels to output the detection results containing object category, confidence level and bounding box coordinates.

[0018] As a preferred embodiment of the TDGA-YOLO-based cryogenic propellant leak detection and location method, in the process of detecting the cryogenic region and generating the cold region mask through the threshold segmentation strategy, the calculation formula for the cryogenic threshold is:

[0019]

[0020] In the formula, T thres The low temperature threshold for the current image; m This is the average value of the current image pixels; s The standard deviation of the current image pixel values; i The coefficient of the standard deviation;

[0021] The formula for calculating the cold region mask is:

[0022]

[0023] In the formula, For cold regions; It is the sigmoid function; s This is the temperature scaling factor; for( x , y The pixel value at the location.

[0024] As a preferred embodiment of the TDGA-YOLO-based cryogenic propellant leak detection and location method, the mathematical expression of the thermal diffusion attention map is:

[0025]

[0026] In the formula, This is the normalized thermal diffusion attention plot; A ( x , y ) represents the attention map after diffusion; min( A ) represents the minimum pixel value of the thermal diffusion attention map; max( A ) represents the maximum pixel value of the thermal diffusion attention map.

[0027] This invention also provides a cryogenic propellant leak detection and location device based on TDGA-YOLO, which, based on the above-mentioned cryogenic propellant leak detection and location method based on TDGA-YOLO, includes:

[0028] YOLOv7 model and TDGA module building units are used to build YOLOv7 models based on backbone networks, neck and detection head frameworks; and to build TDGA modules based on deep learning feature enhancement mechanisms based on physical laws.

[0029] The TDGA-YOLO model building unit is used to integrate the TDGA module into the YOLOv7 model to build the TDGA-YOLO model.

[0030] The TDGA-YOLO model training unit is used to train the TDGA-YOLO model using a cryogenic propellant leakage dataset to obtain a trained TDGA-YOLO model.

[0031] The TDGA-YOLO model processing unit is used to input the target cryogenic propellant leakage data into the trained TDGA-YOLO model for processing and output leakage detection results.

[0032] As a preferred embodiment of the cryogenic propellant leak detection and location device based on TDGA-YOLO, in the YOLOv7 model and TDGA module construction unit, the backbone network adopts an enhanced high-efficiency layer aggregation network; the neck adopts a feature pyramid network structure; and the detection head performs multi-scale prediction through several convolutional layers.

[0033] As a preferred embodiment of the TDGA-YOLO-based cryogenic propellant leak detection and location device, the TDGA-YOLO model processing unit includes the following sub-unit:

[0034] The image feature map acquisition subunit is used to extract features from the input image through the backbone network to obtain an image feature map;

[0035] The thermal diffusion attention map acquisition subunit is used to detect low-temperature regions through a threshold segmentation strategy and generate a cold region mask; based on the cold region mask, the thermal diffusion process is simulated through Gaussian filtering to obtain a thermal diffusion attention map.

[0036] A multi-scale feature map acquisition subunit is used to multiply the heat diffusion attention map with the image feature map to obtain a multi-scale feature map;

[0037] The neck feature map acquisition subunit is used to input the multi-scale feature map into the neck for fusion and output the neck feature map;

[0038] The detection result output subunit is used to classify and judge the candidate regions of the neck feature map through the detection head to obtain the set anchor boxes, categories and confidence levels; and to perform non-maximum suppression processing on the set anchor boxes, categories and confidence levels to output the detection results containing object category, confidence level and bounding box coordinates.

[0039] As a preferred embodiment of the TDGA-YOLO-based cryogenic propellant leak detection and location device, in the thermal diffusion attention map acquisition subunit of the TDGA-YOLO model processing unit, during the process of detecting the cryogenic region and generating the cold region mask through the threshold segmentation strategy, the calculation formula for the cryogenic threshold is as follows:

[0040]

[0041] In the formula, T thres The low temperature threshold for the current image; m This is the average value of the current image pixels; s The standard deviation of the current image pixel values; i The coefficient of the standard deviation;

[0042] The formula for calculating the cold region mask is:

[0043]

[0044] In the formula, For cold regions; It is the sigmoid function; s This is the temperature scaling factor; for( x , y The pixel value at the location.

[0045] As a preferred embodiment of the cryogenic propellant leak detection and location device based on TDGA-YOLO, the mathematical expression of the thermal diffusion attention map in the thermal diffusion attention map acquisition subunit of the TDGA-YOLO model processing unit is:

[0046]

[0047] In the formula, This is the normalized thermal diffusion attention plot; A ( x , y ) represents the attention map after diffusion; min( A ) represents the minimum pixel value of the thermal diffusion attention map; max( A ) represents the maximum pixel value of the thermal diffusion attention map.

[0048] This invention has the following advantages: It constructs a YOLOv7 model based on a backbone network, neck, and detection head framework; it constructs a TDGA module based on a deep learning feature enhancement mechanism using physical laws; it integrates the TDGA module into the YOLOv7 model to construct a TDGA-YOLO model; it trains the TDGA-YOLO model using a cryogenic propellant leakage dataset to obtain a trained TDGA-YOLO model; it inputs the target cryogenic propellant leakage data into the trained TDGA-YOLO model for processing and outputs leakage detection results. This invention achieves efficient and intelligent monitoring of propellant leaks at space launch sites by innovatively integrating physical mechanisms and deep learning technology. Furthermore, it captures cryogenic regions using infrared thermal imaging technology and uses the thermodynamic diffusion equation to simulate the temperature field propagation characteristics to generate a dynamic attention weight map, dynamically enhancing the network's regional response capability to leakage features. Building upon this foundation, a multi-scale feature fusion architecture was designed, combining an enhanced high-efficiency layer aggregation network (E-ELAN) and a feature pyramid network (FPN). Through a top-down and horizontally connected hierarchical fusion mechanism, semantic information and spatial details at different levels were integrated, significantly improving the representation capability of minute leak features. To address the pain point of insufficient interpretability in traditional detection methods, this invention innovatively introduces a physical prior knowledge-guided attention mechanism, ensuring that the model's focus area closely matches the actual thermodynamic characteristics of the leak, giving the detection results clear physical meaning. At the engineering implementation level, this invention employs Soft-NMS to optimize the post-processing workflow, combined with a lightweight network design, achieving millisecond-level response times to meet the real-time requirements of aerospace launch sites. This invention also utilizes temperature sensitivity coefficient adaptive calibration technology to reduce reliance on expensive calibration equipment, enabling the system to maintain high precision operation under complex conditions such as temperature ranges from -183℃ to room temperature, dust interference, and drastic changes in illumination. This invention solves the problems of minute leak detection and multi-point continuous leak detection, offering advantages of high accuracy and precision. Attached Figure Description

[0049] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0050] The structures, proportions, sizes, etc. illustrated in this specification are only for the purpose of assisting those skilled in the art in understanding and reading the content disclosed herein, and are not intended to limit the conditions under which the present invention can be implemented. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportions, or adjustments to the size, without affecting the effects and objectives that the present invention can produce, should still fall within the scope of the technical content disclosed in the present invention.

[0051] Figure 1 This is a flowchart illustrating the cryogenic propellant leak detection and location method based on TDGA-YOLO provided in Embodiment 1 of the present invention.

[0052] Figure 2 This is a schematic diagram illustrating the specific implementation framework of the cryogenic propellant leak detection and location method based on TDGA-YOLO provided in Embodiment 1 of the present invention;

[0053] Figure 3 This is a schematic diagram of the architecture of the cryogenic propellant leak detection and location device based on TDGA-YOLO provided in Embodiment 2 of the present invention. Detailed Implementation

[0054] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] Example 1

[0056] See Figure 1 and Figure 2 Embodiment 1 of the present invention provides a method for detecting and locating cryogenic propellant leaks based on TDGA-YOLO, comprising the following steps:

[0057] S1. Based on the backbone network, neck and detection head framework, construct the YOLOv7 model; based on the deep learning feature enhancement mechanism based on physical laws, construct the TDGA module;

[0058] S2. Integrate the TDGA module into the YOLOv7 model to construct the TDGA-YOLO model;

[0059] S3. Train the TDGA-YOLO model using the cryogenic propellant leakage dataset to obtain the trained TDGA-YOLO model;

[0060] S4. Input the target cryogenic propellant leakage data into the trained TDGA-YOLO model for processing and output the leakage detection results.

[0061] In this embodiment, in step S1, a YOLOv7 model is constructed based on the backbone network, neck and detection head framework; and a TDGA module is constructed based on a deep learning feature enhancement mechanism based on physical laws.

[0062] Specifically, the core of the YOLOv7 model includes three modules: the backbone network, the neck, and the detection head.

[0063] The backbone network employs Enhanced EfficientLayer Aggregation Networks (E-ELAN), which enhances the network's ability to perceive details in the image by strengthening edge information in the feature map.

[0064] Its specific parameters are set as follows: number of layer aggregations: (3,6,9,3); channel expansion ratio: 2 times; activation function: SiLU function.

[0065] The neck section employs a Feature Pyramid Networks (FPN) structure to fuse feature maps from different levels, enabling the detection of targets of varying sizes. Simultaneously, a Path Aggregation Network (PANet) is utilized to improve feature fusion, enhancing both detection accuracy and inference efficiency.

[0066] Its specific parameters are set as follows: number of upsampling channels: 256; number of top-down path channels: 512.

[0067] The detection head uses multiple convolutional layers to generate the final prediction results, enabling multi-scale prediction. It provides five default anchor box sizes: (10×10), (20×20), (40×40), (80×80), and (160×160), supporting multi-class classification and bounding box regression.

[0068] In this embodiment, the TDGA module is a deep learning feature enhancement mechanism based on physical laws. Its core idea is to model the temperature field propagation characteristics of the cryogenic leakage area through the thermodynamic diffusion equation and generate a physical prior attention map to dynamically adjust the feature response of the neural network.

[0069] Specifically, the low-temperature region is first extracted from the infrared image, then Gaussian kernel convolution is used to simulate the thermal diffusion process to generate thermodynamic attention weights, which are finally applied to the feature map of the backbone network in a multiplicative manner.

[0070] In this embodiment, the YOLOv7 model can also be replaced by various object detection algorithm models such as conditional Faster-rcnn.

[0071] In this embodiment, in step S2, the TDGA module is integrated into the YOLOv7 model to construct the TDGA-YOLO model;

[0072] Specifically, a Thermal Diffusion Guided Attention (TDGA) module is embedded at the output of the backbone network in the YOLOv7 model to generate the TDGA-YOLO model. The TDGA-YOLO model dynamically focuses on the subtle temperature changes in the leak area by combining physical mechanism modeling (thermal diffusion equation) with deep learning feature enhancement, thus solving the problem of insufficient sensitivity of traditional AI methods to minor leaks.

[0073] In this embodiment, in step S3, the TDGA-YOLO model is trained using a cryogenic propellant leakage dataset to obtain a trained TDGA-YOLO model;

[0074] Specifically, the training steps for the TDGA-YOLO model are as follows:

[0075] T1. Data Preparation:

[0076] Data source: Infrared video measured at a space launch site (-183℃ liquid oxygen environment);

[0077] Sample size: 2342 labeled images (including samples with leakage areas of 0.5-50 mm²);

[0078] Data augmentation: MixUp (probability 0.5), CutMix (probability 0.3), random rotation (±15°).

[0079] T2, Training Configuration:

[0080] Optimizer: AdamW;

[0081] Learning rate scheduling: cosine annealing strategy (initial lr=1e-3, min lr=1e-5).

[0082] Loss function: CIoU Loss (classification weight 0.5, regression weight 1.5);

[0083] Batch size: 16;

[0084] Training rounds: 50.

[0085] T3, Model Lightweighting:

[0086] Channel pruning: Remove redundant branches (accuracy loss <1%).

[0087] Quantization unlock: Inference speed is increased by 35% after INT8 quantization.

[0088] In this embodiment, through data augmentation and loss function design guided by physical priors, the model's generalization ability to low-temperature and low-contrast leakage scenarios is significantly improved, and the detection accuracy reaches 98.7%.

[0089] In this embodiment, in step S4, the target cryogenic propellant leakage data is input into the trained TDGA-YOLO model for processing, and the leakage detection result is output.

[0090] Specifically, the processing steps of the TDGA-YOLO model are as follows:

[0091] S41. Extract features from the input image through the backbone network to obtain an image feature map;

[0092] Specifically, the backbone network processes the input 512×512 resolution infrared thermal imaging image. During processing, the backbone network employs a four-stage enhanced layer aggregation strategy, with each stage involving (3, 6, 9, 3) aggregations, enhancing edge feature extraction capabilities through cross-layer connections; the channel expansion ratio is 2 times, and the activation function is SiLU (Sigmoid-Weighted Linear Unit) to improve nonlinear representation capabilities. After processing, image feature maps are obtained; these feature maps are multi-level semantic information feature maps (C3, C4, C5), corresponding to the detection requirements of small, medium, and large targets, respectively.

[0093] S42. Detect the low-temperature region using a threshold segmentation strategy to generate a cold region mask; based on the cold region mask, simulate the heat diffusion process using Gaussian filtering to obtain a heat diffusion attention map.

[0094] The temperature of the leaking droplets gradually rises during the dripping process; therefore, a low-temperature threshold needs to be calculated individually for each image. The formula for calculating the low-temperature threshold is:

[0095]

[0096] In the formula, T thres The low temperature threshold for the current image; m This is the average value of the current image pixels; s The standard deviation of the current image pixel values; i The coefficient of the standard deviation;

[0097] The formula for calculating the cold region mask is:

[0098]

[0099] In the formula, For cold regions; It is the sigmoid function; s This is the temperature scaling factor; for( x , y The pixel value at the location.

[0100] The expression for the sigmoid function is:

[0101]

[0102] In the formula, e It is a natural constant (an irrational number of approximately 2.71828); x This is the input value for the sigmoid function.

[0103] In this embodiment, Gaussian filtering achieves smoothing of temperature distribution by convolving the image with a Gaussian kernel, thereby simulating the heat diffusion process.

[0104] The phenomenon of thermal diffusion can usually be described by the heat conduction equation, which has the following continuous form:

[0105]

[0106] In the formula, t For time; T ( x, y, t () represents the temperature field; α The thermal diffusivity; For the Laplace operator.

[0107] In single-frame processing, a steady-state simplification can be performed, and the steady-state solution of the heat conduction equation can be approximated by Gaussian filtering. Its mathematical form is:

[0108]

[0109]

[0110] In the formula, G A Gaussian kernel is used to control the extent of diffusion; d The standard deviation of the Gaussian kernel; Indicates the convolution operation; A ( x , y () is the attention map after diffusion; x The x-coordinate of the pixel; y The vertical coordinate of the pixel is denoted as y.

[0111] in, d Based on the contrast ratio calculation:

[0112]

[0113] In the formula, T max The maximum pixel value of the current image; T min This represents the minimum pixel value of the current image. The δ calculated in this way increases as the contrast decreases, meaning the lower the contrast, the stronger the diffusion.

[0114] Finally, the diffused attention map is normalized to the range [0, 1], and its mathematical form is as follows:

[0115]

[0116] In the formula, This is the normalized thermal diffusion attention plot; A ( x , y ) represents the attention map after diffusion; min( A ) represents the minimum pixel value of the thermal diffusion attention map; max( A ) represents the maximum pixel value of the thermal diffusion attention map.

[0117] S43. Multiply the heat diffusion attention map with the image feature map to obtain a multi-scale feature map;

[0118] Specifically, the heat diffusion attention map The feature maps (C3 / C4 / C5) output by the backbone network are multiplied element-wise, and the key channel features are further enhanced through the attention mechanism (SE-Block) to obtain the enhanced feature maps. The enhanced feature maps are then upsampled and downsampled to generate feature maps of three different scales (P3 / P4 / P5).

[0119] S44. Input the multi-scale feature map into the neck area for fusion, and output the neck feature map;

[0120] Specifically, deep semantic features (P5) are upsampled and fused with shallow spatial features (P4 / P3) to supplement low-level detailed information; horizontal convolution is performed between feature maps of the same scale to enhance local contextual association; 1×1 convolution is used to adjust the number of channels of each level of feature map to a uniform dimension (default 256 channels) to output neck feature map.

[0121] S45. The candidate regions of the neck feature map are classified and judged by the detection head to obtain the set anchor boxes, categories and confidence levels; non-maximum suppression processing is performed on the set anchor boxes, categories and confidence levels to output the detection results containing object categories, confidence levels and bounding box coordinates.

[0122] Specifically, candidate regions are generated from the fused feature map using five default anchor boxes (10×10 to 160×160), covering scenarios ranging from small to large leaks. The candidate regions are classified and judged, and the set anchor boxes, categories, and confidence scores are predicted using convolutional layers. Non-maximum suppression is then applied to the set anchor boxes, categories, and confidence scores, outputting detection results containing object category, confidence score, and bounding box coordinates. Output format: JSON / XML, including leak category (e.g., liquid oxygen leak), confidence score, bounding box coordinates (x, y, w, h), and estimated leak area.

[0123] In summary, this invention constructs a YOLOv7 model based on a backbone network, neck, and detection head framework; builds a TDGA module based on a deep learning feature enhancement mechanism using physical laws; integrates the TDGA module into the YOLOv7 model to construct a TDGA-YOLO model; trains the TDGA-YOLO model using a cryogenic propellant leakage dataset to obtain a trained TDGA-YOLO model; and processes the target cryogenic propellant leakage data into the trained TDGA-YOLO model to output leakage detection results. This invention achieves efficient and intelligent monitoring of propellant leaks at space launch sites by innovatively integrating physical mechanisms and deep learning technology. Furthermore, this invention captures cryogenic regions using infrared thermal imaging technology and uses thermodynamic diffusion equations to simulate the temperature field propagation characteristics to generate a dynamic attention weight map, dynamically enhancing the network's regional response capability to leakage features. Building upon this foundation, a multi-scale feature fusion architecture was designed, combining an enhanced high-efficiency layer aggregation network (E-ELAN) and a feature pyramid network (FPN). Through a top-down and horizontally connected hierarchical fusion mechanism, semantic information and spatial details at different levels were integrated, significantly improving the representation capability of minute leak features. To address the pain point of insufficient interpretability in traditional detection methods, this invention innovatively introduces a physical prior knowledge-guided attention mechanism, ensuring that the model's focus area closely matches the actual thermodynamic characteristics of the leak, giving the detection results clear physical meaning. At the engineering implementation level, this invention employs Soft-NMS to optimize the post-processing workflow, combined with a lightweight network design, achieving millisecond-level response times to meet the real-time requirements of aerospace launch sites. This invention also utilizes temperature sensitivity coefficient adaptive calibration technology to reduce reliance on expensive calibration equipment, enabling the system to maintain high precision operation under complex conditions such as temperature ranges from -183℃ to room temperature, dust interference, and drastic changes in illumination. This invention solves the problems of minute leak detection and multi-point continuous leak detection, offering advantages of high accuracy and precision.

[0124] It should be noted that the method of this disclosure embodiment can be executed by a single device, such as a computer or server. The method of this embodiment can also be applied to a distributed scenario, where multiple devices cooperate to complete the task. In such a distributed scenario, one of these devices may execute only one or more steps of the method of this disclosure embodiment, and the multiple devices will interact with each other to complete the method described.

[0125] It should be noted that the above description describes some embodiments of this disclosure. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in a different order than that shown in the above embodiments and still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0126] Example 2

[0127] See Figure 3 Embodiment 2 of the present invention also provides a cryogenic propellant leak detection and location device based on TDGA-YOLO, comprising:

[0128] YOLOv7 model and TDGA module building unit 001 is used to build the YOLOv7 model based on the backbone network, neck and detection head framework; and to build the TDGA module based on the deep learning feature enhancement mechanism based on physical laws.

[0129] The TDGA-YOLO model building unit 002 is used to integrate the TDGA module into the YOLOv7 model to build the TDGA-YOLO model.

[0130] The TDGA-YOLO model training unit 003 is used to train the TDGA-YOLO model using a cryogenic propellant leakage dataset to obtain a trained TDGA-YOLO model.

[0131] The TDGA-YOLO model processing unit 004 is used to input the target cryogenic propellant leakage data into the trained TDGA-YOLO model for processing and output leakage detection results.

[0132] In this embodiment, in the YOLOv7 model and TDGA module construction unit 001, the backbone network adopts an enhanced high-efficiency layer aggregation network; the neck adopts a feature pyramid network structure; and the detection head performs multi-scale prediction through several convolutional layers.

[0133] In this embodiment, the TDGA-YOLO model processing subunit 004 includes:

[0134] Image feature map acquisition subunit 041 is used to extract features from the input image through the backbone network to obtain an image feature map;

[0135] The thermal diffusion attention map acquisition subunit 042 is used to detect low-temperature regions through a threshold segmentation strategy and generate a cold region mask; based on the cold region mask, the thermal diffusion process is simulated through Gaussian filtering to obtain a thermal diffusion attention map.

[0136] Multi-scale feature map acquisition subunit 043 is used to multiply the heat diffusion attention map with the image feature map to obtain a multi-scale feature map;

[0137] Neck feature map acquisition subunit 044 is used to input the multi-scale feature map into the neck for fusion and output a neck feature map;

[0138] The detection result output subunit 045 is used to classify and judge the candidate regions of the neck feature map through the detection head to obtain the set anchor boxes, categories and confidence levels; to perform non-maximum suppression processing on the set anchor boxes, categories and confidence levels, and to output the detection results containing object category, confidence level and bounding box coordinates.

[0139] In this embodiment, in the thermal diffusion attention map acquisition subunit 042 of the TDGA-YOLO model processing unit 004, during the process of detecting low-temperature regions and generating the cold region mask through the threshold segmentation strategy, the calculation formula for the low-temperature threshold is as follows:

[0140]

[0141] In the formula, T thres The low temperature threshold for the current image; m This is the average value of the current image pixels; s The standard deviation of the current image pixel values; i The coefficient of the standard deviation;

[0142] The formula for calculating the cold region mask is:

[0143]

[0144] In the formula, For cold regions; It is the sigmoid function; s This is the temperature scaling factor; for(x , y The pixel value at the location.

[0145] In this embodiment, in the thermal diffusion attention map acquisition subunit 042 of the TDGA-YOLO model processing unit 004, the mathematical expression of the thermal diffusion attention map is:

[0146]

[0147] In the formula, This is the normalized thermal diffusion attention plot; A ( x , y ) represents the attention map after diffusion; min( A ) represents the minimum pixel value of the thermal diffusion attention map; max( A ) represents the maximum pixel value of the thermal diffusion attention map.

[0148] It should be noted that the information interaction and execution process between the modules of the above system are based on the same concept as the method embodiment in Embodiment 1 of this application, and the resulting technical effects are the same as those in the method embodiment of this application. For details, please refer to the description in the method embodiment shown above in this application, and it will not be repeated here.

[0149] Example 3

[0150] Embodiment 3 of the present invention provides a non-transitory computer-readable storage medium storing program code for a cryogenic propellant leak detection and location method based on TDGA-YOLO. The program code includes instructions for executing the cryogenic propellant leak detection and location method based on TDGA-YOLO of Embodiment 1 or any possible implementation thereof.

[0151] Computer-readable storage media can be any available medium that a computer can access, or a data storage device such as a server or data center that integrates one or more available media. The available medium can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state drives (SSDs)).

[0152] Example 4

[0153] Embodiment 4 of the present invention provides an electronic device, including: a memory and a processor;

[0154] The processor and the memory communicate with each other via a bus; the memory stores program instructions that can be executed by the processor, and the processor can call the program instructions to execute the TDGA-YOLO-based cryogenic propellant leak detection and location method of Embodiment 1 or any possible implementation thereof.

[0155] Specifically, a processor can be implemented in hardware or software. When implemented in hardware, the processor can be a logic circuit, an integrated circuit, etc. When implemented in software, the processor can be a general-purpose processor that reads software code stored in memory. This memory can be integrated into the processor or located outside the processor and exist independently.

[0156] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable system. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means.

[0157] It is obvious to those skilled in the art that the modules or steps of the present invention described above can be implemented using general-purpose computing systems. They can be centralized on a single computing system or distributed across a network of multiple computing systems. Optionally, they can be implemented using program code executable by a computing system, thereby storing them in a storage system for execution by the computing system. In some cases, the steps shown or described can be performed in a different order than those presented herein, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the present invention is not limited to any particular combination of hardware and software.

[0158] Although the present invention has been described in detail above with general descriptions and specific embodiments, modifications or improvements can be made to it, which will be obvious to those skilled in the art. Therefore, all such modifications or improvements made without departing from the spirit of the present invention fall within the scope of protection claimed by the present invention.

Claims

1. A method for detecting and locating cryogenic propellant leaks based on TDGA-YOLO, characterized in that, include: A YOLOv7 model is constructed based on the backbone network, neck and detection head framework; a TDGA module is constructed based on a deep learning feature enhancement mechanism based on physical laws. The TDGA module is integrated into the YOLOv7 model to construct the TDGA-YOLO model. The TDGA-YOLO model was trained using a cryogenic propellant leakage dataset to obtain a trained TDGA-YOLO model. The target cryogenic propellant leakage data is input into the trained TDGA-YOLO model for processing, and the leakage detection results are output. The thermal diffusion guided attention (TDGA) module is embedded at the output of the backbone network in the YOLOv7 model to generate the TDGA-YOLO model. The TDGA-YOLO model dynamically focuses on the subtle temperature change features of the leakage area by combining physical mechanism modeling with deep learning feature enhancement. The processing steps of the TDGA-YOLO model are as follows: The backbone network is used to extract features from the input image to obtain an image feature map; Low-temperature regions are detected using a threshold segmentation strategy to generate a cold region mask. Based on the cold region mask, the heat diffusion process is simulated using Gaussian filtering to obtain a heat diffusion attention map. Multiply the heat diffusion attention map with the image feature map to obtain a multi-scale feature map; The multi-scale feature map is input into the neck area for fusion, and a neck feature map is output. The detection head classifies and judges the candidate regions of the neck feature map to obtain the set anchor boxes, categories and confidence levels; non-maximum suppression processing is applied to the set anchor boxes, categories and confidence levels to output the detection results containing object category, confidence level and bounding box coordinates.

2. The cryogenic propellant leak detection and location method based on TDGA-YOLO according to claim 1, characterized in that, The backbone network adopts an enhanced high-efficiency layer aggregation network; the neck adopts a feature pyramid network structure; and the detection head performs multi-scale prediction through several convolutional layers.

3. The cryogenic propellant leak detection and location method based on TDGA-YOLO according to claim 1, characterized in that, In the process of detecting low-temperature regions and generating the cold region mask using the threshold segmentation strategy, the formula for calculating the low-temperature threshold is: In the formula, T thres The low temperature threshold for the current image; μ This represents the average pixel value of the current image. σ The standard deviation of the current image pixel values; θ The coefficient of the standard deviation; The formula for calculating the cold region mask is: In the formula, For cold regions; It is the sigmoid function; s This is the temperature scaling factor; for( x , y The pixel value at the location.

4. The cryogenic propellant leak detection and location method based on TDGA-YOLO according to claim 3, characterized in that, The mathematical expression for the normalized heat diffusion attention map is: In the formula, This is the normalized thermal diffusion attention plot; A ( x , y ) represents the attention map after diffusion; min( A ) represents the minimum pixel value of the thermal diffusion attention map; max( A ) represents the maximum pixel value of the thermal diffusion attention map.

5. A cryogenic propellant leak detection and location device based on TDGA-YOLO, employing the cryogenic propellant leak detection and location method based on TDGA-YOLO as described in any one of claims 1-4, characterized in that, include: YOLOv7 model and TDGA module building units are used to build YOLOv7 models based on backbone networks, neck and detection head frameworks; and to build TDGA modules based on deep learning feature enhancement mechanisms based on physical laws. The TDGA-YOLO model building unit is used to integrate the TDGA module into the YOLOv7 model to build the TDGA-YOLO model. The TDGA-YOLO model training unit is used to train the TDGA-YOLO model using a cryogenic propellant leakage dataset to obtain a trained TDGA-YOLO model. The TDGA-YOLO model processing unit is used to input the target cryogenic propellant leakage data into the trained TDGA-YOLO model for processing and output leakage detection results.

6. The cryogenic propellant leak detection and location device based on TDGA-YOLO according to claim 5, characterized in that, In the YOLOv7 model and TDGA module construction unit, the backbone network adopts an enhanced high-efficiency layer aggregation network; the neck adopts a feature pyramid network structure; and the detection head performs multi-scale prediction through several convolutional layers.

7. The cryogenic propellant leak detection and location device based on TDGA-YOLO according to claim 6, characterized in that, The TDGA-YOLO model processing unit includes the following sub-units: The image feature map acquisition subunit is used to extract features from the input image through the backbone network to obtain an image feature map; The thermal diffusion attention map acquisition subunit is used to detect low-temperature regions through a threshold segmentation strategy and generate a cold region mask; based on the cold region mask, the thermal diffusion process is simulated through Gaussian filtering to obtain a thermal diffusion attention map. A multi-scale feature map acquisition subunit is used to multiply the heat diffusion attention map with the image feature map to obtain a multi-scale feature map; The neck feature map acquisition subunit is used to input the multi-scale feature map into the neck for fusion and output the neck feature map; The detection result output subunit is used to classify and judge the candidate regions of the neck feature map through the detection head to obtain the set anchor boxes, categories and confidence levels; and to perform non-maximum suppression processing on the set anchor boxes, categories and confidence levels to output the detection results containing object category, confidence level and bounding box coordinates.

8. The cryogenic propellant leak detection and location device based on TDGA-YOLO according to claim 7, characterized in that, In the thermal diffusion attention map acquisition subunit of the TDGA-YOLO model processing unit, during the process of detecting low-temperature regions and generating the cold region mask using the threshold segmentation strategy, the formula for calculating the low-temperature threshold is: In the formula, T thres The low temperature threshold for the current image; μ This represents the average pixel value of the current image. σ The standard deviation of the current image pixel values; θ The coefficient of the standard deviation; The formula for calculating the cold region mask is: In the formula, For cold regions; It is the sigmoid function; s This is the temperature scaling factor; for( x , y The pixel value at the location.

9. The cryogenic propellant leak detection and location device based on TDGA-YOLO according to claim 8, characterized in that, In the thermal diffusion attention map acquisition subunit of the TDGA-YOLO model processing unit, the mathematical expression of the thermal diffusion attention map is: In the formula, This is the normalized thermal diffusion attention plot; A ( x , y ) represents the attention map after diffusion; min( A ) represents the minimum pixel value of the thermal diffusion attention map; max( A ) represents the maximum pixel value of the thermal diffusion attention map.

Images