Performance Prediction Methods and Related Equipment for Space High-Energy Beam Fused Additive Components

By constructing a multimodal database and a lightweight neural network model, the problem of non-destructive, unmanned autonomous inspection and evaluation of high-energy beam filament additive components in the microgravity environment of space was solved, realizing real-time and quantitative quality assessment in orbit and reducing inspection costs and cycles.

CN123118566APending Publication Date: 2026-11-17TECH & ENG CENT FOR SPACE UTILIZATION CHINESE ACAD OF SCI
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202611369865.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-04
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies cannot achieve efficient, unmanned, and autonomous quality inspection and performance evaluation of high-energy beam fused wire additive components in the microgravity and high vacuum environment of space. In particular, they cannot identify internal micron-level hidden defects and predict mechanical properties on orbit, and rely on ground inspection or destructive testing, resulting in high costs and long inspection cycles.

Method used

A multimodal database was constructed, a multi-task neural network model was trained and compressed into a lightweight model, and the internal defects and mechanical properties were predicted simultaneously by acquiring surface morphology images in orbit. Lightweight equipment was used for non-destructive evaluation.

Benefits of technology

It enables efficient quantitative quality assessment in orbit, in real time, and without human intervention, avoiding ground-based re-inspection and destructive testing, reducing launch costs, and is suitable for unmanned autonomous space missions.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123118566A_ABST
    Figure CN123118566A_ABST
Patent Text Reader

Abstract

This invention discloses a method and related equipment for predicting the performance of high-energy beam fused wire additive components in space. It relates to the fields of on-orbit high-energy beam fused wire additive manufacturing, aerospace machine vision non-destructive testing, and intelligent cross-scale performance evaluation of aerospace metal components. The method includes: pre-constructing a multimodal database and training a multi-task neural network model based on the database; then compressing the trained multi-task neural network model into a lightweight model; acquiring surface morphology images of the target high-energy beam fused wire additive component in orbit; inputting the surface morphology images of the target high-energy beam fused wire additive component into the lightweight model to obtain the internal defect quantification parameters and mechanical performance prediction indicators of the target high-energy beam fused wire additive component. This invention can solve the load constraint problem of not being able to carry large-scale mechanical tests and heavy non-destructive testing equipment in space, and eliminate the shortcomings of on-orbit quality assessment relying on human experience.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the fields of on-orbit high-energy beam fused wire additive manufacturing in space, non-destructive testing by aerospace machine vision, and intelligent evaluation of the cross-scale performance of aerospace metal components, and particularly to a method and related equipment for predicting the performance of high-energy beam fused wire additive components in space. Background Technology

[0002] High-energy beam filament additive manufacturing under microgravity vacuum is a disruptive core technology in the aerospace field. Compared with ground-based machining and the launch of prefabricated components into orbit, on-orbit filament additive manufacturing has three irreplaceable advantages: it significantly reduces the volume and weight of rocket launch payloads, allowing for the on-orbit fabrication of arbitrarily complex aerospace structural components based solely on the launch of raw metal wire; it enables in-situ repair of damaged supports, cracked pipes, and failed connections in space stations and deep space probes, eliminating the need for replacing entire modules or equipment; and it supports the long-term unmanned autonomous manufacturing needs of future lunar and Martian research bases, eliminating the need to stockpile massive amounts of spare parts in advance. Currently, ground-based simulated microgravity filament printing experiments and on-orbit verification of small electron beam filament prototypes for space stations have been deployed and completed both domestically and internationally, and the forming process system is gradually maturing. However, the capabilities for quality inspection and performance evaluation lag far behind the development of manufacturing processes, which will become a key bottleneck restricting the engineering implementation of the technology.

[0003] There has been relevant research both domestically and internationally on quality inspection and control in the additive manufacturing process. Invention patent CN117773341A discloses a machine vision-sensing laser filament additive manufacturing forming control method and system. This method uses a charge-coupled device (CCD) camera coaxially mounted with a laser head to acquire images of the front of the molten pool. A convolutional neural network model is built to process these images. During the laser filament additive manufacturing process, the distance from the intersection of the laser filaments to the surface of the deposited layer is calculated in real time. The closed-loop controller adjusts the wire feeding speed based on the deviation between the calculated and set distance values ​​and the trend of this deviation, thereby controlling the amount of metal wire filling on the depositing path and achieving online control of the deposited layer height. Invention patent CN115796026A discloses a machine learning-based active control method for the shape and properties of laser additive manufacturing. This method utilizes finite element method simulation to generate temperature field distribution data, phase field method to simulate and generate microstructure images perpendicular to the printing direction, crystal plasticity theory to calculate the mechanical properties corresponding to the microstructure, and a convolutional neural network-long short-term memory neural network model to predict the forming quality at the next moment. When the predicted forming quality does not meet the requirements, real-time active control of the macroscopic temperature field is achieved by actively adjusting the printing process parameters, thereby regulating the microstructure and mechanical properties of the forming part online. All of the above methods are designed for terrestrial additive manufacturing scenarios, achieving online monitoring and process control of the forming process through visual sensing or physical model simulation.

[0004] However, the aforementioned existing technologies and on-orbit quality control methods all have insurmountable engineering flaws. Existing on-orbit quality control methods fall into three categories: The first category involves carrying components back to Earth for testing after manufacturing. On Earth, industrial computed tomography (CT) scans can be used to accurately obtain internal defects and universal testing machines to assess mechanical properties. However, the cost of round-trip transportation between Earth and space is extremely high, the testing cycle is difficult to coordinate quickly with the return capsule launch window, return channels for deep space exploration missions such as those to the Moon and Mars are scarce, and the component's on-orbit overload and extreme temperature changes during the round-trip process may alter its original metallurgical state, leading to distorted testing data. The second category involves carrying small, lightweight mechanical testing and flaw detection equipment on-orbit. Theoretically, this can achieve in-situ non-destructive and mechanical testing. However, tensile and fatigue testing machines include servo drives and high-precision force sensors, exceeding the cabin's load limits in terms of overall weight and power consumption. Mechanical testing is destructive testing, and the clamping and debugging process is complex and requires manual operation by astronauts, making it impossible to execute autonomously in unmanned deep space exploration scenarios. The third category involves using only simple industrial cameras in orbit for visual observation. Existing in-orbit fuse prototypes, equipped with monocular charge-coupled device cameras, can only qualitatively determine the presence of macroscopic cracks and large pores. They fail to establish a quantitative correlation between visible features such as melt flow texture and surface ripples and internal porosity and under-fusion defects. They cannot quantitatively predict core service indicators such as tensile strength and fatigue life, and many internal micron-level hidden defects cannot be identified. The evaluation results rely on the subjective experience of astronauts and have poor consistency. The methods described in the two patent documents mentioned above are both aimed at ground additive manufacturing processes, relying on real-time monitoring of temperature fields and process parameters for online control. They require continuous data acquisition and feedback control during the printing process, and both require the deployment of additional sensing equipment or complex physical simulation calculations. They cannot be directly applied to the extreme space conditions of space microgravity and high vacuum environments where payload weight, volume, and power consumption are strictly limited, unmanned autonomous operation, and rapid assembly and service of components. Summary of the Invention

[0005] This invention aims to address the shortcomings of existing technologies, specifically providing a method and related equipment for predicting the performance of high-energy beam fused filament additive components in space, as detailed below: 1) In a first aspect, the present invention provides a method for predicting the performance of space high-energy beam fused wire additive components, the specific technical solution of which is as follows: A multimodal database is pre-built that associates surface morphology images, internal defect parameters and mechanical performance indicators of space high-energy beam fused wire additive components. A multi-task neural network model is trained based on the database, and then the trained multi-task neural network model is compressed into a lightweight model. Surface morphology images of the target space high-energy beam fused wire additive component were acquired in orbit; By inputting the surface morphology image of the target space high-energy beam fused wire additive component into the lightweight model, the internal defect quantification parameters and mechanical performance prediction indicators of the target space high-energy beam fused wire additive component are obtained.

[0006] The beneficial effects of the performance prediction method for space high-energy beam fused wire additive components provided by this invention are as follows: By pre-constructing a multimodal database linking surface morphology images, internal defect parameters, and mechanical performance indicators, and training a multi-task neural network model, which is then compressed into a lightweight model, the system can simultaneously output quantitative parameters of internal defects and predicted mechanical performance indicators simply by acquiring surface morphology images of the target component and inputting them into this lightweight model. This completely eliminates the reliance on large mechanical testing machines and heavy non-destructive testing equipment, eliminating the need to return the component to Earth for re-inspection. Quantitative quality assessment can be completed in orbit in real time, effectively solving the detection dilemma in deep space exploration scenarios without a return route. It also eliminates the technical limitations of manual visual observation, which can only make qualitative judgments and cannot identify hidden internal defects. At the same time, it significantly shortens the detection cycle, avoids component scrapping caused by destructive testing, and significantly reduces the cost of round-trip transportation between Earth and space. This provides a feasible lightweight intelligent assessment method for unmanned autonomous operation and rapid installation and service of components in the microgravity and high vacuum environment of space.

[0007] 2) In a second aspect, the present invention also provides a performance prediction system for space high-energy beam fused wire additive components, the specific technical solution of which is as follows: It includes a ground-based offline modeling and database construction module, an on-orbit lightweight image acquisition module, and an AI intelligent prediction module; The ground-based offline modeling and database building module is used to: pre-build a multimodal database that associates surface morphology images, internal defect parameters and mechanical performance indicators of space high-energy beam fused wire additive components, train a multi-task neural network model based on the database, and then compress the trained multi-task neural network model into a lightweight model; The on-orbit lightweight image acquisition module is used to: acquire on-orbit images of the surface morphology of the target space high-energy beam fused wire additive component; The AI ​​intelligent prediction module is used to input the surface morphology image of the target space high-energy beam fused wire additive component into the lightweight model to obtain the internal defect quantification parameters and mechanical performance prediction indicators of the target space high-energy beam fused wire additive component.

[0008] 3) In a third aspect, the present invention also provides an electronic device, the electronic device including a processor coupled to a memory, the memory storing at least one computer program, the at least one computer program being loaded and executed by the processor, so as to enable the electronic device to implement any of the above-mentioned space high-energy beam filament additive component performance prediction methods.

[0009] 4) In a fourth aspect, the present invention also provides a computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, implements any of the above-mentioned methods for predicting the performance of space high-energy beam filament additive components.

[0010] It should be noted that the beneficial effects of the technical solutions of the second to fourth aspects of the present invention and their corresponding possible implementations can be found in the above description of the technical effects of the first aspect and its corresponding possible implementations, and will not be repeated here. Attached Figure Description

[0011] Figure 1 This is a flowchart illustrating a method for predicting the performance of a space high-energy beam fused wire additive component according to an embodiment of the present invention. Figure 2 This is a schematic diagram of a performance prediction system for a space high-energy beam filament additive component according to an embodiment of the present invention. Detailed Implementation

[0012] like Figure 1 As shown, an embodiment of the present invention provides a method for predicting the performance of a space high-energy beam filament additive component, comprising the following steps: S1. A multimodal database is pre-constructed that associates surface morphology images, internal defect parameters and mechanical performance indicators of space high-energy beam fused wire additive components. A multi-task neural network model is trained based on the database, and then the trained multi-task neural network model is compressed into a lightweight model. The process involves pre-constructing a multimodal database that correlates surface morphology images, internal defect parameters, and mechanical performance indicators of space high-energy beam fused wire additive components. A multi-task neural network model is then trained based on this database. The specific implementation process is as follows: S10. Fabrication of space high-energy beam fused wire additive components with multiple materials and process gradients, covering three types of commonly used aerospace structural wires: titanium alloy wire, stainless steel wire, and aluminum alloy wire; and two types of space high-energy beam heat sources: electron beam fused wire heat source and fiber laser fused wire heat source. During the fabrication process, key process parameters such as fused wire feed speed, beam current power, scanning rate, and ambient temperature were adjusted in a gradient manner. Multiple parallel samples were prepared for each set of process parameters, including standard tensile samples, standard fatigue samples, and block-shaped components.

[0013] S11. For each completed sample, three types of binding data are simultaneously collected. The first type of data is surface morphology images. A high-resolution industrial camera is used to capture the complete melt channel morphology of the top and side surfaces of the sample, obtaining raw images containing visual features such as melt channel width, ripple undulation, surface spatter distribution, interlayer overlap continuity, and surface microcracks. The second type of data is internal defect parameters. An industrial computed tomography (CT) scanner is used to perform layer-by-layer tomographic scanning of the sample, quantitatively labeling defects such as internal pore diameter, under-fusion area, interlayer microcrack length, and defect volume fraction. The third type of data is mechanical performance indicators. A universal testing machine is used to test the tensile strength, yield strength, and maximum elongation of the sample; a high-frequency fatigue testing machine is used to obtain the fatigue cycle life of the sample; and an impact testing machine is used to obtain the fracture toughness value of the sample. These measured data are used as ground truth labels for supervised training.

[0014] S12. For each sample, the surface morphology image, internal defect parameters, and mechanical performance indicators are bound and matched to construct an annotated multimodal database. The image data in the database is normalized to unify image size and grayscale standards, and data augmentation is performed to expand the sample size.

[0015] S13. Train a multi-task neural network model based on the constructed multimodal database. This multi-task neural network model adopts a hard parameter sharing architecture, including a shared backbone network and two output branches, specifically: The shared backbone network is responsible for extracting general visual features from the input surface topography image. The backbone network uses a ResNet-50 residual network as its basic architecture. A convolutional layer with a kernel size of 7 pixels × 7 pixels and a stride of 2 pixels is placed at the front of the network. This layer is followed by a batch normalization layer and a linear rectified activation function (RCA). After the convolutional layer, four residual blocks are stacked sequentially, containing 3, 4, 6, and 3 residual units respectively. Each residual unit contains a convolutional layer, a batch normalization layer, an RCA, and a residual connection structure. After processing by the four residual blocks, a global average pooling layer is applied to convert the feature map into a fixed-length feature vector.

[0016] After sharing the backbone network, the network structure is divided into two task-specific output branches. ① The first output branch is the internal defect prediction branch, used to perform regression prediction. This branch receives the feature vector output from the shared backbone network as input, and passes it through two fully connected layers. The first fully connected layer reduces the feature vector dimension to 128 dimensions, and the second fully connected layer further reduces the 128-dimensional feature vector to 32 dimensions. Each fully connected layer is followed by a wired rectified activation function. Finally, an output layer is set, containing multiple neurons, corresponding to the predicted values ​​of internal pore size, under-fusion area, interlayer microcrack length, and defect volume fraction, respectively. ② The second output branch is the mechanical property prediction branch, also used to perform regression prediction. This branch receives the feature vector output from the shared backbone network as input, and passes it through two fully connected layers for feature dimensionality reduction. The first fully connected layer reduces the feature vector dimension to 128 dimensions, and the second fully connected layer reduces the 128-dimensional feature vector to 32 dimensions. Each fully connected layer is followed by a wired rectified activation function. Finally, an output layer is set up, which contains multiple neurons, corresponding to the output of predicted tensile strength, predicted yield strength, predicted maximum elongation, predicted fatigue cycle life, and predicted fracture toughness.

[0017] During training, the total loss function of the multi-task neural network model is obtained by weighted summation of the losses from the internal defect prediction task and the mechanical performance prediction task. Both prediction tasks use the Huber loss function as the loss metric for the regression task. The multi-task neural network model is then subjected to end-to-end supervised training using the constructed multimodal database until convergence, resulting in the trained multi-task neural network model.

[0018] The process of compressing the trained multi-task neural network model into a lightweight model includes: using knowledge distillation, model pruning, and quantization compression to compress the trained multi-task neural network model into a lightweight model. The specific implementation process is as follows: S14. The knowledge distillation method uses a complex, parameter-intensive pre-trained multi-task neural network model as the teacher network. This teacher network performs forward propagation calculations on each surface topography image in the multimodal database, yielding its output. This output is a probability distribution or regression value of softening. The regression value includes predicted values ​​from the internal defect prediction branch (internal pore size, under-fusion area, interlaminar microcrack length, and defect volume fraction) and from the mechanical property prediction branch (tensile strength, yield strength, maximum elongation, fatigue cycle life, and fracture toughness). Based on this, a compact, parameter-efficient student network is constructed. The student network receives the same surface topography images as the teacher network. During training, the loss function consists of two weighted parts: the first part is the error between the student network output and the actual labeled values ​​in the multimodal database, and the second part is the error between the student network output and the teacher network output. The weight parameters of the student network are optimized using a backpropagation algorithm, enabling it to approximate the output behavior of the teacher network while maintaining a small number of parameters and low computational complexity. After training, the teacher network was discarded, and only the student network was retained as the output model for the knowledge distillation method.

[0019] S15. Model pruning uses the student network output via knowledge distillation as the input model, removing redundant connections and channels. Redundant connections are determined by calculating the absolute value of each connection weight, sorting all weights in ascending order, setting a global pruning threshold, and removing connections with weights below this threshold. Redundant channels are determined by calculating the sum of the absolute values ​​of all weights within each output channel of a convolutional layer as its importance score, sorting these scores in ascending order, setting a channel pruning threshold, and removing channels with scores below this threshold. After removing redundant connections and channels, the number of non-zero connections in the network is reduced, decreasing storage size. After model pruning, a multimodal database is used to fine-tune the pruned network to recover the model's prediction accuracy loss caused by pruning. Fine-tuning uses the same loss function as the original training, updating only the weight parameters of the remaining connections in the network; pruned connections and channels are not restored. After fine-tuning the training, a lightweight network model with pruning is obtained.

[0020] S16. The quantization compression method uses the pruned lightweight network model output by the model pruning method as the input model. It converts the floating-point format weight parameters stored in the input model into low-bit-width integer format weight parameters. The floating-point format weight parameters are network weight values ​​represented using 32-bit or 64-bit binary floating-point numbers. The specific operation of quantization compression is as follows: The numerical distribution range of the weight parameters in each layer of the input model is statistically analyzed, the maximum and minimum absolute values ​​of the weight parameters in that layer are determined, and the continuous floating-point value range of the weight parameters is discretized into evenly spaced integer values ​​according to the preset quantization bit width. The quantization bit width is set to 8 bits, that is, the value of each layer of weight parameters is mapped to an integer between 0 and 255. For each layer of weight parameters in the input model, the floating-point values ​​of all weight parameters in that layer are traversed, and each floating-point value is converted into a corresponding 8-bit integer value according to a linear mapping relationship. After quantization conversion, all weight parameters of the input model are stored in 8-bit integer format, and the model file size is reduced to one-quarter of the original floating-point storage size. After quantization and compression are completed, a small number of validation samples from the multimodal database are used to calibrate the quantized model. The error in the model output before and after quantization is minimized by adjusting the scaling factor and zero-point offset of the quantization mapping. After quantization calibration, the final lightweight model is obtained and stored in the storage unit of the onboard embedded chip for subsequent use in the on-orbit inference phase.

[0021] S2. On-orbit acquisition of surface morphology images of the target space high-energy beam fused wire additive component. The specific implementation process is as follows: S20. After the on-orbit fused wire additive manufacturing process is completed, the spacecraft control system initiates the surface topography image acquisition process. The spacecraft control system has pre-stored the coordinate data of the shooting points, including the vertical shooting point directly above the target component and the horizontal shooting point to the side of the target component.

[0022] S21. The spacecraft control system drives the aerospace-grade radiation-hardened charge-coupled device (CCD) camera along a preset trajectory to a vertical shooting position directly above it. After the camera stabilizes, the camera shutter is triggered to capture an image of the top surface morphology of the target component. Subsequently, the spacecraft control system drives the camera to a horizontal shooting position to the side, triggering the camera shutter to capture an image of the side surface morphology of the target component. A narrowband filter is installed at the front of the camera lens. The narrowband filter is used to filter out residual strong light from the high-energy beam fuse equipment and interference light of specific wavelengths from stray light in the space environment, preventing local overexposure of the image.

[0023] S22. The acquired raw surface topography images are transmitted to the onboard image preprocessing unit via a data link. The onboard image preprocessing unit performs real-time noise reduction on the raw images. The noise reduction process uses a median filtering algorithm to remove salt-and-pepper noise introduced by space radiation and the vacuum environment, and a mean filtering algorithm to smooth the slight blurring caused by spacecraft micro-vibrations. After noise reduction, the image size is uniformly scaled to a preset pixel width and preset pixel height, and the grayscale value of each pixel in the image is linearly normalized and mapped to an integer grayscale range of 0 to 255, so that all acquired surface topography images have a uniform size and a uniform grayscale standard. The surface topography images after the above preprocessing are stored in the onboard data storage unit, waiting to be called in subsequent steps.

[0024] Narrowband filters are optical filtering elements that allow light within a specific narrow wavelength range to pass through while blocking light from other wavelength ranges. Narrowband filters achieve wavelength selection through interference or absorption effects, and their full width at half maximum (FWHM) in the transmission band is typically less than 10 nanometers. Median filtering is a nonlinear image filtering method that replaces the gray value of each pixel in an image with the median of the gray values ​​of all pixels within its neighborhood window, effectively removing isolated bright or dark spots while maintaining image edge sharpness. Mean filtering is a linear image filtering method that replaces the gray value of each pixel in an image with the arithmetic mean of the gray values ​​of all pixels within its neighborhood window, smoothing gray-level fluctuations caused by random perturbations and reducing image noise. The onboard image preprocessing unit is a digital signal processing circuit on the spacecraft specifically designed for preliminary processing of image data. It has built-in image processing algorithm firmware with fixed functions, enabling it to autonomously perform operations such as noise reduction, scaling, and gray-level normalization without ground intervention.

[0025] S3. Input the surface morphology image of the target space high-energy beam fused wire additive component into the lightweight model to obtain the internal defect quantification parameters and mechanical property prediction indicators of the target space high-energy beam fused wire additive component. The specific implementation process is as follows: S30. Read the preprocessed surface topography image from the onboard data storage unit. The surface topography image includes the top surface image and the side surface image of the target component. The image size and grayscale standard are consistent with the images in the multimodal database used when training the lightweight model. Both are preset pixel width and preset pixel height, and the grayscale range is 0 to 255.

[0026] S31. The surface topography image is input into a lightweight model deployed in a spaceborne embedded chip. This lightweight model is pre-stored in the read-only memory of the embedded chip before in-orbit flight. During inference, the chip's computing unit loads the model parameters and performs forward propagation calculations. The input image first passes through the shared backbone network of the lightweight model, then sequentially through convolutional layers, batch normalization layers, linear rectified activation functions, and four residual blocks. Finally, it is converted into a fixed-length feature vector through a global average pooling layer. The feature vector contains multi-dimensional topography features extracted from the input image, including melt channel uniformity, surface spatter distribution density, interlayer overlap discontinuity, surface microcrack length, melt channel ripple amplitude, and oxide spot distribution density.

[0027] S32. The feature vector is passed in parallel to the internal defect prediction branch (first output branch) and the mechanical property prediction branch (second output branch) in the lightweight model. In the internal defect prediction branch, the feature vector is sequentially passed through two fully connected layers for feature dimensionality reduction, and finally, the neurons in the output layer generate predicted values ​​for five internal defect quantification parameters, including predicted values ​​for internal pore size, under-fusion area, interlaminar microcrack length, defect volume fraction, and defect risk level. In the mechanical property prediction branch, the feature vector is also sequentially passed through two fully connected layers for feature dimensionality reduction, and finally, the neurons in the output layer generate predicted values ​​for five mechanical property indicators, including predicted values ​​for tensile strength, yield strength, maximum elongation, fatigue cycle life, and fracture toughness. The internal defect quantification parameters and mechanical property prediction indicators, as the final outputs of the lightweight model, are stored in the onboard data storage unit for use in subsequent steps.

[0028] Forward propagation refers to the computational process of passing input data from the input layer of a neural network model, through the computation of each network unit in turn, layer by layer, until the output layer, ultimately yielding the model's output. During forward propagation, each layer performs linear transformations and nonlinear activation operations on the input data.

[0029] Among them, spaceborne embedded chips refer to dedicated integrated circuit chips carried on spacecraft. They have built-in central processing unit (CPU) cores and digital signal processing (DSP) cores, and can independently execute pre-programmed neural network inference programs without relying on ground control commands. Spaceborne embedded chips adopt radiation-hardened designs to withstand the ionizing radiation environment of space, and their power consumption and storage capacity meet the spacecraft payload resource constraints.

[0030] Among them, multi-dimensional morphological features refer to a set of visual descriptive parameters quantized and extracted from surface morphology images, used to characterize appearance attributes such as the geometry of the melt channel, surface texture, defect traces, and spatter distribution during the fused wire additive manufacturing process. These features are automatically learned through convolutional neural network layers, eliminating the need for manually designed feature extraction operators.

[0031] Optionally, the above technical solution also includes: S4. Compare the internal defect quantification parameters and mechanical performance prediction indicators of the target space high-energy beam filament additive component with the design thresholds in the aerospace structural component design threshold database, and determine whether the target space high-energy beam filament additive component is suitable for in-orbit service based on the comparison results.

[0032] The aerospace structural component design threshold database stores the allowable lower limit values ​​of mechanical properties, the maximum allowable defect size values, and the maximum allowable defect volume fraction values ​​for different aerospace structural components. Step S4 specifically includes: The predicted mechanical performance indicators are compared with the lower limit of mechanical performance, and the quantified internal defect parameters are compared with the maximum permissible defect size value and the maximum permissible defect volume fraction value. When all predicted mechanical performance indicators are higher than the lower limit of mechanical performance, and the quantified internal defect parameters are lower than both the maximum permissible defect size value and the maximum permissible defect volume fraction value, the target space high-energy beam fused additive component is determined to be suitable for in-orbit service. When the predicted mechanical performance indicators are lower than the lower limit of mechanical performance, and the quantified internal defect parameters are lower than both the maximum permissible defect size value and the maximum permissible defect volume fraction value, the target space high-energy beam fused additive component needs to be partially repaired and retested. When the quantified internal defect parameters are higher than the maximum permissible defect size value or higher than the maximum permissible defect volume fraction value, the target space high-energy beam fused additive component is determined to be scrapped. The specific implementation process is as follows: S40. Read the internal defect quantification parameters and mechanical property prediction indicators output in step S3 from the onboard data storage unit. The internal defect quantification parameters include predicted values ​​for internal pore diameter, under-fusion area, interlaminar microcrack length, defect volume fraction, and defect risk level. The mechanical property prediction indicators include predicted values ​​for tensile strength, yield strength, maximum elongation, fatigue cycle life, and fracture toughness. Simultaneously, read the allowable lower limit value of mechanical properties, maximum allowable defect size value, and maximum allowable defect volume fraction value corresponding to the type of aerospace structural component to which the current target component belongs from the aerospace structural component design threshold database.

[0033] S41. Compare the predicted values ​​of tensile strength, yield strength, maximum elongation, fatigue cycle life, and fracture toughness with the corresponding lower limits of the allowable mechanical properties one by one to determine whether all predicted mechanical property indicators are higher than the allowable lower limits. Simultaneously, compare the predicted values ​​of internal pore diameter, under-fusion area, and interlaminar microcrack length with the maximum allowable defect scale value to determine whether all the predicted defect scale values ​​are simultaneously lower than the maximum allowable defect scale value; compare the predicted defect volume fraction with the maximum allowable defect volume fraction value to determine whether the predicted defect volume fraction is lower than the maximum allowable defect volume fraction value.

[0034] When all predicted mechanical properties are higher than the lower limit of mechanical properties, and the predicted values ​​of internal pore diameter, under-fusion area, and interlayer microcrack length are simultaneously lower than the maximum allowable defect size, and the predicted defect volume fraction is lower than the maximum allowable defect volume fraction, the target space high-energy beam fused wire additive component is determined to be suitable for in-orbit service. The system generates an instruction allowing in-orbit installation, which, along with the predicted data, is stored in the onboard data storage unit.

[0035] When any one of the predicted values ​​of tensile strength, yield strength, maximum elongation, fatigue cycle life, or fracture toughness is lower than the corresponding lower limit of the allowable lower limit of mechanical properties, and the predicted values ​​of internal pore diameter, under-fusion area, and interlaminar microcrack length are simultaneously lower than the maximum allowable defect size, and the predicted defect volume fraction is lower than the maximum allowable defect volume fraction, the target space high-energy beam fused wire additive component is determined to require local repair and retesting. The system automatically outputs a defect spatial positioning command, which includes the estimated three-dimensional coordinate position of the defect in the component corresponding to the internal defect quantification parameters, triggering the on-orbit fused wire repair welding equipment to perform local fused wire repair welding at the coordinate position. After the repair is completed, steps S2 and S3 are re-executed to obtain new internal defect quantification parameters and new mechanical property prediction indicators for the repaired component, and the comparison and judgment process in this step is executed again.

[0036] When the predicted value of the internal pore diameter, the predicted value of the under-fusion area, the predicted value of the interlayer microcrack length, or the predicted value of the defect volume fraction exceeds the maximum allowable defect size, the target space high-energy beam fused wire additive component is deemed scrapped. The system generates a component scrapping command, triggering the spacecraft control system to initiate the component reprinting process, and re-execute high-energy beam fused wire additive manufacturing according to the original three-dimensional component model.

[0037] The minimum allowable mechanical performance limits refer to the lowest mechanical performance parameters specified for a specific type of aerospace structural component to ensure its structural integrity and load-bearing capacity during its on-orbit service. These minimum limits include lower limits for tensile strength, yield strength, maximum elongation, fatigue cycle life, and fracture toughness. These values ​​are pre-calculated based on the design load spectrum and safety factor of the aerospace structural component. The maximum allowable defect size refers to the upper limits allowed for internal pore diameter, under-fusion area, and interlaminar microcrack length for a specific type of aerospace structural component to prevent failure caused by localized defects during its on-orbit service. When the defect size exceeds these upper limits, the stress concentration at the defect tip will exceed the material tolerance. The maximum allowable defect volume fraction refers to the upper limit of the volume percentage of all defects within a unit volume of material for a specific type of aerospace structural component. When the defect volume fraction exceeds this upper limit, the effective load-bearing cross-sectional area of ​​the material is significantly reduced, and the overall structural strength drops below the design safety boundary. Post-repair re-inspection refers to the process of using on-orbit high-energy beam welding equipment to precisely fill the defective areas of components identified as having minor defects or slight mechanical performance deficiencies. This eliminates defects such as internal porosity, incomplete fusion, or microcracks. Subsequently, the surface morphology image of the repaired component is re-acquired, and the prediction and judgment process is repeated. Aerospace structural components refer to metal load-bearing parts used to transfer loads and maintain the spacecraft's configuration, such as space station module supports, load-bearing pipelines, connecting structures, and equipment mounting bases. Different types of aerospace structural components are assigned different allowable lower limits for mechanical properties, maximum allowable defect size values, and maximum allowable defect volume fraction values ​​during the design phase.

[0038] In another embodiment, the specific process for determining whether a target space high-energy beam filament additive component is suitable for on-orbit service is as follows: 1) The internal defect quantification parameters of the target space high-energy beam fused wire additive component are read from the onboard data storage unit. These parameters include predicted values ​​for internal pore diameter, under-fusion area, interlaminar microcrack length, and defect volume fraction. Based on these parameters, a virtual defect spatial distribution map of the target component is constructed. This map records the spatial coordinates and equivalent geometric dimensions of each type of defect within the component in the form of a three-dimensional tensor. Specifically: A spatial rectangular coordinate system is established with the geometric center of the target component's 3D model as the origin. The internal space of the component is discretized into a uniformly spaced voxel mesh with a side length of 0.1 mm. For each type of defect, a defect identifier value is filled into the voxel mesh at the corresponding spatial coordinates based on the defect's equivalent geometric dimensions. A defect identifier value of 1 represents a porosity defect, a defect identifier value of 2 represents an under-fusion defect, a defect identifier value of 3 represents an interlaminar microcrack defect, and voxel meshes at defect-free locations are filled with a defect identifier value of 0. For each defect, its equivalent geometric dimensions are simplified according to the following rules: porosity defects are simplified to spheres, with the sphere diameter equal to the predicted internal pore diameter; under-fusion defects are simplified to flattened ellipsoids, with the major axis length of the ellipsoid equal to the square root of the predicted under-fusion area; and interlaminar microcrack defects are simplified to flattened elliptical plates, with the major axis length of the flattened elliptical plate equal to the predicted interlaminar microcrack length. The spatial coordinates and equivalent geometric dimensions of each defect are written into the corresponding voxel mesh to form a three-dimensional tensor. The first dimension of this three-dimensional tensor corresponds to the voxel index in the X-axis direction of the spatial rectangular coordinate system, the second dimension corresponds to the voxel index in the Y-axis direction of the spatial rectangular coordinate system, and the third dimension corresponds to the voxel index in the Z-axis direction of the spatial rectangular coordinate system. Each voxel position stores a defect identifier value and the corresponding equivalent geometric dimension value, thereby completing the construction of the virtual defect spatial distribution map.

[0039] 2) The virtual defect spatial distribution map is input into a defect evolution prediction model pre-trained on the ground. This model is established based on the relationship between the initial defect morphology and the corresponding fatigue crack propagation rate and creep cavity growth rate from historical ground test data. The model outputs the post-defect morphological parameters of the target component under a preset on-orbit service duration. These parameters include the equivalent crack length after fatigue propagation and the equivalent cavity diameter after creep growth. Specifically: Fatigue crack propagation and creep test data of multiple sets of standard titanium alloy, stainless steel, and aluminum alloy specimens were collected. Each set of test data included initial defect morphology parameters, corresponding fatigue cyclic load amplitudes, corresponding test temperatures, and corresponding crack propagation rates or void growth rates. A recurrent neural network was used as the basic model architecture, with initial defect morphology parameters, fatigue cyclic load amplitudes, and test temperatures as input features, and crack propagation rates and void growth rates as output labels for supervised training until the model converged. After inputting a virtual defect spatial distribution map into the defect evolution prediction model, the model performed evolution calculations on defects at each voxel location in the map. For interlaminar microcrack defects, the fatigue crack propagation was calculated based on the estimated alternating stress amplitude at its spatial location and the on-orbit temperature conditions; for porosity defects, the creep void growth was calculated based on the estimated average stress at its spatial location and the on-orbit temperature conditions. After traversing all voxel meshes, the defect evolution prediction model outputs the morphological parameters of the target component after defect evolution under a preset on-orbit service duration. The morphological parameters after defect evolution include the equivalent length of the crack after fatigue propagation and the equivalent diameter of the cavity after creep growth.

[0040] 3) Calculate the first growth factor between the equivalent crack length in the defect evolution morphological parameters and the predicted value of the interlaminar microcrack length in the internal defect quantification parameters; calculate the second growth factor between the equivalent cavity diameter in the defect evolution morphological parameters and the predicted value of the internal pore diameter in the internal defect quantification parameters; correlate the first and second growth factors with the predicted values ​​of fatigue cycle life and fracture toughness in the mechanical performance prediction indicators, respectively, to obtain the remaining bearing capacity coefficient and remaining fracture resistance coefficient of the target component at the end of the preset on-orbit service period. Specifically: The first growth factor is obtained by dividing the equivalent length of the fatigue-propagated crack by the predicted length of the interlaminar microcrack. The second growth factor is obtained by dividing the equivalent diameter of the cavities after creep growth by the predicted internal pore size. The specific method of correlation calculation is as follows: the reciprocal of the predicted fatigue cycle life is set as the fatigue damage accumulation rate factor, and the first growth factor is multiplied by the fatigue damage accumulation rate factor to obtain the fatigue damage contribution value; the reciprocal of the predicted fracture toughness is set as the fracture sensitivity factor, and the second growth factor is multiplied by the fracture sensitivity factor to obtain the creep damage contribution value; the remaining bearing capacity coefficient is obtained by subtracting the weighted sum of the fatigue damage contribution value and the creep damage contribution value from the value of 1, where the weight of the fatigue damage contribution value is 0.6 and the weight of the creep damage contribution value is 0.4; the remaining fracture resistance coefficient is obtained by subtracting the weighted sum of the creep damage contribution value and the fatigue damage contribution value from the value of 1, where the weight of the creep damage contribution value is 0.7 and the weight of the fatigue damage contribution value is 0.3.

[0041] 4) Compare the remaining load-bearing capacity coefficient and the remaining fracture resistance coefficient with the preset load-bearing capacity retention threshold and the preset fracture resistance retention threshold, respectively. When the remaining load-bearing capacity coefficient is higher than the preset load-bearing capacity retention threshold and the remaining fracture resistance coefficient is higher than the preset fracture resistance retention threshold, the target space high-energy beam fused wire additive component is determined to be suitable for on-orbit service. When the remaining load-bearing capacity coefficient is higher than the preset load-bearing capacity retention threshold but the remaining fracture resistance coefficient is lower than the preset fracture resistance retention threshold, the target component is determined to need to be repaired by local welding and then re-evaluated. When the remaining load-bearing capacity coefficient is lower than the preset load-bearing capacity retention threshold, the target component is determined to be scrapped and the reprinting process is initiated.

[0042] In the above embodiments, although the steps are numbered S1, S2, etc., they are only specific embodiments given by the present invention. Those skilled in the art can adjust the execution order of S1, S2, etc. according to the actual situation. The scheme after adjusting the order is also within the protection scope of the present invention. It can be understood that in some embodiments, some or all of the above embodiments may be included.

[0043] like Figure 2 As shown, an embodiment of the present invention provides a performance prediction system for high-energy beam filament additive components in space, including a ground offline modeling and library construction module, an on-orbit lightweight image acquisition module, and an AI intelligent prediction module. The ground-based offline modeling and database building module is used to: pre-build a multimodal database that associates surface morphology images, internal defect parameters and mechanical performance indicators of space high-energy beam fused wire additive components, train a multi-task neural network model based on the database, and then compress the trained multi-task neural network model into a lightweight model; The on-orbit lightweight image acquisition module is used to: acquire on-orbit images of the surface morphology of the target space high-energy beam fused wire additive component; The AI ​​intelligent prediction module is used to input the surface morphology image of the target space high-energy beam fused wire additive component into the lightweight model to obtain the internal defect quantification parameters and mechanical performance prediction indicators of the target space high-energy beam fused wire additive component.

[0044] Optionally, the above technical solution also includes a performance output and on-orbit service safety assessment module. This module has a built-in aerospace structural component design threshold database. It is used to: compare the quantitative parameters of internal defects and predicted mechanical properties of the target space high-energy beam fused wire additive component with the design thresholds in the aerospace structural component design threshold database, and determine whether the target space high-energy beam fused wire additive component is suitable for on-orbit service based on the comparison results. By comparing the quantitative parameters of internal defects and predicted mechanical properties with the design thresholds in the aerospace structural component design threshold database, and determining whether the target space high-energy beam fused wire additive component is suitable for on-orbit service based on the comparison results, a clear conclusion on whether the component can be directly installed in orbit can be automatically output. This eliminates the uncertainty of manual experience assessment, makes the quality assessment results objective and repeatable, meets the high-reliability quality control standards for manned spaceflight, and provides a quantitative basis for subsequent on-orbit repair or remanufacturing decisions.

[0045] Optionally, in the above technical solution, the aerospace structural component design threshold database stores the allowable lower limit value of mechanical properties, the maximum allowable defect size value, and the maximum allowable defect volume fraction value corresponding to different aerospace structural components; The performance output and on-orbit service safety assessment module is specifically used to: compare the predicted mechanical performance indicators with the allowable lower limit of mechanical performance, and compare the internal defect quantification parameters with the maximum allowable defect size value and the maximum allowable defect volume fraction value; when all the predicted mechanical performance indicators are higher than the allowable lower limit of mechanical performance, and the internal defect quantification parameters are lower than the maximum allowable defect size value and the maximum allowable defect volume fraction value, the target space high-energy beam fused additive component is determined to be suitable for on-orbit service; when the predicted mechanical performance indicators are lower than the allowable lower limit of mechanical performance, and the internal defect quantification parameters are lower than the maximum allowable defect size value and the maximum allowable defect volume fraction value, the target space high-energy beam fused additive component needs to be partially repaired and retested; when the internal defect quantification parameters are higher than the maximum allowable defect size value or higher than the maximum allowable defect volume fraction value, the target space high-energy beam fused additive component is determined to be scrapped. By comparing the predicted mechanical performance indicators with the allowable lower limit of mechanical performance, and comparing the quantitative parameters of internal defects with the maximum allowable defect size value and the maximum allowable defect volume fraction value, a three-level automatic judgment suitable for in-orbit service, retesting after partial repair, and scrapping is performed based on the three comparison results. This enables precise classification and disposal of each target component, avoiding the scrapping of qualified components and the misjudgment of components with excessive defects into service. It provides a clear decision-making basis for in-orbit repair and remanufacturing, ensuring the quality and safety reliability of components in in-orbit service.

[0046] Optionally, in the above technical solution, the ground offline modeling and library building module is specifically used to: compress the trained multi-task neural network model into a lightweight model using knowledge distillation, model pruning, and quantization compression methods.

[0047] Continuing with the description of the system of the present invention, an embodiment of the present invention provides a performance prediction system for high-energy beam fused filament additive components in space, comprising a ground-based offline modeling and library construction module, an on-orbit lightweight image acquisition module, a morphological feature extraction and artificial intelligence intelligent prediction module, and a performance output and on-orbit service safety determination module, along with a complete ground pre-training process and an on-orbit real-time inference process. The overall system architecture consists of three parts: a ground-based offline system, an on-orbit space subsystem, and a wireless transmission link for the lightweight model on the satellite.

[0048] The ground-based offline system comprises a multi-wire sample preparation unit, a high-energy beam fused wire printing platform, an industrial computed tomography (CT) scanner, a universal testing machine, a high-resolution ground-based visual acquisition camera, a database server, and an artificial intelligence model training workstation. The space-to-ground transmission link transmits lightweight compressed neural network model packages to the spacecraft via an onboard wireless transmission channel. The space-based on-orbit subsystem comprises an on-orbit fused wire additive manufacturing equipment, an aerospace-grade radiation-resistant charge-coupled device (CCD) camera, an onboard local data storage unit, an onboard embedded chip, and a performance output and on-orbit safety assessment output terminal. The overall data flow logic is as follows: ground-based sample acquisition, sample annotation, and multimodal database construction are completed; then, multi-task neural network model training and lightweight compression are performed; the lightweight compressed neural network model package is uploaded to the spacecraft; after on-orbit printing is completed, surface images of the components are automatically acquired; the onboard embedded chip performs local offline analysis of the surface images, synchronously outputting defect levels, mechanical performance indicators, and service safety margins; and finally, outputting a determination command regarding whether the component can be put into service in orbit.

[0049] The ground-based offline modeling and database construction module is used to pre-build a multimodal database that correlates surface morphology images, internal defect parameters, and mechanical performance indicators of space-based high-energy beam fused wire additive components. Based on this database, a multi-task neural network model is trained, and then the trained multi-task neural network model is compressed into a lightweight model. All the work of the ground-based offline modeling and database construction module is completed in advance on the ground; on-orbit, only the lightweight analysis model is called, without consuming onboard computing power for training.

[0050] The specific process of pre-constructing a multimodal database in the ground offline modeling and library building module is as follows: First, prepare space high-energy beam fused wire additive component samples with multiple materials and process gradients, covering three types of commonly used aerospace structural wires, namely titanium alloy wire, stainless steel wire and aluminum alloy wire, covering two types of space high-energy beam heat sources, namely electron beam fused wire heat source and fiber laser fused wire heat source. Gradiently adjust the core process parameters such as fused wire feed speed, beam current power, scanning rate and ambient temperature, and prepare multiple parallel samples for each set of process parameters, including standard tensile samples, standard fatigue samples and block-shaped components. Second, three types of binding data are simultaneously collected for each prepared sample: The first type of data is surface morphology images, which are captured by a high-resolution industrial camera to obtain the complete melt channel morphology of the top and side surfaces of the sample, including visual features such as melt channel width, ripple undulation, surface splash distribution, interlayer overlap continuity, and surface microcracks; The second type of data is internal defect parameters, which are obtained by performing layer-by-layer tomographic scanning of the sample using an industrial computed tomography (CT) scanner, and quantitatively labeling defects such as internal pore diameter, under-fusion area, interlayer microcrack length, and defect volume fraction; The third type of data is mechanical performance indicators, which are obtained by testing the tensile strength, yield strength, and maximum elongation of the sample using a universal testing machine, obtaining the fatigue cycle life of the sample using a high-frequency fatigue testing machine, and obtaining the fracture toughness value of the sample using an impact testing machine. These measured data are used as ground truth labels for supervised training. Third, the surface morphology image, internal defect parameters and mechanical performance indicators of each sample are bound and matched to construct an annotated multimodal database. The image data in the database is normalized to unify the image size and grayscale standard, and data augmentation is performed to expand the sample size.

[0051] The offline ground modeling and database construction module trains a multi-task neural network model based on the constructed multimodal database. This multi-task neural network model adopts a hard parameter sharing architecture, consisting of a shared backbone network and two task-specific output branches. The shared backbone network is responsible for extracting general visual features from the input surface topography image. The backbone network uses a ResNet-50 residual network as its basic architecture. At the very beginning of the network, a convolutional layer with a kernel size of 7 pixels by 7 pixels and a stride of 2 pixels is placed. This layer is followed by a batch normalization layer and a linear rectified activation function. After the convolutional layer, four residual blocks are stacked sequentially, containing 3, 4, 6, and 3 residual units respectively. Each residual unit contains a convolutional layer, a batch normalization layer, a linear rectified activation function, and a residual connection structure. After processing by the four residual blocks, a global average pooling layer is applied to convert the feature map into a fixed-length feature vector. After the shared backbone network, the network structure is divided into two task-specific output branches. The first output branch is the internal defect prediction branch, which is used to perform regression prediction tasks. It receives the feature vector output by the shared backbone network as input and passes through two fully connected layers. The first fully connected layer reduces the feature vector dimension to 128 dimensions, and the second fully connected layer further reduces the 128-dimensional feature vector to 32 dimensions. After each fully connected layer, a linear rectified activation function is set. Finally, an output layer is set, which contains multiple neurons and outputs the predicted values ​​of internal pore size, under-fusion area, interlayer microcrack length, and defect volume fraction, respectively. The second output branch is the mechanical property prediction branch, also used to perform regression prediction tasks. It receives the feature vector output from the shared backbone network as input, and sequentially passes it through two fully connected layers for feature dimensionality reduction. The first fully connected layer reduces the feature vector dimension to 128 dimensions, and the second fully connected layer reduces the 128-dimensional features to 32 dimensions. Each fully connected layer is followed by a linear rectified activation function. Finally, an output layer is set, which contains multiple neurons, corresponding to the predicted values ​​of tensile strength, yield strength, maximum elongation, fatigue cycle life, and fracture toughness, respectively. During training, the total loss function of the multi-task neural network model is obtained by weighted summation of the loss from the internal defect prediction task and the mechanical property prediction task. Both prediction tasks use the Huber loss function as the loss metric for the regression task. The multi-modal database is used to perform end-to-end supervised training on the above multi-task neural network model until the model converges, resulting in the trained multi-task neural network model.

[0052] The ground offline modeling and database construction module employs knowledge distillation, model pruning, and quantization compression to compress the trained multi-task neural network model into a lightweight model. The knowledge distillation method uses a complex, parameter-intensive pre-trained multi-task neural network model as the teacher network. This teacher network performs forward propagation calculations on each surface topography image in the multimodal database, yielding its output. This output consists of regression values, including predicted values ​​for internal pore size, under-fusion area, interlaminar microcrack length, and defect volume fraction from the internal defect prediction branch, and predicted values ​​for tensile strength, yield strength, maximum elongation, fatigue cycle life, and fracture toughness from the mechanical property prediction branch. Based on these, a... The student network, with its compact structure and few parameters, receives the same surface topography image as input as the teacher network. During training, the loss function consists of two weighted parts: the first part is the error between the student network output and the real labeled data in the multimodal database, and the second part is the error between the student network output and the teacher network output. The weight parameters of the student network are optimized through backpropagation, enabling the student network to approximately fit the output behavior of the teacher network while maintaining a small number of parameters and low computational complexity. After training, the teacher network is discarded, and only the student network is retained as the output model for knowledge distillation. The model pruning method uses the student network output by knowledge distillation as the input model. Redundant connections or channels in the input model are removed. Redundant connections are determined by calculating the absolute value of each connection weight, sorting all weights in ascending order, and setting a global pruning threshold. Connections with weights below the threshold are considered redundant and removed from the network. Redundant channels are determined by calculating the sum of the absolute values ​​of all weights in each output channel of the convolutional layer as the channel's importance score, sorting the importance scores in ascending order, and setting a channel pruning threshold. Channels with importance scores below the threshold are considered redundant and removed from the network. After removing redundant connections and channels, the number of non-zero connections in the network is reduced, and the storage volume is decreased. After model pruning, a multimodal database is used to fine-tune the pruned network to recover the model prediction accuracy loss caused by the pruning operation. Fine-tuning uses the same loss function as the original training, only updating the weight parameters of the remaining connections in the network. Pruned connections and channels are not restored. After fine-tuning, a lightweight pruned network model is obtained.The quantization compression method uses the pruned lightweight network model output by the model pruning method as the input model. It converts the floating-point format weight parameters stored in the input model into low-bit-width integer format weight parameters. The floating-point format weight parameters are network weight values ​​represented using 32-bit or 64-bit binary floating-point numbers. The specific operation of quantization compression involves statistically analyzing the numerical distribution range of the weight parameters in each layer of the input model, determining the maximum and minimum absolute values ​​of the weight parameters in that layer, and discretizing the continuous floating-point value range of the weight parameters into evenly spaced integer values ​​according to a preset quantization bit depth. The quantization bit depth is set to 8 bits, that is, mapping the value of each layer's weight parameter to an integer between 0 and 255. For each layer of weight parameters in the input model, the floating-point values ​​of all weight parameters in that layer are traversed, and each floating-point value is converted into a corresponding 8-bit integer value according to a linear mapping relationship. After quantization conversion, all weight parameters of the input model are stored in 8-bit integer format, reducing the model file size to one-quarter of the original floating-point storage size. After the quantization compression is completed, a small number of verification samples from the multimodal database are used to quantize and calibrate the quantized model. The error of the model output before and after quantization is minimized by adjusting the scaling factor and zero offset of the quantization mapping. After quantization calibration, the final lightweight model is obtained. The lightweight model is stored in the storage unit of the spaceborne embedded chip for subsequent on-orbit inference.

[0053] The on-orbit lightweight image acquisition module is used to acquire surface morphology images of the target space high-energy beam fused-wire additive component in orbit. The aerospace-grade radiation-hardened charge-coupled device (CCD) camera serves as the sensing device, requiring no auxiliary flaw detection or mechanical support equipment. It is equipped with a narrow-band filter to eliminate interference from the strong light of the high-energy beam, a low-power lens, and a lightweight mounting bracket. The total weight and power consumption of the entire acquisition hardware system meet the weight and power constraints of the spacecraft payload. After the on-orbit fused-wire additive component is formed, the spacecraft control system initiates the surface morphology image acquisition process. The spacecraft control system pre-stores the coordinate data of the shooting points, including a vertical shooting point directly above the target component and a horizontal shooting point to the side of the target component. The spacecraft control system drives the aerospace-grade radiation-hardened CCD camera along a preset trajectory to the vertical shooting point directly above. After the camera stabilizes, the camera shutter is triggered, acquiring an image of the top surface morphology of the target component. Subsequently, the spacecraft control system drives the camera to the horizontal shooting point to the side, triggering the camera shutter to acquire an image of the side surface morphology of the target component. A narrowband filter is installed at the front of the camera lens. This filter removes residual strong light from the high-energy beam filament equipment and interference from specific wavelengths of stray light in the space environment, preventing localized overexposure in the image. After the filament additive manufacturing single-layer printing is completed or the entire component is formed, the spacecraft control system automatically drives the camera to a fixed shooting position directly above and to the side of the component to acquire complete surface morphology images from multiple angles. The acquired raw surface morphology images are transmitted to the onboard image preprocessing unit via a data link. The onboard image preprocessing unit performs real-time noise reduction on the raw images. The noise reduction process uses a median filtering algorithm to remove salt-and-pepper noise introduced by space radiation and the vacuum environment, and a mean filtering algorithm to smooth the slight blurring caused by spacecraft micro-vibrations. After noise reduction, the image size is uniformly scaled to a preset pixel width and preset pixel height, and the grayscale value of each pixel in the image is linearly normalized and mapped to an integer grayscale range of 0 to 255, ensuring that all acquired surface morphology images have a uniform size and a uniform grayscale standard. The preprocessed surface topography image is stored in the onboard data storage unit and directly read into the onboard embedded chip. The entire process requires no manual operation by astronauts and is autonomously triggered for data acquisition.

[0054] The morphology feature extraction and AI-powered intelligent prediction module is used to input the surface morphology image of the target space high-energy beam fused wire additive component into a lightweight model to obtain the internal defect quantitative parameters and mechanical performance prediction indicators of the target space high-energy beam fused wire additive component. The spaceborne embedded chip loads the ground-pre-trained lightweight model and performs two levels of computation on the on-orbit images: The first level of computation involves automatic extraction of multi-dimensional morphology features. The lightweight model automatically analyzes all key visual features in the image, including fused channel uniformity, interlayer overlap discontinuity, surface spatter distribution density, surface microcrack length, fused channel ripple amplitude, and oxide spot distribution density, among other multi-dimensional quantitative visual features. The second level of computation is multi-task parallel real-time inference. Based on the extracted multi-dimensional surface morphology features, two sets of prediction results are simultaneously output through a pre-trained cross-scale correlation model: the internal defect prediction group includes porosity, under-fusion, equivalent size of interlayer cracks, defect volume ratio, and defect risk level; the mechanical performance prediction group includes quantitative indicators such as elastic modulus, yield strength, tensile strength, fatigue limit, fracture toughness, and creep life.

[0055] The specific process of morphological feature extraction and AI intelligent prediction module inference is as follows: Preprocessed surface morphological images are read from the onboard data storage unit. These images include top and side views of the target component. The image size and grayscale standards are consistent with those of the images in the multimodal database used to train the lightweight model, both having preset pixel widths and heights, and grayscale ranges from 0 to 255. The surface morphological images are then input into the lightweight model deployed in the onboard embedded chip. This lightweight model is pre-stored in the read-only storage space of the onboard embedded chip before in-orbit flight. During inference, the computing unit within the chip loads the model parameters and performs forward propagation calculations. The input image first passes through the shared backbone network in the lightweight model, sequentially through convolutional layers, batch normalization layers, linear rectified activation functions, and four residual blocks. Finally, it is converted into a fixed-length feature vector through a global average pooling layer. This feature vector contains multi-dimensional morphological features extracted from the input image. The feature vectors are passed in parallel to the internal defect prediction branch and the mechanical property prediction branch in the lightweight model. In the internal defect prediction branch, the feature vectors are sequentially passed through two fully connected layers for feature dimensionality reduction. Finally, neurons in the output layer generate predicted values ​​for internal defect quantification parameters, including predicted values ​​for internal pore size, under-fusion area, interlaminar microcrack length, defect volume fraction, and defect risk level. In the mechanical property prediction branch, the feature vectors are similarly sequentially passed through two fully connected layers for feature dimensionality reduction. Finally, neurons in the output layer generate predicted values ​​for mechanical properties, including predicted values ​​for tensile strength, yield strength, maximum elongation, fatigue cycle life, and fracture toughness. The internal defect quantification parameters and mechanical property prediction indicators, as the final outputs of the lightweight model, are stored in the onboard data storage unit. The entire inference process is completed offline locally on the spacecraft, eliminating the need for real-time image transmission from Earth and avoiding the problems of deep space communication latency and insufficient communication bandwidth.

[0056] The performance output and on-orbit service safety assessment module incorporates a spacecraft structural component design threshold database. This module compares the quantified internal defect parameters and predicted mechanical performance indicators of the target space high-energy beam fused filament additive component with the design thresholds in the database, and determines whether the component is suitable for on-orbit service based on the comparison results. The database stores the allowable lower limit values ​​for mechanical performance, the maximum allowable defect size, and the maximum allowable defect volume fraction for different spacecraft structural components. The database also pre-loads the minimum allowable lower limit values ​​for mechanical performance, the maximum allowable defect size, and the maximum allowable defect volume fraction for different cabin supports, load-bearing pipelines, and connection structures.

[0057] The performance output and on-orbit service safety assessment module is specifically used to: compare the predicted mechanical performance indicators with the allowable lower limit of mechanical performance, and compare the internal defect quantification parameters with the maximum allowable defect size value and the maximum allowable defect volume fraction value; when all the predicted mechanical performance indicators are higher than the allowable lower limit of mechanical performance, and the internal defect quantification parameters are lower than the maximum allowable defect size value and the maximum allowable defect volume fraction value, the target space high-energy beam fused additive component is determined to be suitable for on-orbit service; when the predicted mechanical performance indicators are lower than the allowable lower limit of mechanical performance, and the internal defect quantification parameters are lower than the maximum allowable defect size value and the maximum allowable defect volume fraction value, the target space high-energy beam fused additive component needs to be partially repaired and retested; when the internal defect quantification parameters are higher than the maximum allowable defect size value or higher than the maximum allowable defect volume fraction value, the target space high-energy beam fused additive component is determined to be scrapped.

[0058] The specific judgment process of the performance output and on-orbit service safety judgment module is as follows: First, the internal defect quantification parameters and mechanical performance prediction indicators output by the morphology feature extraction and artificial intelligence intelligent prediction module are read from the onboard data storage unit. The internal defect quantification parameters include the predicted values ​​of internal pore diameter, under-fusion area, interlayer microcrack length, defect volume fraction, and defect risk level. The mechanical performance prediction indicators include the predicted values ​​of tensile strength, yield strength, maximum elongation, fatigue cycle life, and fracture toughness. At the same time, the allowable lower limit value of mechanical performance, the maximum allowable defect size value, and the maximum allowable defect volume fraction value corresponding to the type of aerospace structural component to which the current target component belongs are read from the aerospace structural component design threshold database. Second, a comparison calculation is performed. The predicted values ​​of tensile strength, yield strength, maximum elongation, fatigue cycle life, and fracture toughness are compared one by one with the corresponding lower limit values ​​in the allowable lower limit values ​​of mechanical properties to determine whether all the predicted mechanical property indicators are higher than the allowable lower limit values ​​of mechanical properties. At the same time, the predicted values ​​of internal pore diameter, under-fusion area, and interlaminar microcrack length are compared with the maximum allowable defect scale value to determine whether the predicted values ​​of the above defect scales are simultaneously lower than the maximum allowable defect scale value. Finally, the predicted value of defect volume fraction is compared with the maximum allowable defect volume fraction value to determine whether the predicted value of defect volume fraction is lower than the maximum allowable defect volume fraction value. Third, when all predicted mechanical performance indicators are higher than the lower limit of mechanical performance, and the predicted values ​​of internal pore diameter, under-fusion area, and interlayer microcrack length are lower than the maximum allowable defect size, and the predicted defect volume fraction is lower than the maximum allowable defect volume fraction, all predicted mechanical performance is higher than the design threshold and the defect level is level 1 or 2. The target space high-energy beam fused wire additive component is determined to be suitable for in-orbit service, and an instruction allowing in-orbit installation is generated. This instruction, along with the predicted data, is stored in the onboard data storage unit.Fourth, when any one of the predicted values ​​of tensile strength, yield strength, maximum elongation, fatigue cycle life, or fracture toughness is lower than the corresponding lower limit of the lower limit of mechanical properties, and the predicted values ​​of internal pore diameter, under-fusion area, and interlaminar microcrack length are simultaneously lower than the maximum allowable defect size, and the predicted value of defect volume fraction is lower than the maximum allowable defect volume fraction, the defect level is level 2 or 3 and the key mechanical properties are slightly lower than the standard. It is determined that the target space high-energy beam fused wire additive component needs to be retested after local repair. The system automatically outputs a defect spatial positioning command, which includes the estimated three-dimensional coordinate position of the defect in the component corresponding to the internal defect quantification parameters. This triggers the on-orbit fused wire repair welding equipment to perform local fused wire repair welding at the coordinate position. After the repair is completed, the functions of the on-orbit lightweight image acquisition module and the morphological feature extraction and artificial intelligence intelligent prediction module are re-executed to obtain the new internal defect quantification parameters and new mechanical property prediction indicators of the repaired component. The performance output is then compared and judged again in the on-orbit service safety judgment module. Fifth, when the predicted value of the internal pore diameter is higher than the maximum allowable defect size value, or the predicted value of the under-fusion area is higher than the maximum allowable defect size value, or the predicted value of the interlayer microcrack length is higher than the maximum allowable defect size value, or the predicted value of the defect volume fraction is higher than the maximum allowable defect volume fraction value, the defect level is level 4 or the key mechanical indicators are significantly lower than the design threshold, and there is a risk of failure in on-orbit service. The target space high-energy beam fused wire additive component is determined to be scrapped, a component scrapping command is generated, and the spacecraft control system is triggered to start the component reprinting process, and high-energy beam fused wire additive manufacturing is re-executed according to the original three-dimensional component model.

[0059] The performance output and on-orbit service safety assessment module also automatically generates a standardized assessment report. The report includes the component's unique number, forming process parameters, original surface acquisition images, internal defects and mechanical property prediction status, and the final safety assessment conclusion. The data is stored locally and can be transmitted back to the ground control center for long-term archiving as needed.

[0060] This invention belongs to the interdisciplinary technical fields of on-orbit high-energy beam fused wire additive manufacturing in space, aerospace machine vision non-destructive testing, and intelligent evaluation of the cross-scale performance of aerospace metal components. Its specific application areas include in-situ manufacturing and structural damage repair for near-Earth space stations, unmanned in-situ 3D printing of metals in lunar and Martian deep space, rapid fabrication of emergency load-bearing structures for manned spaceflight, and non-contact, non-destructive quality evaluation of on-orbit metal components. This invention targets two mainstream high-energy beam fused wire forming processes in space: electron beam fused wire additive manufacturing and laser fused wire additive manufacturing. It provides a lightweight intelligent prediction method for various commonly used aerospace structural filaments, such as titanium alloys, stainless steel, and aluminum alloys, based solely on component surface morphology images. This method eliminates the need for on-orbit mechanical testing machines or component re-inspection upon return to Earth. This invention differs from traditional ground-based offline testing, industrial computed tomography (CT) flaw detection, and universal materials laboratory mechanical testing systems. It is specifically adapted to the extreme aerospace conditions of space microgravity, high vacuum, strict limitations on payload weight, volume, and power consumption, unmanned autonomous operation, and rapid assembly and commissioning of components. Its beneficial effects are as follows: 1) This invention pioneers a lightweight, on-orbit performance evaluation system for additively manufactured space components, eliminating the need for mechanical testing and return-to-Earth inspection. It breaks through the traditional approach that relies on large mechanical testing machines, industrial computed tomography (CT) scans, and component return-to-Earth re-inspection. Performance evaluation is achieved solely through surface morphology images of the component, adapting to the stringent constraints of weight, volume, and power consumption in space payloads. The entire on-orbit added testing hardware consists of only one aerospace-grade radiation-resistant charge-coupled device (CCD) camera. It eliminates the need for large-scale electromechanical testing equipment, on-orbit mechanical testing machines, heavy flaw detection equipment such as industrial CT scans and ultrasonic sensors, and component return-to-Earth re-inspection. A complete, standardized performance evaluation report is output instantly on-orbit. Compared to traditional solutions that incorporate small mechanical testing machines and miniature industrial CT scans, this invention only adds a CCD vision device, resulting in an on-orbit payload weight of less than 500 grams. The overall power consumption is low, minimizing the impact on spacecraft core power and payload resources. It is suitable for various manned space stations and small, lightweight deep-space probes, solving the payload constraint problem of not being able to carry large-scale mechanical testing and heavy non-destructive testing equipment in space. The neural network model is compressed and optimized in advance on the ground. It relies on the onboard embedded chip for local real-time analysis and inference, without relying on the deep space remote communication link, to achieve unmanned autonomous intelligent evaluation throughout the entire process. No human intervention is required throughout the process. It is fully adaptable to the extreme aerospace conditions of space microgravity, high vacuum, strict limitations on payload weight, volume and power consumption, unmanned autonomous operation and rapid assembly and service of components.

[0061] 2) This invention constructs a three-level cross-scale quantitative correlation mechanism from surface morphology images to internal defect parameters and then to mechanical performance indicators, enabling the visual characterization of hidden internal quality. Existing on-orbit inspection technologies can only qualitatively observe surface morphology with the naked eye. This invention trains a multi-task neural network model using a massive ground-based bound and labeled dataset to establish a precise mathematical mapping relationship between surface micromorphology, internal defects, and mechanical properties. It can quantitatively predict internal micro-defects that cannot be directly observed and the overall service performance, connecting the complete technical link of cross-scale quantitative evaluation of surface morphology images, internal defect parameters, and mechanical performance indicators. It upgrades from traditional manual visual qualitative rough observation to artificial intelligence quantitative full-indicator intelligent prediction, which can identify micron-level surface micro-features and hidden internal micro-defects that are indistinguishable to the naked eye, filling the gap in cross-scale quantitative correlation technology from surface morphology features to internal metallurgical defects and macroscopic mechanical properties.

[0062] 3) This invention achieves universal compatibility with dual heat sources and multiple aerospace alloy wires, and is compatible with both electron beam and laser-based fused wire additive manufacturing processes, which are mainstream in space applications. During the ground training phase, the model simultaneously covers two mainstream high-energy beam heat sources in orbit and three types of core structural alloys for space stations. The multimodal database comprehensively includes the variation patterns of surface morphology images, internal defect parameters, and mechanical performance indicators under different processes and materials. A single lightweight model can be adapted to different printing equipment and wires, eliminating the need for separate remodeling for individual equipment and materials. This significantly reduces the workload of ground-based pre-research and facilitates batch application on multiple manned space stations and deep space exploration platforms. It also establishes a universal intelligent evaluation framework compatible with both electron beam and laser dual heat sources and multiple types of aerospace alloy wires.

[0063] 4) This invention achieves a significant leap in the efficiency of closed-loop on-orbit manufacturing operations. Traditional component return-to-Earth inspection methods have evaluation cycles measured in months and years. This invention automatically acquires images upon component printing, completing a full set of performance predictions and safety assessments in orbit in real time. The reasoning time for a single image is less than 1 second, enabling a closed-loop operation mode where evaluation is completed immediately upon completion of manufacturing, and installation and service are initiated immediately upon successful evaluation. This overcomes the application limitations of long return-to-Earth inspection cycles and high costs associated with on-orbit component inspections, supporting rapid response to emergency repairs of space station structures and continuous, uninterrupted unmanned manufacturing operations in deep space. The automatic image acquisition upon component printing and the real-time completion of a full set of performance predictions and safety assessments realize a closed-loop process encompassing space-based filament additive manufacturing, visual image acquisition, AI-powered intelligent performance prediction, defect classification and service assessment, and direct on-orbit installation and use. This supports the implementation of emergency repairs for space stations and unmanned in-situ manufacturing projects in deep space, and strongly supports specific application areas such as in-situ manufacturing and structural damage repair for near-Earth space stations, rapid preparation of emergency load-bearing structures for manned spaceflight, and non-contact, non-destructive quality evaluation of on-orbit metal components.

[0064] 5) This invention offers unique engineering value for unmanned deep space exploration. Deep space exploration missions such as those to the Moon and Mars lack return capsules, rendering traditional return-to-Earth testing methods completely ineffective. This invention autonomously completes quality assessment throughout the entire process in orbit, requiring no human intervention and no component return to Earth. It solves the fundamental industry challenges of the lack of return channels in deep space and the inability to conduct destructive mechanical testing in orbit. It is the only feasible performance evaluation technology solution for in-situ metal manufacturing in future extraterrestrial research bases, supporting major national aerospace projects such as manned Mars landings and the operation and maintenance of long-term lunar research stations. It directly serves the specific application areas of unmanned in-situ metal 3D printing in deep space missions on the Moon and Mars.

[0065] 6) This invention significantly reduces the life-cycle cost of aerospace structural components. It eliminates the cost of round-trip launch and return transportation of components, saving substantial launch costs per component inspection; it eliminates the need for separately developing and launching dedicated heavy-duty on-orbit mechanical and flaw detection equipment, significantly reducing the development and launch costs of aerospace equipment; it avoids the scrapping of qualified components due to destructive on-orbit mechanical testing, saving on expensive aerospace-grade alloy wire; it allows for the early prediction of high-risk defective components, preventing the installation of structures with safety hazards in orbit, and avoiding the enormous costs of on-orbit structural failure repair and cabin damage.

[0066] 7) This invention significantly improves the level of aerospace safety management and control. It accurately quantifies the safety margin of components during on-orbit service, eliminating the shortcomings of on-orbit quality assessment that rely on human experience, qualitative judgment alone, and the inability to quantify safety margins. It forms a standardized, reproducible, and long-term archiveable on-orbit quality assessment specification, eliminating safety assessment errors caused by human experience judgment, fully meeting the high reliability and rigorous quality control standards of manned spaceflight, and providing reliable assurance for the long-term on-orbit service safety of various aerospace structural components.

[0067] 8) This invention, through comparison of alternative solutions and demonstration of its technological superiority, comprehensively surpasses all existing traditional technical solutions in six core dimensions: lightweighting, testing timeliness, material and process versatility, deep space scenario adaptability, full-cycle economy, and unmanned autonomous operation. Traditional component return-to-Earth ground testing solutions have zero on-orbit hardware weight increase, but testing takes months to years and cannot be performed autonomously in orbit; on-orbit mechanical and industrial computed tomography (CT) scanning equipment solutions require hundreds of kilograms of heavy equipment and require astronauts to operate and clamp it; solutions relying solely on manual visual camera observation cannot quantitatively predict mechanical properties or identify internal micro-defects. This invention, with an on-orbit hardware weight increase of only less than 500 grams, utilizes an aerospace-grade radiation-resistant charge-coupled device (CCD) camera, completes a full set of performance predictions in 1 second, is a non-contact, non-destructive testing method that preserves the component's integrity, quantitatively outputs all mechanical indicators, quantitatively predicts internal micro-defects through surface morphology images, is fully adaptable to unmanned deep-space in-situ manufacturing, has extremely low hardware costs and no additional transportation expenses, and features fully automated triggering and AI-driven autonomous judgment throughout the entire process.

[0068] Furthermore, the system and method embodiments provided in the above embodiments belong to the same concept, and their specific implementation process can also be found in the method embodiments, which will not be repeated here.

[0069] An electronic device according to an embodiment of the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements any of the above-mentioned methods for predicting the performance of space high-energy beam filament additive components.

[0070] An embodiment of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements any of the above-described methods for predicting the performance of space high-energy beam filament additive components.

[0071] The above description is merely a preferred embodiment of the present invention and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this invention is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-disclosed concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this invention.

[0072] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for predicting the performance of space high-energy beam fused wire additive components, characterized in that, include: A multimodal database is pre-constructed that associates surface morphology images, internal defect parameters, and mechanical performance indicators of space high-energy beam fused wire additive components. A multi-task neural network model is trained based on the database, and then the trained multi-task neural network model is compressed into a lightweight model. Surface morphology images of the target space high-energy beam fused wire additive component were acquired in orbit; The surface morphology image of the target space high-energy beam fused wire additive component is input into the lightweight model to obtain the internal defect quantification parameters and mechanical performance prediction index of the target space high-energy beam fused wire additive component.

2. The method for predicting the performance of a space high-energy beam fused wire additive component according to claim 1, characterized in that, Also includes: The internal defect quantification parameters and mechanical performance prediction indicators of the target space high-energy beam filament additive component are compared with the design thresholds in the aerospace structural component design threshold database, and the suitability of the target space high-energy beam filament additive component for in-orbit service is determined based on the comparison results.

3. The method for predicting the performance of a space high-energy beam fused wire additive component according to claim 2, characterized in that, The aerospace structural component design threshold database stores the allowable lower limit of mechanical properties, the maximum allowable defect size value, and the maximum allowable defect volume fraction value for different aerospace structural components. The internal defect quantification parameters and mechanical performance prediction indicators of the target space high-energy beam filament additive component are compared with the design thresholds in the aerospace structural component design threshold database. Based on the comparison results, it is determined whether the target space high-energy beam filament additive component is suitable for in-orbit service, including: The predicted mechanical performance indicators are compared with the lower limit of mechanical performance, and the quantified internal defect parameters are compared with the maximum permissible defect size value and the maximum permissible defect volume fraction value. When all the predicted mechanical performance indicators are higher than the lower limit of mechanical performance, and the quantified internal defect parameters are lower than both the maximum permissible defect size value and the maximum permissible defect volume fraction value, the target space high-energy beam fused filament additive component is determined to be suitable for in-orbit service. When the predicted mechanical performance indicators are lower than the lower limit of mechanical performance, and the quantified internal defect parameters are lower than both the maximum permissible defect size value and the maximum permissible defect volume fraction value, the target space high-energy beam fused filament additive component needs to be partially repaired and retested. When the quantified internal defect parameters are higher than the maximum permissible defect size value or higher than the maximum permissible defect volume fraction value, the target space high-energy beam fused filament additive component is determined to be scrapped.

4. The method for predicting the performance of a space high-energy beam filament additive component according to any one of claims 1 to 3, characterized in that, The trained multi-task neural network model is compressed into the lightweight model by means of: using knowledge distillation, model pruning and quantization compression to compress the trained multi-task neural network model into a lightweight model.

5. A performance prediction system for space high-energy beam fused wire additive components, characterized in that, It includes a ground-based offline modeling and database construction module, an on-orbit lightweight image acquisition module, and an AI intelligent prediction module; The ground offline modeling and database building module is used to: pre-build a multimodal database that associates the surface morphology image, internal defect parameters and mechanical performance indicators of the space high-energy beam fused wire additive component, train a multi-task neural network model based on the database, and then compress the trained multi-task neural network model into a lightweight model. The on-orbit lightweight image acquisition module is used to: acquire on-orbit surface morphology images of the target space high-energy beam fused wire additive component; The AI ​​intelligent prediction module is used to: input the surface morphology image of the target space high-energy beam fused wire additive component into the lightweight model to obtain the internal defect quantification parameters and mechanical performance prediction index of the target space high-energy beam fused wire additive component.

6. The performance prediction system for a space high-energy beam fused wire additive component according to claim 5, characterized in that, It also includes a performance output and on-orbit service safety assessment module, which has a built-in aerospace structural component design threshold database. The performance output and on-orbit service safety assessment module is used to: compare the internal defect quantification parameters and mechanical performance prediction indicators of the target space high-energy beam filament additive component with the design thresholds in the aerospace structural component design threshold database, and determine whether the target space high-energy beam filament additive component is suitable for on-orbit service based on the comparison results.

7. The performance prediction system for a space high-energy beam filament additive component according to claim 6, characterized in that, The aerospace structural component design threshold database stores the allowable lower limit of mechanical properties, the maximum allowable defect size value, and the maximum allowable defect volume fraction value for different aerospace structural components. The performance output and on-orbit service safety determination module is specifically used to: compare the predicted mechanical performance indicators with the allowable lower limit of mechanical performance, and compare the internal defect quantification parameters with the maximum allowable defect size value and the maximum allowable defect volume fraction value; when all the predicted mechanical performance indicators are higher than the allowable lower limit of mechanical performance, and the internal defect quantification parameters are simultaneously lower than the maximum allowable defect size value and the maximum allowable defect volume fraction value, the target space high-energy beam fused additive component is determined to be suitable for on-orbit service; when the predicted mechanical performance indicators are lower than the allowable lower limit of mechanical performance, and the internal defect quantification parameters are simultaneously lower than the maximum allowable defect size value and the maximum allowable defect volume fraction value, the target space high-energy beam fused additive component needs to be partially repaired and retested; when the internal defect quantification parameters are higher than the maximum allowable defect size value or higher than the maximum allowable defect volume fraction value, the target space high-energy beam fused additive component is determined to be scrapped.

8. A performance prediction system for space high-energy beam filament additive components according to any one of claims 5 to 7, characterized in that, The ground offline modeling and library building module is specifically used to: compress the trained multi-task neural network model into a lightweight model using knowledge distillation, model pruning, and quantization compression methods.

9. An electronic device, characterized in that, The invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the performance prediction method for a space high-energy beam filament additive component as described in any one of claims 1 to 4.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the performance prediction method for a space high-energy beam filament additive component as described in any one of claims 1 to 4.

Citation Information

Patent Citations

  • Laser additive manufacturing shape active control method based on machine learning

    CN115796026A

  • Machine vision sensing laser fuse wire additive manufacturing forming control method and system

    CN117773341A