An intelligent judgment method for defects in a rocket engine combustion chamber

Abstract

The present invention relates to an intelligent judgment method for defects in a rocket engine combustion chamber, aiming at defects such as the debonding of the interface between the grain and the shell during the charging process of the grain of a solid rocket engine, and the bubbles in the grain caused by differences in the casting process. An intelligent judgment method for defects in a rocket engine combustion chamber is developed. By using image enhancement and non-local deep features to find more similar features in non-local blocks, the image quality is effectively improved. Aiming at the problem of small defect samples, geometric transformation methods such as mirroring, rotation, translation, distortion, filtering, and contrast adjustment are adopted to enhance the network's learning ability for small-sample defects. To reduce the computational complexity of processing high-dimensional data, the visual attention mechanism and incremental learning method can not only make the network more focused on finding useful information that is more similar to the target features, but also effectively cope with the diversity and generalization of defect samples, greatly improving the intelligent judgment efficiency and accuracy of grain defects.

Description

Technical Field

[0001] The present invention belongs to the technical fields of computer vision and image processing, and relates to a method for intelligent interpretation of defects in a rocket engine combustion chamber. Background Art

[0002] A solid rocket motor (SRM) is the main power device of a missile weapon. Its unique advantages include simple structure, high safety and stability, and easy maintenance, and it is widely used in the aerospace field. As the power fuel of a solid rocket, defects such as debonding at the charge interface and bubbles in the charge caused by the casting process are likely to occur during the manufacturing process of the engine grain. In addition, during transportation or storage, the grain is affected by various combined actions, making problems such as pores, cracks, and debonding at the interface between the grain and the casing more complex. These problems will lead to a reduction in its safety and, in severe cases, affect its service life. Traditional defect detection methods mainly rely on manual visual inspection and interpretation based on the naked eye and experience. With the increasing demand for batch production tasks, relying solely on manual interpretation requires a large amount of manpower and time, and it is difficult to ensure the consistency of products. How to effectively analyze SRM defects and formulate scientific and effective defect treatment plans is an urgent issue to be solved in the field of aerospace technology.

[0003] Currently, for the defect detection of the SRM combustion chamber, the commonly used non-destructive testing methods mainly include: ray scanning method, ultrasonic testing method, laser holographic (speckle) non-destructive testing method, etc. The radiographic method uses a ray-sensitive material placed behind the test piece to be penetrated to receive rays with different intensity distributions after passing through the test piece. Since the ray intensity is proportional to the photographic action of the film emulsion under normal conditions, a latent image is formed on the film under the action of the rays. After darkroom processing, the structural image of the object will be displayed. According to the shape and unevenness of the blackness of the image on the negative film, it is possible to evaluate whether there are defects in the test piece to be detected and the nature, shape, and position of the defects. The pulse echo method mainly uses the strong reflection performance of ultrasonic waves for fault detection. It is very effective for detecting interface debonding and the bonding quality of the casing and the liner, and is currently the most widely used ultrasonic method. The laser holographic (speckle) non-destructive testing method utilizes the characteristic that the surface displacement of the test component changes when an external force is applied, and its displacement change is closely related to its internal defects and stress distribution, and can accurately detect the defect problems of composite components.

[0004] Most of the existing publicly available research on automated CT image interpretation techniques stays at relatively traditional image processing algorithms, with low robustness. Traditional methods basically follow the idea of "manually designed features + classifier". This selection strategy has problems such as poor pertinence, high time complexity, and window redundancy. The present invention provides an intelligent interpretation method for SRM combustion chamber defects based on machine vision, which not only realizes the controllability and stability of CT image detection, but also realizes the acquisition and retention of information, and improves the efficiency of typical defect detection in this application field. Summary of the Invention

[0005] Technical Problems to be Solved

[0006] In order to avoid the deficiencies of the prior art, the present invention proposes an intelligent interpretation method for rocket engine combustion chamber defects, which better solves the problem of intelligent auxiliary interpretation of typical defects (interface debonding and pores) in the SRM combustion chamber, and further improves the quality reliability and safety of SRM products. First of all, through in-depth analysis of the noise characteristics generated during the generation, transmission, and processing of CT images, the present invention studies the noise reduction processing technology for SRM combustion chamber CT images; on the basis of traditional image filtering methods, a method for simultaneously removing mosaics and noise based on residual learning is particularly proposed, and the visual attention mechanism and non-local feature enhancement convolution are used to extract and enhance the non-local feature information in the image, realizing the quality improvement of SRM combustion chamber CT images. Secondly, the present invention studies data enhancement methods for typical defects such as interface debonding and pores in SRM; develops an algorithm for detecting pores and debonding based on the classic Faster-RCNN object detection model; and uses the optimization method of pre-training and fine-tuning to improve the network convergence speed and detection accuracy. The intelligent and accurate interpretation of typical defects in SRM combustion chamber CT images is realized. The present invention discloses an intelligent interpretation method for SRM combustion chamber defects based on CT images, which mainly improves the interpretation accuracy and efficiency of typical defects (interface debonding and pores) in the SRM combustion chamber.

[0007] Specifically, the purpose of the present invention is to improve the following aspects.

[0008] 1. Most of the existing research on CT image interpretation techniques stays at relatively traditional image processing algorithms, with low robustness;

[0009] 2. The idea of traditional manually designed features and classifiers is only applicable to specific defect detection, with low generalization;

[0010] 3. Some algorithms use the method of continuously sliding a sliding window to select regions on the image. This selection strategy has no pertinence, high time complexity, and window redundancy;

[0011] 4. For defects with blurred edges, small sizes, or narrow gray-scale variations, the detection and segmentation accuracy is low.

[0012] Technical solution

[0013] Theoretical research is carried out on the bottleneck problems existing in the typical defect interpretation technology of the SRM combustion chamber CT image and the three aspects of research content, and corresponding models or methods are constructed mainly by using the theories and technologies of artificial intelligence and computer vision. First, through in-depth feature analysis and research on typical defects such as interface debonding and grain pores in the SRM combustion chamber, as well as signals such as noise, the CT image quality of interface debonding and grain pores in the SRM combustion chamber is optimized, and a quality improvement technology for the SRM combustion chamber CT image is proposed. Secondly, focus on overcoming the difficult technical problems faced by the intelligent interpretation of SRM combustion chamber defects based on CT images, collect and establish a small sample data set of typical defects, construct a small sample machine learning method based on a small amount of typical defect label data, and establish a high-precision interpretation model for typical combustion chamber defects in large-scale CT images. Finally, on this basis, develop an intelligent interpretation software for the CT image of typical defects in the SRM combustion chamber to meet the project requirements of the functions and indicators expected for defect interpretation. Adopt a research method combining theoretical analysis, logical reasoning, machine learning, and computer simulation experiments to solve the technical difficulties faced by typical defects in the SRM combustion chamber at the theoretical and technical levels.

[0014] 1. Image quality improvement method based on non-local feature deep neural network

[0015] The quality improvement of CT images is related to the accuracy and robustness of subsequent SRM typical defect interpretation. Using the non-local feature deep neural network module can find more similar features in a larger non-local block, and this method can effectively improve the image quality, which plays a key role in the index performance of the entire SRM combustion chamber typical defect discrimination system.

[0016] 2. Small sample learning method for SRM typical defects based on multi-strategy fusion

[0017] The small sample problem of typical defects in the SRM combustion chamber is a significant feature of this project. How to effectively handle this problem is related to the effectiveness and accuracy of the defect discrimination system. The coping methods include basic image processing operations such as mirroring, rotation, translation, distortion, filtering, and contrast adjustment; there are also methods based on pre-trained networks or transfer learning, and reasonable network structure design methods.

[0018] 3. SRM typical defect discrimination method based on attention mechanism and incremental learning

[0019] The introduction of the visual attention mechanism can reduce the computational complexity of processing high-dimensional input data, and enable the network to focus more on finding useful information that is more similar to the target features, improving the discriminability of the expression; while the incremental learning method can effectively address the increasing defective samples and generalization problems. Therefore, an improved method of YOLO or SSD based on the visual attention mechanism and incremental learning is proposed to perform defect interpretation on the CT images of the SRM combustion chamber.

[0020] Generally speaking, the present invention is an intelligent defect interpretation method for a rocket engine combustion chamber, which constructs corresponding models or methods mainly using artificial intelligence, computer vision theories and technologies. Specifically, a research method combining theoretical analysis, logical reasoning, machine learning and computer simulation experiments is adopted to solve the technical difficulties faced by typical defects in the SRM combustion chamber at the theoretical and technical levels.

[0021] The technical solution of the present invention is as follows:

[0022] An intelligent defect interpretation method for a rocket engine combustion chamber, characterized by the following steps:

[0023] Step 1: Cut the original image of the ultra-high-resolution CT image of the SRM combustion chamber into several small images with a width and height of m*n, and divide them into training samples and test samples, where the overlop ratio is set to 0.2 and the step size is 512;

[0024] The cutting method is: start from the upper left corner of the original image and cut according to m*n, and record the upper left coordinates of the cut small images for use in label positioning;

[0025] Step 2: Adopt a method of simultaneously removing mosaic and noise based on residual learning for the training samples, and perform end-to-end mapping learning between the noisy, low-resolution space and the clear, high-resolution space to obtain the denoised training samples;

[0026] Step 3: Use the visual attention mechanism and non-local feature enhanced convolution to extract and enhance the non-local feature information in the training samples for denoising, and combine a high-precision feature point detection method with non-linear optimization to eliminate the interference pixel points during CT image defect detection, thereby enhancing the image quality;

[0027] Step 4: Based on the Faster RCNN object detection model integrated in the MMDetection framework, select the ResNet50 residual network as the backbone feature extraction network of the model in the Backbone part, and add an SE attention mechanism module to enhance the effect of the network model in extracting defect features:

[0028] In the Neck part, a feature pyramid module is introduced to fuse the feature maps of different scales generated by the Backbone part, obtaining multi-scale feature maps;

[0029] In the Head part, taking the multi-scale feature maps obtained from the Neck part as input, using the Region Proposal Network (RPN) to generate anchor boxes and candidate boxes, realizing a rough localization of the defective targets. According to the confidence ranking, setting a threshold of 0.5 to perform Non-Maximum Suppression (NMS) once, removing the candidate boxes exceeding the threshold, and obtaining the final candidate boxes; then sending the candidate boxes to the RoI Pooling module for defective class division and coordinate regression;

[0030] Step 5: Train the network according to the above-built steps to obtain a trained model, and test the trained model with test samples;

[0031] The original image of the SRM combustion chamber ultra-high-resolution CT image to be judged is input into the trained and tested model after being processed through Steps 1 to 3 for prediction, and the class and position information of the grain defects are output.

[0032] The m*n is 640×640.

[0033] For the training samples of the cut-out several small images, image data augmentation is performed using data augmentation methods including but not limited to flipping, adjusting contrast, and random cropping.

[0034] The training samples and test samples are allocated according to 7:3.

[0035] Beneficial effects

[0036] An intelligent judgment method for defects in a rocket engine combustion chamber proposed by the present invention aims at defects such as the debonding of the interface between the grain and the shell and the grain bubbles caused by the difference in casting process during the charging process of the grain of a solid rocket motor (SRM). An intelligent judgment method for defects in a rocket engine combustion chamber is developed. This method, aiming at the CT image quality problem, effectively improves the image quality by using image enhancement and non-local deep features to find more similar features in non-local blocks. Aiming at the small sample problem of defects, geometric transformation methods such as mirroring, rotation, translation, distortion, filtering, and contrast adjustment are used to enhance the network's learning ability for small sample defects. To reduce the computational complexity of processing high-dimensional data, the visual attention mechanism and incremental learning method are used, which can not only make the network more focused on finding useful information more similar to the target features but also effectively cope with the diversity and generalization of defect samples, greatly improving the intelligent judgment efficiency and accuracy of grain defects.

[0037] The present invention has the following advantages compared with the existing typical defect detection methods for SRM combustion chambers:

[0038] 1. By introducing deep learning methods into the traditional defect detection method for SRM combustion chamber CT images, the present invention realizes the intelligent interpretation requirements for defect detection, and has obvious advantages in terms of the stability and controllability of defect detection;

[0039] 2. The intelligent analysis and evaluation technology for SRM combustion chamber defect detection proposed by the present invention can efficiently and accurately evaluate the defect problems and internal quality of SRM combustion chambers through intelligent defect interpretation technology.

[0040] 3. By applying artificial intelligence means, the present invention not only solves the problem that the manually designed defect features have low robustness to diverse changes, but also improves the limitation problem of low detection and segmentation accuracy of traditional image processing methods.

[0041] 4. The present invention is not only applicable to the defect problems of SRM combustion chambers, but can also be further extended to the defect detection requirements of other types of aerospace structural parts. The typical defect intelligent interpretation method for SRM combustion chambers is simple to deploy and has very important application value and engineering significance. Brief Description of the Drawings

[0042] Figure 1 : Flowchart of the present invention Detailed Embodiment

[0043] The present invention will be further described in combination with embodiments and drawings:

[0044] The present invention provides an intelligent interpretation technology for typical defects in CT images of rocket motor grains. First, by deeply analyzing the noise characteristics generated during the generation, transmission, and processing of CT images, the noise reduction processing technology for SRM combustion chamber CT images is studied; on the basis of traditional image filtering methods, a method for simultaneously removing mosaics and noise based on residual learning is particularly proposed, and the visual attention mechanism and non-local feature enhancement convolution are used to extract and enhance the non-local feature information in the images, realizing the quality improvement of SRM combustion chamber CT images. Secondly, data enhancement methods for typical defects such as interface debonding and pores in SRMs are studied; an algorithm for detecting pores and debonding based on the classic Faster-RCNN object detection model is developed; the optimization method of pre-training and fine-tuning is adopted to improve the network convergence speed and detection accuracy. The intelligent and accurate interpretation of typical defects in SRM combustion chamber CT images is realized. The present invention discloses an intelligent interpretation method for SRM combustion chamber defects based on CT images, mainly improving the interpretation accuracy and efficiency of typical defects (interface debonding and pores) in SRM combustion chambers.

[0045] Refer toFigure 1 , the implementation steps of the present invention are as follows:

[0046] Step 1: Cut the original image of the ultra-high-resolution CT image of the SRM combustion chamber into several small images with a width and height of 640×640, and divide them into training samples and test samples according to a ratio of 7:3, where the overlop ratio is set to 0.2 and the step size is 512;

[0047] For the training samples of the cut small images, image data augmentation is performed using data augmentation methods including but not limited to flipping, adjusting contrast, and random cropping.

[0048] The cutting method is as follows: Starting from the upper left corner of the original image, cut the image according to 640×640, record the upper left corner coordinates of the cut small images for use in label positioning; according to the cutting idea, obtain the coordinates of the upper left and lower right points of the cut small images in the original image through the sliding window method. Finally, judge whether the original image annotation box is inside the small image according to the coordinates, so as to determine the new coordinate position of the box relative to the small image, ensure that there are defective targets in the cut small images, and at the same time eliminate the pure background. In summary, the entire process of offline cutting is realized.

[0049] For typical defective targets in the ultra-high-resolution CT image of the SRM combustion chamber, if reshaped into small images and then fed into the network for training, the defective targets will become very small, making it difficult to identify. Directly training with large images will make it difficult to meet the GPU video memory requirements, and the original image will consume too much CPU time, resulting in a serious increase in training time and a slowdown in the inference speed.

[0050] Step 2: Use the method of simultaneously removing mosaic and noise based on residual learning for the training samples to perform end-to-end mapping learning between the noisy, low-resolution space and the clear, high-resolution space, and obtain the denoised training samples;

[0051] Step 3: To better utilize the non-local similarity feature information in the optical image to improve the image quality, a method of extracting and enhancing the non-local feature information in the training samples for denoising is proposed by combining the visual attention mechanism and non-local feature enhanced convolution, and using a high-precision feature point detection method combined with non-linear optimization to eliminate the interference pixel points during CT image defect detection, thereby enhancing the image quality;

[0052] Step 4: In the feature extraction stage, the conventional method is to input the segmented images into ResNet or VGGNet for feature extraction. Due to the problems of multi-scale and small targets in the defective images, methods such as using multi-layer fusion feature maps or increasing the resolution of the feature maps are used for optimization. The generated feature maps are input into the feature pyramid module for feature fusion, and feature maps with the same number of output channels are output.

[0053] Then it is sent into the generation network for foreground and background classification and bbox regression.

[0054] The HyperNet method is used to fuse the features of the deep, middle, and shallow feature networks, which improves the detection accuracy. At the same time, this method can make the feature details more abundant, facilitating the detection of small defective targets.

[0055] Specifically in the embodiment:

[0056] Based on the Faster RCNN object detection model integrated in the MMDetection framework, the ResNet50 residual network is selected as the backbone feature extraction network of the model in the Backbone part, and the SE attention mechanism module is added to enhance the effect of the network model in extracting defective features:

[0057] In the Neck part, a feature pyramid module is introduced to fuse the feature maps of different scales generated by the Backbone part to obtain multi-scale feature maps;

[0058] In the original Faster RCNN method, in order to avoid overlapping boxes when generating proposals in the RPN, the non-maximum suppression method is used to delete all boxes with an IoU greater than the threshold, which sometimes causes missed detections of occluded defective targets. Therefore, on the original method, instead of simply deleting the detection boxes with an IoU greater than the threshold, the confidence score is reduced, so as to reduce the missed detections of defects without increasing the computational complexity.

[0059] In the Head part, taking the multi-scale feature maps obtained in the Neck part as the input, the Region Proposal Network (RPN) is used to generate anchor boxes and candidate boxes to achieve a rough localization of the defective targets. According to the confidence ranking, a threshold of 0.5 is set to perform NMS once, and the candidate boxes exceeding the threshold are eliminated to obtain the final candidate boxes; then the candidate boxes are sent to the RoI Pooling module for defective category classification and coordinate regression;

[0060] Step 5: Train the network according to the above-built steps to obtain a trained model, and test the trained model with test samples;

[0061] The original image of the SRM combustion chamber ultra-high-resolution CT image to be judged is input into the trained and tested model after being processed through steps 1 to 3, and the category and position information of the grain defects are predicted and output.

[0062] Due to the class imbalance phenomenon existing in two types of typical defects (interface debonding, propellant bubbles) in CT images, mainly manifested as fewer debonding defects, therefore, sample enhancement and data resampling methods were adopted to increase the quantity of defect data, different weights were correspondingly added to different classes in the loss function, and Focal loss was used in the classification loss of the RCNN part in the original method, which alleviated the class imbalance phenomenon. To sum up, to meet the actual needs, a corresponding detection model was trained after a series of improvements to the original method, and then the trained model was used in the test link to detect the defect type and locate the defect target, so as to verify the effectiveness of the model.

[0063] The effects of the present invention can be further illustrated by the following experiments.

[0064] 1. Experimental conditions

[0065] The experiments of the present invention were carried out using the Python language on a central processing unit of Intel Xeon@Gold 5218R 2.1GHz CPU, with 128G of memory and an Ubuntu18.04 operating system.

[0066] 2. Experimental data

[0067] The data used in the experiments is classified data and is not publicly used.

[0068] 3. Experimental content

[0069] First of all, the two-stage detection network Faster-RCNN was selected as the Baseline for experiments. The recognition accuracy of the two-stage algorithm for targets is better than that of the single-stage detection method, and it is also verified through experiments that the network structure meets the real-time requirements. The optimization method of pre-training - fine-tuning was used to improve the network convergence speed and detection accuracy. To prove the effectiveness of the present invention, the two-stage detection method used in the present invention was compared with the single-stage detection YOLOv5 method, and the detection accuracy and real-time performance of this method are higher than those of the single-stage detection method. The comparison results are shown in Table 1.

[0070] Table 1

[0071]

[0072] As can be seen from Table 1, compared with the comparison method, the present invention can significantly improve the performance indicators of interface debonding and bubble defects, which also shows the effectiveness of the network model, thus proving the good practicability of the present invention for the detection of typical defects in the SRM combustion chamber of CT images.

Claims

1. An intelligent judgment method for defects in a rocket engine combustion chamber, characterized in that The steps are as follows: Step 1: Cut the original image of the ultra-high-resolution CT image of the SRM combustion chamber into several small images with a width and height of m*n, and divide them into training samples and test samples. The overlop ratio is set to 0.2, and the step size is 512; The image cutting method is: starting from the upper left corner of the original image, cut the image according to m*n, and record the upper left corner coordinates of the cut small images for use in label positioning; Step 2: Use the method of simultaneously removing mosaic and noise based on residual learning for the training samples, and perform end-to-end mapping learning between the noisy and low-resolution space and the clear and high-resolution space to obtain the denoised training samples; Step 3: Use the visual attention mechanism and non-local feature enhanced convolution to extract and enhance the non-local feature information in the training samples for denoising. In the way of combining the high-precision feature point detection method and non-linear optimization, eliminate the interference pixel points during CT image defect detection, so as to enhance the image quality; Step 4: Based on the Faster RCNN object detection model integrated in the MMDetection framework, select the ResNet50 residual network as the backbone feature extraction network of the model in the Backbone part, and add an SE attention mechanism module to enhance the effect of the network model in extracting defect features: In the Neck part, introduce a feature pyramid module to fuse the feature maps of different scales generated by the Backbone part to obtain multi-scale feature maps; In the Head part, using the multi-scale feature maps obtained from the Neck part as the input, adopt the Region Proposal Network (RPN) to generate anchor boxes and candidate boxes, and achieve a rough positioning of the defect targets. According to the confidence ranking, set the threshold to 0.5 and perform NMS once. Eliminate the candidate boxes exceeding the threshold to obtain the final candidate boxes; then send the candidate boxes to the RoIPooling module for defect category classification and coordinate regression; Step 5: Train the network according to the steps built above to obtain a trained model, and test the trained model with the test samples; The original image of the ultra-high-resolution CT image of the SRM combustion chamber to be interpreted is input into the trained and tested model after being processed by Steps 1 to 3 for prediction, and the category and position information of the grain defects are output.

2. The intelligent defect interpretation method for a rocket engine combustion chamber according to claim 1, wherein: The m*n is 640×640.

3. The intelligent defect interpretation method for a rocket engine combustion chamber according to claim 1, wherein: For the training samples of the cut small images, image data augmentation methods including but not limited to flipping, adjusting contrast, and random cropping are used for image data augmentation.

4. The intelligent defect interpretation method for a rocket engine combustion chamber according to claim 1, wherein: The training samples and test samples are allocated according to 7:3.

Images