Aerospace equipment-oriented digital-analog fusion dual-drive digital-real fusion evaluation method

Abstract

The invention discloses a digital-analog fusion dual-drive digital-real fusion evaluation method for aerospace equipment, and belongs to the technical field of intelligent manufacturing. The method comprises the following steps: generating enhanced product data covering an extreme working condition through a conditional generative adversarial network, and constructing an environmental parameter matrix to realize systematic characterization of an assembly working condition; secondly, a dynamic data scheduler and a layered reward mechanism are designed, and intelligent fusion of the real product data and the enhanced product data in the reinforcement learning training process is achieved; and finally, constructing a two-way cooperative training framework, combining the feature extraction capability of the graph neural network with the decision-making capability of reinforcement learning, and improving the model generalization capability through element course learning of environment perception. In the online evaluation stage, a hybrid prediction mechanism based on uncertainty estimation is adopted, and the prediction reliability of the system under different working conditions is ensured. According to the method, the quality prediction accuracy under the small sample condition is remarkably improved, and the method is particularly suitable for the fields of aerospace, precise instruments and the like with extremely high assembly precision requirements.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent manufacturing, in particular to a digital-analog fusion dual-drive digital-physical fusion evaluation method for aerospace equipment. The method is especially suitable for manufacturing industries such as automobile manufacturing, precision electronics, aerospace, etc. that have extremely high requirements for assembly precision and consistency. Through the deep fusion of digital testing and physical assembly processes, accurate and advanced prediction of product quality is achieved. BACKGROUND

[0002] In the existing production line assembly quality control system, methods based on statistical process control (SPC) and detection techniques based on fixed thresholds are mainly relied on. Although these traditional methods have played an important role in the industrialization process, they have shown obvious limitations when faced with modern manufacturing modes with high complexity, multiple varieties, and small batches. The SPC method mainly monitors key parameters of the production process to achieve quality control, but it is essentially a post-analysis method and cannot achieve early prediction and early warning of quality problems.

[0003] In recent years, with the development of artificial intelligence technology, quality prediction methods based on machine learning have been gradually applied to industrial practice. These methods analyze historical production data to build prediction models, which can theoretically achieve early detection of quality problems. However, in actual application, these methods face multiple challenges. The first problem is the severe shortage of high-quality training samples, especially in the acquisition of defect samples. Due to the continuous improvement of the maturity of modern manufacturing processes, the pass rate of production lines is usually maintained at a high level, resulting in a very limited number of defect samples. This sample imbalance problem seriously restricts the model's ability to learn fault patterns, directly affecting the prediction accuracy.

[0004] Secondly, existing models have obvious deficiencies in understanding complex assembly relationships. In a typical assembly process, product quality is often not determined by a single part parameter, but by the complex interaction between multiple components. This interaction presents typical topological structure characteristics, including assembly sequence relationship, fitting tolerance chain, mechanical transmission path, etc. Traditional machine learning models, such as support vector machines and random forests, cannot effectively capture this structured and relational feature dependency, resulting in limited prediction accuracy when dealing with complex assemblies.

[0005] Furthermore, existing quality assessment methods generally lack forward-looking decision-making capabilities. Most methods remain at the "perception-diagnosis" level, that is, judging whether the current state is abnormal by analyzing existing data. This approach is essentially still passive quality control. With the development of digital twin technology, manufacturing enterprises have been able to build digital mirrors of production lines, but how to effectively combine the simulation advantages of digital space with actual data from the physical world to form a closed-loop control of "prediction-decision-optimization" remains a technical challenge for the industry.

[0006] Taking the assembly process of aero-engine rotors as an example, existing methods typically employ vibration monitoring or dimensional inspection for quality verification. This approach can only identify problems after assembly is complete and cannot intervene in real-time during the assembly process. A few machine learning-based methods attempt to predict assembly quality by analyzing press-fit force curves, but due to a lack of modeling capabilities for the complex mechanical relationships between parts and insufficient training data on rare failure modes (such as misalignment and fretting wear), their predictions are often unreliable. Summary of the Invention

[0007] The core technical problem this invention aims to solve is to address the shortcomings of existing production line assembly quality assessment methods in three dimensions: data foundation, relationship modeling, and predictive capability. It provides a dual-driven, data-real fusion assessment method for aerospace equipment, integrating data and model data. Specifically, it needs to overcome the following technical bottlenecks: first, how to effectively learn complex failure modes under small sample conditions; second, how to accurately capture complex topological relationships during the assembly process; and third, how to achieve a shift from passive detection to proactive predictive decision-making. This will solve the problems of insufficient prediction accuracy caused by the scarcity of defect samples, difficulty in modeling complex assembly relationships, and lack of forward-looking decision-making capabilities in existing quality prediction technologies.

[0008] Compared with existing technologies, the innovation of this invention is mainly reflected in breakthroughs at three levels: at the data level, efficient data augmentation is achieved by introducing conditional generative adversarial networks; at the training level, a dynamic fusion mechanism for collaborative real and digital data is established; and at the model level, a bidirectional collaborative architecture of graph neural networks and reinforcement learning is constructed. These innovations enable this invention to better adapt to the stringent quality control requirements of modern intelligent manufacturing, providing an effective technical path for achieving true predictive quality control.

[0009] The core of this invention lies in constructing a complete technical system comprising three major modules: data augmentation, data fusion, and model fusion. Firstly, in the data augmentation module, an environmental parameter matrix is ​​introduced to systematically represent assembly conditions. Then, using conditional generative adversarial networks (CGANs), augmented product data covering extreme conditions and high-risk decision boundaries is generated based on actual assembled product data, effectively solving the bottleneck problem of scarce defect samples.

[0010] In the data fusion module, a dynamic triple adaptive adjustment mechanism was designed. Through a dynamic data scheduler based on course learning, the sampling ratio of actual assembly data to augmented assembly data is intelligently adjusted during reinforcement learning training, achieving progressive training from basic learning to extreme challenges. Simultaneously, a hierarchical reward function is employed, adding exploration rewards and cross-data source consistency rewards on top of basic rewards, guiding the agent to improve generalization ability while maintaining decision reliability.

[0011] The model fusion module employs a bidirectional collaborative meta-learning framework, which is the core innovation of this invention. A graph neural network is used to model the assembly's graph structure and extract deep feature vectors; the reinforcement learning agent then makes decisions based on these features and environmental parameters. The key innovation lies in establishing a bidirectional collaborative training loop. When the agent receives a negative reward, the graph neural network parameters are optimized through backpropagation of feature alignment loss, forcing the network to enhance its ability to represent risky states. Combined with meta-curriculum learning based on environmental perception, the training data is clustered into multiple curriculum units according to working conditions, and the meta-learning framework improves the model's adaptation speed under new working conditions.

[0012] The online prediction phase employs a hybrid prediction mechanism based on uncertainty estimation. Through dual-path parallel reasoning and real-time uncertainty assessment, it dynamically integrates robust prediction based on practical knowledge and exploratory prediction based on virtual knowledge to ensure that the system maintains optimal prediction performance under different operating conditions.

[0013] Technical solution of the present invention:

[0014] A dual-drive data-real fusion evaluation method for aerospace equipment, comprising:

[0015] S110 uses actual product data from the physical production line as seeds and generates enhanced product data using conditional generative adversarial network CGAN.

[0016] S120 utilizes real assembly data and augmented assembly data to perform graph structure modeling on the assembly using a graph neural network, and concatenates the output deep features with environmental parameters as the state input of the reinforcement learning agent; the real assembly data represents a fusion dataset consisting of real product samples and their corresponding environment; the augmented assembly data represents a fusion dataset consisting of augmented product samples and their corresponding environment.

[0017] S130 adopts an environment-aware meta-curriculum learning strategy, which clusters enhanced assembly data and actual assembly data based on the environmental parameter matrix to construct curriculum units, and conducts collaborative training of graph neural networks and reinforcement learning agents within this framework.

[0018] S140, in the online prediction stage, based on the deep features extracted by the graph neural network, a hybrid prediction mechanism based on uncertainty estimation is used to make quality predictions. The hybrid prediction mechanism uses dual-path parallel reasoning and integrates the results of the robust path based on actual knowledge and the exploration path based on reinforcement to output the final prediction results and confidence.

[0019] Compared with existing technologies, this invention has three significant advantages: it solves the problem of small-sample learning through a data augmentation system with environmental parameterization; it achieves a joint improvement in model representation and decision-making capabilities through bidirectional collaborative training; and its hybrid prediction mechanism based on uncertainty estimation ensures the reliability and adaptability of the system in practical applications. This invention is particularly suitable for fields with high assembly precision requirements, such as aerospace, automotive manufacturing, and precision electronics, providing effective technical support for production line quality control. Attached Figure Description

[0020] Figure 1 This is a framework diagram of a dual-drive data-real fusion evaluation method for aerospace equipment. Detailed Implementation

[0021] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0022] This invention provides a dual-drive digital-real fusion evaluation method for aerospace equipment, which can be divided into the following key steps in its implementation process. Figure 1 This is a framework diagram for a dual-drive data-real fusion evaluation method for aerospace equipment. (Example:) Figure 1 As shown, the method includes:

[0023] S110 uses actual product data from the physical production line as seeds and generates enhanced product data using Conditional Generative Adversarial Network (CGAN).

[0024] This step is performed in the data augmentation module.

[0025] Taking the tightening station of the electromagnetic pilot valve plug bolt as an example, S110 can include: deploying high-precision torque sensors and angle encoders on the physical production line to collect torque-angle curves under normal operating conditions as actual product data. In specific implementation, the data acquisition frequency is set to 50Hz to ensure the capture of the complete assembly dynamic process. The collected actual product data is input into a conditional generative adversarial network (GAN). The generator of the GAN uses the actual product data as a condition and learns the data distribution characteristics through a five-layer fully connected network to generate synthetic data under extreme parameter conditions as enhanced product data. This includes operating conditions where the torque value is outside the normal range of 85Nm-115Nm (such as 70Nm or 130Nm) and abnormal patterns (such as slippage due to sudden angle changes). The discriminator of the GAN uses a convolutional neural network structure, distinguishing between real and generated data through four convolutional layers. During training, the learning rate is set to 0.0002, the batch size to 32, and after 5000 iterations, the distribution difference between the generated data and the real data is less than 5%.

[0026] A complete data representation method was designed for data-real fusion evaluation. This method defines the following core data structures:

[0027] Environmental parameter matrix b represents the total number of assembly steps, c represents the total number of environmental parameter types (such as torque and temperature), j is the index of the assembly step (j=1,2,…,b), and k is the index of the environmental parameter type (k=1,2,…,c). This represents the normalized value of the k-th environmental parameter at the j-th assembly stage; N represents the environmental stage vector, indicating which stages the assembly includes. Represents an environmental parameter vector, defining the type of environmental parameter being considered (such as torque and temperature).

[0028] Actual product data Enhance product data 'a' represents the total number of parts in a product, 'i' is the index of the part (i=1,2,…,a), and 'j' is the index of the assembly stage (j=1,2,…,b). and These represent the production data of the i-th part in the j-th stage of the actual and enhanced product data, respectively.

[0029] Actual assembly data , representing a fusion dataset consisting of p actual product samples and their corresponding environments;

[0030] Enhanced assembly data , represents a fusion dataset consisting of q augmented product samples and their corresponding environments, and typically q is much larger than p;

[0031] Environmental parameter vector It is the baseline vector for all parameter types, and in the corresponding actual product data, there are actual installation environment parameters. p represents the number of samples, corresponding to satisfy This represents normal operating conditions; based on the corresponding enhanced product data, environmental parameters are enhanced. q represents the number of samples, and q is much larger than p, corresponding to satisfy It represents extended operating conditions covering extreme and abnormal conditions.

[0032] S120 uses real assembly data and augmented assembly data to perform graph structure modeling of the assembly using a graph neural network, and concatenates the output depth features with environmental parameters as the state input of the reinforcement learning agent.

[0033] This step is performed in the model fusion module. S120 includes:

[0034] S120-1 abstracts the assembly into a graph structure. Taking the assembly of an electromagnetic pilot valve as an example, 14 nodes are defined, including the valve body, valve cover, 2 screw plugs, and 7 expansion plugs. Node attributes include component model, theoretical torque value, and real-time torque curve characteristic value. 42 edges are defined, such as screw plug-valve cover and screw plug-valve body. Edge attributes include assembly sequence and fit tolerance.

[0035] S120-2 utilizes real and augmented assembly data to learn the graph structure of the assembly using a graph neural network (GNN), outputting a deep feature vector. This deep feature vector is then concatenated with environmental parameters as the state input to the reinforcement learning agent. The GNN employs a graph attention network (GAT) with a three-layer structure and hidden layer dimensions of 64, 128, and 64 respectively. It uses the LeakyReLU activation function and aggregates neighborhood information through a multi-head attention mechanism (8 heads), ultimately outputting a 128-dimensional deep feature vector H. This deep feature vector H is then combined with the environmental parameter vector of the current stage. After being concatenated, they together serve as the state input for the reinforcement learning (RL) agent. .

[0036] S130 employs an environment-aware meta-curriculum learning strategy, which clusters enhanced assembly data and actual assembly data based on the environmental parameter matrix to construct curriculum units, and conducts collaborative training of graph neural networks and reinforcement learning agents within this framework.

[0037] Meta-curriculum learning of environmental perception is first based on the environmental parameter matrix. , training set The augmented assembly data and actual assembly data were clustered into multiple course units, such as "high torque - normal friction" and "normal torque - low friction," which have clear physical meanings. During training, a meta-learning framework was employed. The inner loop performed 1000 fast adaptations on each course unit, while the outer loop computed meta-gradients on the mixed validation set to optimize the model's generalization ability across courses. The mixed validation set was a validation dataset composed of a proportional mix of actual assembly data and augmented assembly data. The meta-batch size was set to 4, and the meta-learning rate to 0.001, significantly improving the model's adaptation speed under new operating conditions.

[0038] The reinforcement learning agent employs the Proximal Policy Optimization (PPO) algorithm. Both the policy network and value network use a three-layer fully connected structure. A discount factor γ = 0.99 and a generalized advantage estimation parameter λ = 0.95 are set. Each update executes for 10 epochs, with a learning rate of 0.0001. The reinforcement learning agent is trained by interacting with the environment, collecting 4096 samples in each training step.

[0039] In terms of data scheduling and reward design during reinforcement learning training, the dynamic data scheduler adopts a course-based scheduling strategy. In the early training phase (first 10,000 iterations), it samples real assembly data with a 70% probability to ensure the stability of policy learning. In the middle training phase (10,000-30,000 iterations), the sampling ratio of augmented assembly data is gradually increased to 50%. In the later training phase (after 30,000 iterations), the sampling ratio of augmented assembly data is increased to 70%, focusing on training decision-making capabilities under extreme conditions. The reward function design adopts a hierarchical structure: when using real assembly data, a correct prediction is rewarded with a base reward of +1, and an incorrect prediction is penalized with -1. When using augmented assembly data, successfully avoiding virtual faults (such as over-tightening or under-tightening) is rewarded with an additional exploration reward of +0.5 on top of the base reward. Simultaneously, a consistency reward is set across data sources: when the agent makes the same correct decision regarding different data source states with feature similarity exceeding 0.9, an additional collaborative reward of +0.25 is given.

[0040] In the bidirectional collaborative training loop, the system establishes a reverse optimization channel between the GNN and RL. Specifically, when the reinforcement learning agent is in a certain state... When a negative reward is obtained, not only are the parameters of its own policy network updated, but the gradient of the feature alignment loss is also calculated. The trainable parameters of the graph neural network are updated via backpropagation based on the gradient of the feature alignment loss, wherein the alignment loss... The aim is to significantly improve the feature vectors of the model that lead to negative reward states under new operating conditions in this way. With the feature vector that leads to a positive reward state The distance between them This represents the set of all trainable parameters of the GNN. This mechanism forces the GNN to improve its ability to represent risk states. Specifically, the feature alignment loss is used as one of the training objectives of the graph neural network, forming a loss function together with the original feature learning objectives in the graph neural network, namely its internal supervised learning loss and self-supervised feature learning loss. Gradient descent is then used to backpropagate and update the graph neural network parameters.

[0041] S140, in the online prediction stage, based on the deep features extracted by the graph neural network, a hybrid prediction mechanism based on uncertainty estimation is used to make quality predictions. The hybrid prediction mechanism uses dual-path parallel reasoning and integrates the results of the robust path based on actual knowledge and the exploration path based on reinforcement to output the final prediction results and confidence.

[0042] Among them, the robust path outputs prediction results based on actual assembly data. The exploration path is based on the prediction results of enhanced assembly data output. At the same time, calculate the current state. With training set The uncertainty estimate U is obtained by taking the feature cosine similarity of all states: the current state The corresponding actual assembly data is input into the graph neural network to obtain the corresponding deep feature vector. ,Should The current state The mathematical encoding representation, and then the training set Each sample in the graph neural network is used to obtain a corresponding deep feature vector. Calculate cosine similarity Finally, an uncertainty estimate is obtained. The training set It is the complete dataset used to train graph neural networks and reinforcement learning agents, consisting of both live and augmented assembly data. The final prediction output is guided by uncertainty: when... When (high certainty), adopt ;when When (low certainty), adopt In other cases, a weighted average is used. The system also outputs the confidence level. and the sources of information for decision-making.

[0043] The entire implementation process is monitored in real time through a Unity3D-based visual interface. The interface dynamically displays key metrics such as data augmentation effects, the ratio of the two types of data used during training, model prediction confidence, and changes in feature alignment loss during bidirectional collaborative training loops. The system incorporates an anomaly detection mechanism; when the standard deviation of the prediction confidence for 10 consecutive samples exceeds 0.15, the model recalibration process is automatically triggered. Once the model's prediction accuracy on the validation set remains stable above 95% for 10 consecutive epochs, and the loss function converges to below 0.1, it can be containerized and deployed to the production line edge server for practical use.

[0044] Application Examples

[0045] To verify the applicability of this method in the assembly of precision fluid control components, it was implemented and validated on the pilot valve assembly line of an aerospace hydraulic system manufacturing company. This assembly line involves the precise fitting of multiple components such as valve body, valve core, spring, and sealing ring, with extremely stringent requirements for sealing performance, response consistency, and durability. Furthermore, in actual production, defect samples are scarce, and the assembly error propagation path is complex.

[0046] (1) Implementation configuration and data preparation

[0047] a. High-precision pressure sensors, torque sensors, and micro-displacement measuring instruments were deployed at key work stations such as valve body assembly, valve core selection, and bolt tightening to collect actual product data under normal working conditions, resulting in 280 sets of valid assembly samples.

[0048] b. Based on CGAN, enhance product data were generated, simulating risky operating conditions such as extreme pressure (0.5~35MPa), temperature gradient (-40℃~120℃), particulate contamination, and minor damage to sealing rings, generating 4500 sets of enhanced samples.

[0049] c. The environmental parameter matrix is ​​designed as 9×5, corresponding to 9 assembly stages and 5 types of environmental parameters (pressure, temperature, vibration, cleanliness, and torque).

[0050] (2) Model training and validation

[0051] a. The method of this invention is used for training, with the GAT hidden layer dimension set to [64, 128, 64], and the PPO algorithm parameters are consistent with the above description.

[0052] b. The comparison methods include: traditional threshold alarm methods, XGBoost-based quality classification models, and GNN+RL baseline models that do not incorporate augmented data and bidirectional co-training.

[0053] c. Dataset partitioning: training set (280 real-world models + 4500 augmented models), validation set (70 real-world models), and test set (70 real-world models, including 15 sets of simulated internal leakage and jamming fault samples).

[0054] (3) Implementation results

[0055] a. The assembly quality prediction accuracy of traditional threshold alarms is 78.2%, the internal leakage fault detection rate is 70.0%, the average single prediction time is less than 50 milliseconds, and the adaptation error to unknown working conditions is high.

[0056] b. XGBoost has an assembly quality prediction accuracy of 87.7%, an internal leakage fault detection rate of 85.3%, an average single prediction time of approximately 200 milliseconds, and a medium adaptation error for unknown operating conditions.

[0057] c. The assembly quality prediction accuracy of the graph neural network and reinforcement learning baseline but without reinforcement learning method is 86.4%, the internal leakage fault detection rate is 78.9%, the average single prediction time is 250 milliseconds, and the adaptation error to unknown working conditions is medium.

[0058] d. The assembly quality prediction accuracy of the method of the present invention is 92.1%, the internal leakage fault detection rate is 92.6%, the average single prediction time is 300 milliseconds, and the adaptation error to unknown working conditions is low;

[0059] (4) Effect Analysis

[0060] a. The method of the present invention is significantly better than the comparative method in terms of assembly quality prediction accuracy and internal leakage fault detection rate, especially showing good identification ability under the condition of a very small number of fault samples.

[0061] b. Through meta-curriculum learning based on environmental perception, the model can quickly adapt to the new "low temperature-high pressure" coupled operating conditions, reducing prediction errors by approximately 30%.

[0062] c. During the online prediction phase, the system automatically activated the exploration path for a batch of sealing rings with minor damage (U=0.28), providing early warning and successfully preventing three post-assembly leakage accidents.

[0063] d. The system ran continuously for 4 months, and the model showed good stability. The standard deviation of the prediction confidence level was always below 0.1, which met the high reliability operation requirements of the aviation hydraulic valve production line.

[0064] The contents not described in detail in this specification are existing technologies known to those skilled in the art.

[0065] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A digital-analog fusion dual-drive digital-real fusion evaluation method for aerospace equipment, characterized in that, Comprise: S110, using conditional generative adversarial network CGAN to generate enhanced product data by taking physical production line installation product data as seed; S120, using installation assembly data and enhanced assembly data, using graph neural network to model the assembly body in graph structure, and splicing the output deep features and environmental parameters as the state input of the reinforcement learning agent; The installation assembly data represents a fusion data set composed of installation product samples and their corresponding environments; The enhanced assembly data represents a fusion data set composed of enhanced product samples and their corresponding environments; S130, using environment-aware meta-course learning strategy, clustering enhanced assembly data and installation assembly data based on environmental parameter matrix to construct course units, and training graph neural network and reinforcement learning agent in this framework; S140, in the online prediction stage, based on the deep features extracted by the graph neural network, a hybrid prediction mechanism based on uncertainty estimation is used for quality prediction; The hybrid prediction mechanism outputs the final prediction result and confidence by parallel inference of double paths and fusion of the results of robust path based on installation knowledge and exploration path based on enhancement.

2. The space equipment-oriented digital-analog fusion dual-drive digital-real fusion evaluation method according to claim 1, characterized in that, For numerical-real fusion evaluation, a data representation method is designed, which defines the following core data structures: environment parameter matrix , b represents the total number of assembly links, c represents the total number of environment parameter types, j is the index of the assembly link, j = 1, 2, …, b, k is the index of the environment parameter type, k = 1, 2, …, c, represents the normalized value of the kth environment parameter on the jth assembly link; N represents the environment link vector, indicating which links are included in the assembly, represents the environment parameter vector, and the type of environment parameter considered is defined; Assembly product data , enhanced product data , a represents the total number of components contained in a product, i is the index of the component, i = 1, 2, …, a, j is the index of the assembly link, j = 1, 2, …, b, and respectively represent the production data of the i-th component in the j-th assembly link in the assembly product data and the enhanced product data; Physical assembly data , representing a fused dataset of p physical product samples with their corresponding environments; Augmented assembly data , representing a fused dataset of q augmented product samples with their corresponding environments, and q is greater than p; environment parameter vector is the reference vector of all parameter types, on the corresponding implementation product data, there are implementation environment parameters , p represents the number of samples, and the corresponding satisfies , which represents normal working conditions; on the corresponding enhanced product data, enhanced environment parameters , q represents the number of samples, and q is greater than p, and the corresponding satisfies , which represents the extended working conditions covering extremes and abnormalities.

3. The space equipment-oriented digital-analog fusion dual-drive digital-real fusion evaluation method according to claim 1, characterized in that, S110 includes: deploying high-precision torque sensors and angle encoders in the physical production line to collect torque-angle curves under normal working conditions as installation product data; The collected installation product data is input into the conditional generative adversarial network, and the generator of the conditional generative adversarial network learns the data distribution characteristics through a five-layer fully connected network under the condition of installation product data, and generates synthetic data under extreme parameter conditions as enhanced product data.

4. The space equipment-oriented digital-analog fusion dual-drive digital-real fusion evaluation method according to claim 1, characterized in that, S120 includes: S120-1, abstract the assembly body into a graph structure; S120-2, using installation assembly data and enhanced assembly data, learning the assembly body graph structure through graph neural network GNN, outputting deep feature vectors, splicing the output deep features and environmental parameters as the state input of the reinforcement learning agent.

5. The space equipment-oriented digital-analog fusion dual-drive digital-real fusion evaluation method according to claim 4, characterized in that, The graph neural network GNN adopts the graph attention network GAT, sets a three-layer network structure, the hidden layer dimensions are 64, 128 and 64 respectively, uses LeakyReLU activation function, aggregates neighborhood information through multi-head attention mechanism, and finally outputs a 128-dimensional deep feature vector.

6. The space equipment-oriented digital-analog fusion dual-drive digital-real fusion evaluation method according to claim 1, characterized in that, S130 includes: the dynamic data scheduler adopts a learning-based scheduling strategy, samples the real assembly data with a probability of 70% at the initial stage of training to ensure the stability of policy learning; gradually increases the sampling proportion of enhanced assembly data to 50% in the middle stage of training; increases the sampling proportion of enhanced assembly data to 70% in the later stage of training; the feedback reward signal design adopts a hierarchical structure: when using real assembly data, a basic reward of +1 is given for correct prediction, and a penalty of -1 is given for incorrect prediction; when using enhanced assembly data, an exploration reward of +0.5 is additionally given for successfully avoiding virtual faults in addition to the basic reward; at the same time, a consistency reward across data sources is set, when the agent makes the same and correct decision for different data source states with a feature similarity of more than 0.9, an additional cooperative reward of +0.25 is given.

7. The space equipment-oriented digital-analog fusion dual-drive digital-real fusion evaluation method according to claim 6, characterized in that, In S130, when the reinforcement learning agent obtains a negative reward in a state not only updates its own policy network parameters, but also calculates the gradient of the feature alignment loss and updates the trainable parameters of the graph neural network GNN based on the gradient of the feature alignment loss, wherein the alignment loss aims to increase the distance between the feature vector leading to a negative reward state and the feature vector leading to a positive reward state, representing a set of all trainable parameters of the GNN.

8. The space equipment-oriented digital-analog fusion dual-drive digital-real fusion evaluation method according to claim 1, characterized in that, In S130, the meta-course learning based on environment perception is first based on the environment parameter matrix The training set is clustered into multiple course units; in the training process, a meta-learning framework is adopted, the inner loop is executed 1000 times of fast adaptation on each course unit, the outer loop calculates the meta-gradient on the mixed validation set, and the generalization ability of the model across courses is optimized; the mixed validation set is a validation dataset composed of real assembly data and enhanced assembly data in proportion.

9. The space equipment-oriented digital-analog fusion dual-drive digital-real fusion evaluation method according to claim 8, characterized in that, The reinforcement learning agent adopts the proximal policy optimization (PPO) algorithm, the policy network and the value network both adopt a three-layer fully connected structure, the discount factor γ is set to 0.99, the generalized advantage estimation parameter λ is set to 0.95, 10 epochs are performed for each update, and the learning rate is set to 0.0001.

10. The space equipment-oriented digital-analog fusion dual-drive digital-real fusion evaluation method according to claim 2, characterized in that, In S140, the robust path outputs the prediction result based on the real assembly data , and the exploration path outputs the prediction result based on the augmented assembly data ; calculate the cosine similarity of the features of the current state with all states in the training set to obtain the uncertainty estimate U, wherein the calculation process of U includes: inputting the corresponding real assembly data of the current state into the graph neural network to obtain the corresponding deep feature vector , wherein the deep feature vector is the mathematical encoding representation of the current state , and then inputting each sample in the training set into the graph neural network to obtain the corresponding deep feature vector , calculating the cosine similarity , and finally obtaining the uncertainty estimate ; wherein the training set is the complete data set used to train the graph neural network and the reinforcement learning agent, which is composed of the real assembly data and the augmented assembly data; the final prediction output is guided by the uncertainty: when , it indicates high certainty, and is adopted; when , it indicates low certainty, and is adopted; in other cases, the weighted average is adopted.

Images