Space antenna fault detection and identification method based on three-dimensional Gaussian sputtering

Abstract

The invention belongs to the technical field of space antenna fault detection and identification, and relates to a space antenna fault detection and identification method based on three-dimensional Gaussian sputtering, which comprises the following steps of: performing coarse positioning on the position of a space antenna hinge by using an optimized YOLOv5 model; according to the position of the space antenna hinge, tracking of the space antenna hinge and calibration of a viewpoint pose are achieved through continuous image streams; constructing a 3DGS model of an intact space antenna hinge by using a three-dimensional Gaussian sputtering 3DGS technology, performing defect detection on the tracked and calibrated space antenna hinge, and quantitatively analyzing the size of the defect of the space antenna hinge; according to the invention, monitoring blind areas and fault omission can be reduced; shooting points are accurately estimated, and antenna details are restored with high precision; time resources are saved, and the detection efficiency is improved; three-dimensional Gaussian sputtering is combined with a loss function, a fault area is accurately recognized, and misjudgment and missed judgment are reduced.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of space antenna fault detection and identification, and relates to a space antenna fault detection and identification method based on three-dimensional Gaussian sputtering. BACKGROUND

[0002] In the field of space antenna fault detection and identification, there are many problems in the prior art that need to be solved:

[0003] Insufficient monitoring comprehensiveness and real-time performance: Traditional detection methods mostly rely on a small number of sensors or manual inspection, which makes it impossible to discover subtle damage and structural deformation of the antenna in complex space environments in time, affecting the normal operation and service life of the antenna.

[0004] Inaccurate shooting point estimation: When determining the position and angle of the shooting device relative to the antenna, traditional methods cannot accurately calculate and estimate the pose, and it is also difficult to reasonably sample pixels and determine the line of sight direction. As a result, the antenna surface cannot be accurately rendered, making it difficult to accurately analyze the antenna based on accurate image data.

[0005] Difficulty in dividing key investigation areas: Due to the lack of accurate analysis of antenna surface features and problem probability evaluation methods, traditional methods cannot specifically investigate potential fault areas during detection, and a large amount of time is wasted on ineffective detection of normal areas, reducing fault detection efficiency.

[0006] Weak fault area self-recognition ability: In the prior art, when comparing ideal and actual antenna images to identify fault areas, manual comparison or simple algorithms are mostly used, which makes it difficult to automatically and accurately identify fault areas with significant differences, resulting in missed and false detections.

[0007] Inaccurate fault determination and parameter determination: For the identified fault areas, traditional methods lack accurate determination standards and algorithms to determine fault types. In the measurement and calculation of key parameters such as fault size and shape, due to the limited image data processing capability, detailed and accurate information cannot be provided, which brings great inconvenience to subsequent maintenance scheme development and processing work. SUMMARY

[0008] The purpose of the present application is to overcome the deficiencies in the prior art and provide a space antenna fault detection and identification method based on three-dimensional Gaussian sputtering.

[0009] In order to achieve the purpose of the present application, the following technical solutions are adopted.

[0010] A space antenna fault detection and identification method based on three-dimensional Gaussian sputtering, comprising the following steps:

[0011] S1, using the optimized YOLOv5 model to coarsely position the position of the space antenna hinge;

[0012] S2, according to the position of the space antenna hinge obtained in step S1, using a continuous image stream to realize tracking of the space antenna hinge and calibration of the viewpoint pose;

[0013] S3, using three-dimensional Gaussian sputtering 3DGS technology to construct a 3DGS model of the intact space antenna hinge, and detecting defects of the space antenna hinge tracked and calibrated in step S2, and quantitatively analyzing the size of the defects of the space antenna hinge.

[0014] As a preferred scheme of the application, the specific process of the coarse positioning is to input an image containing the space antenna hinge into the optimized YOLOv5 model, the optimized YOLOv5 model extracts image features through its backbone network, and then uses the attention mechanism module inserted into the backbone network to strengthen the extracted image features, so as to quickly identify the position of the space antenna hinge in the image.

[0015] As a preferred scheme of the application, the attention mechanism module is inserted before the SPPF module at the end of the backbone network of the YOLOv5 model.

[0016] As a preferred scheme of the application, the attention mechanism module is an SE module, an ECA module or a CB AM module.

[0017] As a preferred scheme of the application, the specific implementation process of step S2 includes the following steps:

[0018] S31, according to the position of the space antenna hinge obtained in step S1, visually analyzing and marking the target hinge;

[0019] S32, using neural radiation field technology to model a three-dimensional implicit representation of the target hinge;

[0020] S33, using the idea of the inverse neural radiation field algorithm to estimate the six-degree-of-freedom pose of the viewpoint under the continuous image stream.

[0021] As a preferred scheme of the application, the visual significance analysis uses a preset threshold as an evaluation standard, and at the same time, introduces the proportion of the significant area in the image as a judgment index for judgment;

[0022] The visual significance marking is to use the prior information of the size of the target hinge and the internal parameters of the camera to calculate the observation radius after completing the significance analysis, and mark the observation area range of the significant target hinge.

[0023] As a preferred scheme of the present application, the evaluation criterion is obtained by the common law, and the specific obtaining process is: setting a threshold value k, the probability of a pixel being classified as "background" is p b (k), the average gray value of the pixel assigned to the background is m b (k); the probability of a pixel being classified as "foreground" is p f (k), the average gray value of the pixel assigned to the foreground is m f (k); the cumulative average value corresponding to the threshold value k is m, and the total average gray value of the entire image is m G ; therefore:

[0024] m G =p b (k)m b (k)+p f (k)m f (k)

[0025] σ 2 =p b (k)(m b (k)-m G ) 2 +p f (k)(m f (k)-m G ) 2

[0026] By calculating the variance, the value k that minimizes the variance is the threshold τ in the Otsu method.

[0027] As a preferred scheme of the present application, the obtaining process of the judgment index is: the number of feature points in the salient region is detected by using the ORB algorithm, a pixel in the image is selected as a center pixel, M adjacent pixels uniformly distributed around the center pixel are selected, the gray values of the adjacent pixels and the center pixel are compared, if there are N continuous pixels whose gray values are higher or lower than that of the center pixel, the center pixel is identified as a feature point, and finally the response value of each feature point is calculated by non-maximum suppression NMS to eliminate repeated points and weak points.

[0028] As a preferred scheme of the present application, the calculation formula of the observation radius is:

[0029]

[0030] Wherein, A p represents the scale of the physical imaging plane, B p represents the scale of the pixel plane, and R represents the scale of the hinge target.

[0031] As a preferred scheme of the present application, the three-dimensional implicit representation modeling is expressed as:

[0032] Fθ =(x,d)→(c,σ);

[0033] wherein σ is the volume density and c is the color value of the target object.

[0034] As a preferred scheme of the present application, the estimation of the viewpoint six-degree-of-freedom pose is achieved by minimizing the photometric loss between the NeRF rendered image and the observed image.

[0035] As a preferred scheme of the present application, the obtaining process of the 3DGS model comprises the following steps:

[0036] S91, data acquisition: photographing the intact space antenna hinge from multiple different angles and positions to obtain a series of images containing complete space antenna hinge information;

[0037] S92, structure from motion (SFM): processing the multi-view images collected in step S91 using the structure from motion algorithm to calculate the camera pose and the 3D point cloud representation of the scene;

[0038] S93, conversion to Gaussian: converting each point in the point cloud data obtained in step S92 into a Gaussian distribution;

[0039] S94, model training: each discrete point in the point cloud calculated in step S93 is given the attribute of Gaussian distribution, so that it becomes a Gaussian splat with a certain shape, size, color and transparency; the random gradient descent method is used for optimizing the Gaussian parameters; the Gaussian is rendered into an image using the differentiable Gaussian rasterization, then the rendered image is compared with the original photographed image to calculate the loss function; the parameters of the Gaussian are adjusted according to the loss, and automatic densification and pruning operations are performed at the same time; after multiple iterations of training, the intact hinge 3DGS model is obtained.

[0040] As a preferred scheme of the present application, the specific process of defect detection comprises the following steps:

[0041] S101, obtaining an image to be detected: photographing the space antenna hinge to be detected to obtain an image containing the space antenna hinge to be detected;

[0042] S102, iNeRF pose estimation: using the iNeRF method to estimate the pose of the space antenna hinge in the obtained image to be detected;

[0043] S103, rendering the intact model with the same pose: according to the pose of the hinge to be detected estimated by iNeRF, the intact hinge 3DGS model generated before is rendered to obtain the rendered image of the intact hinge under the same pose;

[0044] S104, feature extraction and comparison: respectively extracting pre-training features from the to-be-detected image and the rendered image, comparing the extracted features, and calculating the mean square error between them;

[0045] S105, abnormal score map construction: constructing an abnormal score map according to the calculated mean square error;

[0046] S106, defect detection: by setting a suitable threshold, the abnormal score map is binarized, thereby realizing the detection and positioning of defects.

[0047] Beneficial effects

[0048] The present application can monitor comprehensively, real-time omnidirectional monitoring of antenna arrays, comprehensive and accurate data acquisition, reducing monitoring blind spots and fault omissions; accurate shooting point estimation, reverse neural radiation field processing data, optimized rendering results, high-precision restoration of antenna details, conducive to discovering subtle faults; efficient detection, algorithm analysis divides key exploration areas, avoids comprehensive detection, saves time resources, and improves detection efficiency; accurate fault area identification, three-dimensional Gaussian sputtering combined with loss function, automatic comparison of images, accurate identification of fault areas, and reduction of misjudgment and missed judgment; detailed repair basis: accurate fault type determination, accurate fault size and shape measurement, and help to develop efficient repair schemes; complete and efficient system: each step is closely linked to form a complete system, and the fault detection and identification performance is comprehensively improved. BRIEF DESCRIPTION OF DRAWINGS

[0049] Figure 1 is a flowchart of the present application;

[0050] Figure 2 is a camera observation target hinge model diagram for simplification;

[0051] Figure 3 is a NeRF network schematic diagram;

[0052] Figure 4 is a six-degree-of-freedom pose estimation flowchart under continuous image flow;

[0053] Figure 5 is a 3DGS reconstruction result diagram of a perfect hinge;

[0054] Figure 6 is an abnormal score map. DETAILED DESCRIPTION

[0055] The present application is further illustrated in combination with the embodiments and the drawings.

[0056] For monitoring comprehensiveness and real-time and key exploration division, the present application provides a YOLOv5 small target detection based on attention mechanism; for shooting point pose estimation, the present application provides a spatial antenna hinge tracking and view point pose calibration based on a reversed neural radiation field; for autonomous identification of fault area, fault determination and parameter determination, the present application provides a spatial antenna defect detection and size quantitative analysis based on three-dimensional Gaussian sputtering (3DGS).

[0057] YOLOv5: YOLOv5 (You Only Look Once version 5) is a target detection algorithm designed to quickly and accurately identify target objects in images or videos. YOLOv5 extracts features from input images through a backbone network, capturing information at different scales and levels in the image. Then, the extracted features are further fused and processed using a neck network to enhance the expression ability of the features. Finally, the head network directly predicts the class, location, and confidence of the target objects in the image based on the processed features. It realizes end-to-end target detection without the need for complex region proposal steps like traditional methods, and can guarantee detection accuracy while having high detection speed, widely used in security monitoring, autonomous driving, industrial detection and many other fields.

[0058] Neural Radiance Fields: Neural Radiance Fields (NeRF) is a deep learning technology for three-dimensional scene modeling and rendering. It aims to learn the latent representation of data and generate realistic three-dimensional scene images. NeRF uses a fully connected neural network to map three-dimensional spatial coordinates and viewing directions as input to the color and density information of the corresponding points. Using volume rendering technology, the integral of the light emitted from the camera in space is calculated to obtain the final pixel color, realizing high-quality representation and rendering of three-dimensional scenes. Its training process calculates the mean square error of the synthesized pixel color and the real image pixel color as the loss function, optimizing the neural network parameters to make the rendering result approach the real image. Compared with traditional three-dimensional representation methods such as point cloud, mesh and voxel, it only needs a set of two-dimensional images to construct a three-dimensional scene, which can effectively avoid the problems of high data storage and processing cost, distortion of texture mapping, difficulty in rendering complex lighting and material, etc. NeRF has the ability to synthesize high-quality new views, can generate images with high detail and realism, and retains delicate light changes; uses continuous functions to represent, can sample at any position in space, avoiding the discontinuity of discrete sampling.

[0059] Three-dimensional Gaussian Splatting (3DGS): Three-dimensional Gaussian Splatting (3DGS) is an advanced technique in the field of computer graphics for rendering and reconstructing three-dimensional scenes. It has attracted much attention. 3DGS constructs and renders three-dimensional scenes through millions of tiny translucent ellipsoids, known as "Gaussian splats". Unlike traditional methods that rely on polygonal meshes or complex neural networks, it reconstructs the scene based on these splat points. Each Gaussian splat carries information such as position, color, size, and transparency, which can highly restore the realism of the scene after combination. In terms of rendering, similar to photogrammetry techniques, the object is first captured from multiple angles using photos or video frames, the camera positions are determined and a preliminary three-dimensional point cloud of the scene is generated (structure from motion method). The points in the point cloud are converted into Gaussian splat points with specific attributes: position determines its orientation in three-dimensional space; shape and size control stretching or scaling to represent fine details; color stores color information and can change with viewing angle; transparency is used to control the mixing of multiple layers of splats. Subsequently, these splat points undergo an optimization process, including differentiable rasterization (projecting three-dimensional Gaussian splats onto a two-dimensional plane simulating camera perspective), measuring the difference between rendered images and original images using a loss function, fine-tuning splats to reduce differences with the help of optimization algorithms, and adaptive density control (removing unnecessary splats or increasing splats in detail areas), finally generating a three-dimensional image that highly matches the original scene for rendering. The advantages of 3DGS are significant, with high fidelity, efficiency, and speed. In terms of fidelity, it excels in capturing fine details and complex lighting effects such as reflection and refraction, achieving highly realistic rendering; in terms of efficiency, compared to polygonal meshes and large data neural networks, Gaussian splats are simple, requiring less storage and computing resources; in terms of speed, efficient data structures and optimized rendering processes enable real-time or near-real-time rendering, making it suitable for interactive applications and efficiently handling complex scenes with a large number of splat points.

[0060] The small target detection based on attention mechanism YOLOv5 aims to optimize the detection effect of YOLOv5 on small targets such as space antenna hinges, and coarsely locate the position of the space antenna hinge, facilitating subsequent fault detection work, which includes the following steps:

[0061] 1) The attention mechanism module is used to optimize the YOLOv5 backbone network;

[0062] 2) Determine the insertion position of the attention mechanism module in the backbone network;

[0063] 3) Compare the influence of different attention mechanism modules on the performance of the model after insertion;

[0064] 4) Based on the optimized model, the coarse positioning of the space antenna hinge is realized.

[0065] The space antenna hinge small target detection method based on the attention mechanism optimized YOLOv5 has the following main features:

[0066] (1) In step 1), several different types of attention mechanism modules are selected, including the attention module (SE module), ECA (Efficient Channel Attention) attention mechanism, and CBAM (Convolutional Block Attention Module). These are channel attention mechanisms or hybrid domain attention mechanisms. Taking the SE module as an example, it obtains the weights of different feature channels by strengthening important channels and weakening secondary channels. In the operation process, first, perform global average pooling (Global Average Pooling) on the feature map layer to obtain the global feature compression vector of the current feature map. This step is called the Squeeze operation. Then, through a two-layer fully connected Bottleneck structure, the weight of each channel in the feature map is obtained, and then the weight is normalized through the Sigmoid activation function. Finally, the normalized weight is applied to the original feature map through the Scale operation.

[0067] (2) In step 2), considering the different effects of the YOLOv5 backbone network after extracting features through different modules, the attention mechanism module is inserted at the end of the backbone network before the SPPF module. This is because after 4 C3 modules of feature extraction, the feature map can better replace the basic features of the original image. At this point, the attention mechanism is used on the feature map output by the last C3 module, which can better extract features related to small targets and improve the detection ability of small targets.

[0068] (3) In step 3), the performance of the model after inserting different attention mechanism modules is compared and analyzed. The SE module increases the computational complexity but improves the model accuracy. The ECA module realizes information interaction between channels based on the method of adaptively selecting convolution kernel size, which can significantly reduce the model complexity while maintaining performance. It performs one-dimensional convolution on the feature map after the pooling layer, and the convolution kernel size is determined by the channel dimension. The CBAM module is a hybrid domain attention mechanism that can perform attention operations in the channel and spatial dimensions, including two independent sub-modules: the channel attention module and the spatial attention module. It can save parameters and computing power and can be integrated into existing network architectures as a plug-and-play module. After inserting these several attention mechanism modules into the same position of the YOLOv5 backbone network, the accuracy, recall rate, and other indicators of different models in detecting space antenna hinge small targets are compared through experiments, and the influence of different modules on the performance of the model is analyzed.

[0069] (4) When step 4) is performed, the YOLOv5 model is optimized to coarsely locate the space antenna hinge. In actual application scenarios, an image containing the space antenna hinge is input into the optimized model. The model extracts image features through a backbone network, and in combination with the reinforcing effect of the inserted attention mechanism module on small target features, the approximate position of the hinge in the image can be quickly identified, providing convenience for subsequent loop shooting work and facilitating further accurate detection and analysis of the state of the hinge.

[0070] The space antenna hinge tracking and viewpoint pose calibration based on a continuous image stream is mainly achieved through a series of steps to track the space antenna hinge and accurately calibrate the viewpoint pose, and the specific process is as follows:

[0071] 1) Perform target visual saliency analysis and marking;

[0072] 2) Construct a three-dimensional implicit model of the hinge target;

[0073] 3) Estimate the six-degree-of-freedom pose of the viewpoint based on a continuous image stream.

[0074] The space antenna hinge tracking and viewpoint pose calibration based on a continuous image stream has the following main features:

[0075] When step 1) is performed, considering the challenge of continuous image stream frame-by-frame analysis to the onboard computing power of the aircraft, a saliency detection method based on time process is designed. The first detection is started within a predetermined time interval, and if no hinge target is detected, the time interval for the next detection is reduced. In each analysis process, a preset threshold is used as the evaluation standard, and the Otsu method and the ORB feature detection algorithm are highly coupled. The Otsu method is based on image gray characteristics to calculate the number of pixels at each gray level in the image. Let the threshold be k, the probability of pixels being classified as "background" be p b (k), the average gray level of pixels assigned to the background be m b (k); the probability of pixels being classified as "foreground" be p f (k), and the average gray level of pixels assigned to the foreground be m f (k). In addition, the cumulative mean value corresponding to the threshold is p G (k), and the total average gray level of the entire image is m G . Therefore:

[0076] m b =p b (k)m f (k)+p f (k)m 2 (k)

[0077] σ b(k)(m b (k)-m G ) 2 +p f (k)(m f (k)-m G ) 2

[0078] The value k that minimizes the variance is the threshold τ in the Otsu method by calculating the variance. At the same time, the proportion of the salient region in the image is introduced as a judgment index to avoid the decrease of positioning accuracy caused by long observation distance. The number of feature points in the salient region is detected by using the ORB algorithm, a pixel is selected as the center pixel in the image, a group of (usually 16) adjacent pixels uniformly distributed around it are selected, the gray values of adjacent pixels and the center pixel are compared, if there are N (usually set to 12) consecutive pixels whose gray values are higher or lower than that of the center pixel, the center pixel is identified as a feature point, and finally the response value of each feature point is calculated by non-maximum suppression (NMS) to eliminate repeated points and weak points. After completing the saliency judgment, the prior information of the hinge target scale and the camera intrinsic parameter are used to calculate the observation radius, and the salient target observation region range is delimited, the observation radius L ob The calculation method is as follows:

[0079]

[0080] Wherein, A p represents the physical imaging plane scale, B p represents the pixel plane scale, and R represents the scale of the hinge target. Simplify the camera observation target hinge model, as shown in Figure 2 .

[0081] (2) When performing step 2), a three-dimensional implicit representation model of the hinge target is established by using the neural radiation field (Neural Radiance Field, NeRF) technology. The NeRF technology represents a continuous scene as a five-dimensional vector-valued function, the input of which includes a three-dimensional position (x, y, z) and a two-dimensional observation direction The output includes the color value c=(r, g, b) of the target object and the volume density σ. In practical applications, the observation position is represented as a three-dimensional Cartesian unit vector d. NeRF uses an approximation technique to represent this continuous five-dimensional scene and optimizes the related weights, aiming to map each input five-dimensional coordinate to its corresponding volume density and color of the light emission direction. NeRF uses an eight-layer fully connected neural network F θ =(x, d)→(c, σ) to approximate the scene, and the structure of the eight-layer fully connected neural network is as shown in Figure 3 .

[0082] To predict the color value of each sampling pixel, NeRF samples along the emitted light ray from the camera center through the image plane. In order to reduce the computational cost caused by large-scale point sampling, NeRF adopts a coarse-to-fine hierarchical sampling method. Since only a few points on the light ray have an impact on the final rendering result, and most areas have no color, NeRF first uniformly samples n coarse points on the light ray, then finds the color distribution according to the transmittance of each sampling point, and selects the points n fine that need to be fine-sampled based on the volume density σ obtained by coarse sampling, and then estimates the color value C(r) of each pixel using the classic volume rendering technique:

[0083]

[0084] where p near and p far represent the boundary values of the sampling depth p along the light ray r, and T(r, p) represents the object occlusion rate, expressed as:

[0085]

[0086] Finally, the optimal color value is obtained by optimizing the loss function between the output value of volume rendering and the color true value. The loss function is as follows.

[0087]

[0088] (3) When step 3) is performed, the viewpoint six-degree-of-freedom pose estimation under the image continuous flow is realized based on the idea of inverse neural radiance field (iNeRF) algorithm. This method is mainly based on the idea of inverse neural radiance field (iNeRF) algorithm, as shown below. Figure 4 High-precision six-degree-of-freedom pose estimation is achieved by minimizing the photometric loss between the NeRF rendered image and the observed image. Assuming that the hinge already has a pre-trained NeRF three-dimensional model with weight Θ, and the camera internal parameters are known, iNeRF uses the observed image I and its corresponding camera pose T ∈ SE(3) to estimate the current pose

[0089]

[0090] where L is the L2 loss. During the entire optimization process, the weight Θ of NeRF remains unchanged, and the loss function is minimized by continuously adjusting T.

[0091] In the six-degree-of-freedom space of SE(3) manifold, the loss function is non-convex, and using it in the optimization loop leads to high computational cost for the whole NeRF rendering process. To alleviate this problem, a gradient-based optimization method is designed to handle the prior T of pose estimation. To ensure that the pose estimation T i The prior is kept within the SE(3) manifold, and the pose prior needs to be parameterized using exponential coordinates. Given the initial pose estimate prior T i It can be represented as:

[0092]

[0093] Where S=[ω,ν] T represents the screw axis, and represents the amplitude. With this parameterization, the goal is to determine the best relative transformation starting from the initial estimated pose T0. The gradient is obtained by iteratively differentiating the loss function with respect to MLP, and then the estimated relative transformation is updated using the gradient. The Adam optimizer is adopted, with exponentially decaying learning rates of β1=0.9 and β2=0.999, respectively. For each observed image, Sθ is initialized near 0, where each element is randomly drawn from a zero-mean normal distribution.

[0094] In a typical differentiable rendering and comparison process, each study aims to utilize the gradient of all output pixel contributions in the rendered image. However, in NeRF, the value of each output pixel is calculated by weighting the values of n sampling points along each ray r∈R during the ray marching process. Therefore, calculating and backpropagating the loss of all pixels in the image requires a large amount of GPU memory. Therefore, the present invention explores the strategy of selecting a set of sample rays R to evaluate the loss function at each optimization step. For example, random sampling, randomly sampling pixel positions on the image to calculate the corresponding rays, and after observing the target object, the calculation can be accelerated without sacrificing positioning accuracy; interest point sampling, first using an interest point detector to locate candidate pixel positions in the observed image, then sampling points from them, and if the detected interest points are insufficient, falling back to random sampling, but this strategy is prone to local minimum; interest region sampling, sampling within an expanded mask centered on the interest point, and by morphological dilation, the sampling region is expanded, which can speed up the optimization when the ray batch is small, and setting a larger number will fall back to random sampling.

[0095] The spatial antenna defect detection and size quantitative analysis based on 3D Gaussian sputtering aims to realize the defect detection and fault size quantitative analysis of the spatial antenna through a series of technical means, and the specific steps are as follows:

[0096] 1) Construct a perfect hinge 3DGS model using three-dimensional Gaussian sputtering (3DGS) technology;

[0097] 2) Based on the constructed model, the fault detection of the hinge to be detected is carried out, and the size quantitative analysis is realized.

[0098] The above-mentioned space antenna defect detection and size quantitative analysis method based on 3D Gaussian sputtering, its main features are:

[0099] When performing step 1), it needs to go through the sub-steps of data acquisition, motion recovery structure (SFM), conversion to Gaussian, and model training. In the data acquisition stage, the intact hinge ring is photographed from multiple different angles and positions to obtain a series of images containing complete hinge information. During the shooting process, the shooting environment should be stable and the light should be uniform, so as to ensure the image quality and provide a reliable data basis for subsequent accurate modeling. In the motion recovery structure (SFM) link, the motion recovery structure algorithm (such as Colmap library) is used to process the collected multi-view images. This algorithm finds out the same feature points in different images, and then calculates the camera pose and the rough 3D point cloud representation of the scene by using the geometric constraints between adjacent views. This process provides basic geometric information for constructing a 3D model. Each point in the point cloud data obtained by SFM is converted into a Gaussian distribution. At this time, only the position and color information can be inferred from the SFM data, and further training is needed to obtain a more accurate model. Each discrete point in the point cloud is given the attribute of Gaussian distribution, so that it becomes a Gaussian splat with certain shape, size, color and transparency, which preliminarily has more rich visual performance ability. The model training uses the stochastic gradient descent method to optimize the Gaussian parameters. The Gaussian is rendered into an image using differentiable Gaussian rasterization, and then the rendered image is compared with the original shooting image to calculate the loss function. According to the loss, the parameters of the Gaussian are adjusted, and automatic densification and pruning operations are performed at the same time. If the gradient of a certain Gaussian is large (indicating a large error), it is split or cloned; if the Gaussian is small, it is cloned; if it is large, it is split; if the opacity is too low, it is removed. After several iterations of training, an accurate perfect hinge 3DGS model is obtained.

[0100] In step 2), the main steps include image acquisition, iNeRF pose estimation, rendering the intact model at the same pose, feature extraction and comparison, and abnormal score map construction and defect detection, etc. to realize size quantitative analysis. The image containing the hinge to be detected is acquired by taking a photo of the hinge to be detected. The shooting conditions are as consistent as possible with those when the intact model is constructed to reduce errors. The iNeRF (Inverting Neural Radiance Fields for Pose Estimation) method is used to estimate the pose of the hinge in the acquired image to be detected. Based on the principle of neural radiation field (NeRF), the iNeRF optimizes the unknown camera pose by keeping the parameters of the trained NeRF unchanged, so that the rendered image and the image to be detected are optimally matched in luminosity, thereby obtaining the accurate pose of the hinge to be detected. According to the pose of the hinge to be detected estimated by iNeRF, the previously generated 3D GS model of the intact hinge is rendered, as shown in FIG. 8, to obtain the rendered image of the intact hinge at the same pose. The pre-trained features are extracted from the image to be detected and the rendered image, respectively, for example, using the pre-trained EfficientNet-b4 model. The extracted features are compared, and the mean squared error (MSE) between them is calculated. The mean squared error is an indicator to measure the similarity of two images, and the formula is as follows: Figure 5

[0101] The mean squared error is particularly sensitive to large errors. If there is a large difference in the image (such as a large area of noise), the MSE value will increase significantly, but it is not sensitive to the perceived difference in visual quality. According to the calculated mean squared error, an abnormal score map is constructed, as shown in FIG. 9. The area with a larger mean squared error indicates a larger difference from the intact model, which may have defects. By setting a suitable threshold, the abnormal score map is binarized to realize the detection and positioning of defects.

[0102] Figure 6

[0103] For the detected defect area, the size, shape, and other specific parameters of the defect are further determined by measuring and calculating the image data, providing detailed information for subsequent repair processing. In calculating the size and shape, professional image processing algorithms and measurement tools are used to ensure the accuracy of the results.

[0104] The preferred embodiments of the embodiments of the present application are described above with reference to the accompanying drawings, but this does not limit the scope of the rights of the embodiments of the present application. Any modifications, equivalent replacements and improvements made by those skilled in the art without departing from the scope and essence of the embodiments of the present application shall be within the scope of the rights of the embodiments of the present application.​​

Claims

1. A method for fault detection and identification of space antennas based on three-dimensional Gaussian sputtering, characterized in that: Includes the following steps: S1. Use the optimized YOLOv5 model to coarsely locate the position of the space antenna hinge. S2. Based on the position of the spatial antenna hinge obtained in step S1, the spatial antenna hinge is tracked and the viewpoint pose is calibrated using a continuous image stream. S3. Construct a complete 3DGS model of the space antenna hinge using 3D Gaussian sputtering 3DGS technology, perform defect detection on the space antenna hinge tracked and calibrated in step S2, and quantitatively analyze the size of the defects in the space antenna hinge.

2. The method for fault detection and identification of space antennas based on three-dimensional Gaussian sputtering according to claim 1, characterized in that: The specific process of coarse localization involves inputting an image containing the space antenna hinge into an optimized YOLOv5 model. The optimized YOLOv5 model extracts image features through its backbone network, and then uses an attention mechanism module inserted into the backbone network to enhance the extracted image features, enabling rapid identification of the space antenna hinge's position in the image; wherein: The attention mechanism module is inserted before the SPPF module at the end of the YOLOv5 model backbone network; The attention mechanism module is an SE module, an ECA module, or a CBAM module.

3. The method for fault detection and identification of space antennas based on three-dimensional Gaussian sputtering according to claim 1, characterized in that: The specific implementation process of step S2 includes the following steps: S31. Based on the position of the space antenna hinge obtained in step S1, perform visual saliency assessment and marking on the target hinge. S32. The target hinge is modeled in three dimensions using neural radiation field technology. S33. The idea of ​​the reverse neural radiation field algorithm is used to estimate the six-degree-of-freedom pose of the viewpoint in a continuous image stream.

4. The method for fault detection and identification of space antennas based on three-dimensional Gaussian sputtering according to claim 3, characterized in that: The visual saliency assessment uses a preset threshold as the evaluation standard, and at the same time, introduces the proportion of salient regions in the image as an assessment indicator. The visual saliency marking is performed after the saliency assessment is completed. The observation radius is calculated using prior information about the target hinge scale and camera internal parameters, and the observation area of ​​the saliency target hinge is delineated and marked.

5. The method for fault detection and identification of space antennas based on three-dimensional Gaussian sputtering according to claim 4, characterized in that: The evaluation criteria are obtained from the general law, specifically through the following process: Let the threshold be k, and the probability of a pixel being classified as "background" be p. b (k), the average gray level of the pixels assigned to the background is m. b (k); the probability that a pixel is classified as "foreground" is p. f (k), the average gray level of the pixels assigned to the foreground is m. f (k); the cumulative mean value corresponding to threshold k is m, and the total average gray level of the entire image is m. G ;therefore: m G =p b (k)m b (k)+p f (k)m f (k) σ 2 =p b (k)(m b (k)-m G ) 2 +p f (k)(m f (k)-m G ) 2 The value k that minimizes the variance is the threshold τ in Otsu's method. Meanwhile, the ORB algorithm is used to obtain the evaluation index: the ORB algorithm is used to detect the number of feature points in the salient region. A pixel is selected from the image as the center pixel, and M neighboring pixels are selected evenly distributed around it. The gray values ​​of the neighboring pixels are compared with those of the center pixel. If there are N consecutive pixels with gray values ​​higher or lower than the gray value of the center pixel, the center pixel is identified as a feature point. Finally, the response value of each feature point is calculated by non-maximum suppression (NMS) to eliminate duplicate points and weak points.

6. The method for fault detection and identification of space antennas based on three-dimensional Gaussian sputtering according to claim 4, characterized in that: The formula for calculating the observation radius is as follows: Among them, A p B represents the physical imaging plane scale. p R represents the pixel plane scale, and R represents the scale of the hinge target.

7. The method for fault detection and identification of space antennas based on three-dimensional Gaussian sputtering according to claim 3, characterized in that: The three-dimensional implicit representation modeling is expressed as follows: F θ =(x,d)→(c,σ); In the formula, σ is the volume density and c is the color value of the target object.

8. The method for fault detection and identification of space antennas based on three-dimensional Gaussian sputtering according to claim 3, characterized in that: The estimation of the six-DOF pose of the viewpoint is achieved by minimizing the photometric loss between the NeRF rendered image and the observed image.

9. The method for fault detection and identification of space antennas based on three-dimensional Gaussian sputtering according to claim 1, characterized in that: The process of obtaining the 3DGS model includes the following steps: S91. Data Acquisition: Take pictures of the intact space antenna hinge from multiple different angles and positions to obtain a series of images containing complete information about the space antenna hinge; S92, Structure of Motion (SFM): The Structure of Motion algorithm is used to process the multi-view images acquired in step S91 to calculate the camera pose and the 3D point cloud representation of the scene. S93. Convert to Gaussian: Convert each point in the point cloud data obtained in step S92 into a Gaussian distribution; S94. Model Training: Each discrete point in the point cloud calculated in step S93 is assigned the properties of a Gaussian distribution, making it a Gaussian splat with a certain shape, size, color and transparency. The model training uses stochastic gradient descent to optimize the Gaussian parameters; differentiable Gaussian rasterization is used to render the Gaussian image, and then the rendered image is compared with the original captured image to calculate the loss function; the Gaussian parameters are adjusted according to the loss, and automatic densification and pruning operations are performed simultaneously; after multiple iterations of training, a complete hinge 3DGS model is obtained.

10. The method for fault detection and identification of space antennas based on three-dimensional Gaussian sputtering according to claim 1, characterized in that: The specific process of defect detection includes the following steps: S101. Image acquisition: Take a picture of the space antenna hinge to be detected and acquire an image containing the space antenna hinge to be detected. S102, iNeRF Pose Estimation: The iNeRF method is used to estimate the pose of the spatial antenna hinge in the acquired image to be detected; S103. Rendering a intact model with the same pose: Based on the pose of the hinge to be detected estimated by iNeRF, render the previously generated intact hinge 3DGS model to obtain the rendered image of the intact hinge with the same pose. S104. Feature Extraction and Comparison: Extract pre-trained features from the image to be detected and the rendered image respectively, compare the extracted features, and calculate the mean square error between them. S105, Anomaly Fraction Plot Construction: Construct an anomaly fraction plot based on the calculated mean squared error; S106. Defect Detection: By setting an appropriate threshold, the anomaly score map is binarized to achieve defect detection and location.

Images