Lunar space decision method and device based on multi-task unified neural network model

Abstract

The application relates to the technical field of near-moon space decision-making, and provides a near-moon space decision-making method and device based on a multi-task unified neural network model. The method adds a task mark to an input sequence by performing task mark embedding processing on acquired multi-source heterogeneous data; a shared expression is obtained by encoding the input sequence by using a Transformer model; a target shared representation including a target task mark is obtained from a shared latent space; an output sequence is obtained by decoding the target shared representation by using an adaptive decoder; and a target near-moon space decision is determined based on the output sequence by using a target task head. The method can integrate efficient multi-task processing and accurate uncertainty quantification, and improves the efficiency and reliability of near-moon space decision-making.

Description

Technical Field

[0001] This application relates to the field of near-lunar space decision-making technology, and in particular to a near-lunar space decision-making method and apparatus based on a multi-task unified neural network model. Background Technology

[0002] In the near-lunar space environment, multi-task learning and uncertainty quantification are problems in complex decision-making tasks. Specifically, these tasks include, but are not limited to, orbital communication relay, lunar surface base station scheduling, transportation vehicle charging station selection and planning, and impactor defense task allocation. These tasks have complex logical connections and data dependencies, and require extremely high reliability in decision-making. Existing decision-making methods mainly include methods based on single-task deep learning models and methods based on deterministic multi-task learning frameworks.

[0003] Single-task deep learning models treat each specific decision-making task as an independent optimization problem and train a dedicated deep learning model for it. For example, for rail communication relay tasks, a Transformer model can be trained to predict the optimal relay path; for transport vehicle charging station selection, another independent model can be trained to predict the optimal charging location. The core implementation of these models utilizes the self-attention mechanism in the Transformer architecture. The self-attention mechanism effectively captures long-distance dependencies in the data by calculating the correlations between different positions in the input sequence, generating feature representations highly adapted to specific tasks. The advantage of this single-task approach is its clear model design objective, allowing for targeted optimization of the model architecture, hyperparameters, and training data, thus achieving better performance on specific tasks. However, the essence of this approach is "each doing its own job," with each model working independently and no information sharing between tasks. This leads to significant parameter redundancy between models and fails to leverage the inherent correlations between tasks to improve overall performance.

[0004] Methods based on single-task deep learning models suffer from the following problems: First, there is significant redundancy in model parameters. For multiple similar but independent tasks, each model needs to learn similar feature representations from scratch, which not only increases computational and storage costs but also wastes valuable training resources. Second, generalization ability is limited. When faced with tasks with limited data or imbalanced training samples, independent models often struggle to learn robust features and are prone to overfitting. Due to the lack of knowledge sharing between tasks, the model cannot acquire additional regularization information from other related tasks, making it difficult to make reliable decisions in unknown or complex new environments.

[0005] The core idea of ​​deterministic multi-task learning frameworks is to allow all related tasks to share a common underlying representation. This is typically achieved through a shared encoder and multiple task-specific decoders or prediction heads. Input data from all tasks (e.g., data from different sensors or task descriptions) is uniformly fed into this shared encoder. The encoder learns a general representation that captures common features across all tasks through self-attention. This shared representation is then passed to different prediction heads, each predicting for a specific task. For example, one prediction head might output the decision for communication relay, while another might output the decision for charging station selection. This approach, by forcing the sharing of underlying parameters, achieves implicit knowledge transfer between tasks, theoretically enabling a single model to efficiently handle multiple tasks simultaneously. The function of this approach is to leverage the correlation between tasks to regularize the model, avoid overfitting, and improve the model's generalization ability. Its implementation is based on a "shared-separated" architecture: sharing general knowledge at the lower levels and separating at higher levels to handle task-specific details.

[0006] The core flaw of deterministic multi-task learning frameworks lies in their lack of ability to quantify prediction uncertainty. These models, such as the multi-task Transformer, employ deterministic weights and parameters during training, outputting a single point estimate. This means that when making a prediction, the model cannot provide any information about the confidence or uncertainty of that prediction. In high-risk, low-tolerance applications such as near-lunar space decision-making, this deficiency is fatal. For example, in impactor defense task allocation, if the model predicts a defense strategy is effective but its confidence is low, the decision-maker may not be aware of this information, potentially leading to incorrect decisions. When the model faces new situations that differ significantly from the training data, its predictions may deviate completely, but because the uncertainty cannot be quantified, the model will still confidently output an incorrect answer, resulting in catastrophic consequences.

[0007] Meanwhile, these deterministic multi-task learning frameworks also have shortcomings in handling complex task dependencies. Although they achieve knowledge sharing between tasks through shared encoders, a single shared encoder may not be able to balance the needs of all tasks when facing multi-source heterogeneous data and complex task dependencies in near-lunar space. When the optimization objective of one task conflicts with the optimization objectives of other tasks, the shared encoder may be forced to compromise between different tasks, resulting in the performance of all tasks being affected and failing to reach the optimal level. Especially when dealing with strongly coupled tasks, existing multi-task learning techniques often struggle to find a unified adaptive mechanism to dynamically balance the relationship between shared representations and task specificity, leading to poor model performance when handling complex connections between tasks.

[0008] These shortcomings collectively demonstrate that existing technologies are insufficient to meet the comprehensive requirements of high reliability, high accuracy, and uncertainty quantification for near-lunar space decision-making. Summary of the Invention

[0009] In view of this, embodiments of this application provide a near-lunar space decision-making method and apparatus based on a multi-task unified neural network model to solve the problems of insufficient reliability and accuracy of near-lunar space decision-making in the prior art and the inability to quantify uncertainty.

[0010] A first aspect of this application provides a near-lunar space decision-making method based on a multi-task unified neural network model, including:

[0011] Acquire multi-source heterogeneous data; multi-source heterogeneous data should include at least near-lunar space data and lunar surface data;

[0012] Multi-source heterogeneous data is preprocessed and task-labeled embedding is performed to obtain labeled input sequences; task labels are used to identify different decision-making tasks in lunar space.

[0013] The shared encoder of the Transformer model obtains a shared representation based on the labeled input sequence; each shared representation includes at least the task label of this shared representation;

[0014] Self-shared potential space acquisition includes a target-shared representation with target mission labels; the target mission labels correspond to the target near-lunar space decision-making tasks;

[0015] The target shared representation is decoded using an adaptive decoder of the Transformer model to obtain the output sequence;

[0016] In a multi-task head using a Transformer model, the target task head determines the near-lunar space decision based on the output sequence.

[0017] Among them, the target mission head corresponds to the lunar space decision-making task; the shared encoder, adaptive decoder and each linear layer in the target mission head are Bayesian linear layers.

[0018] A second aspect of this application provides a near-lunar space decision-making device based on a multi-task unified neural network model, comprising:

[0019] The acquisition module is configured to acquire multi-source heterogeneous data; the multi-source heterogeneous data includes at least near-lunar space data and lunar surface data.

[0020] The preprocessing module is configured to preprocess multi-source heterogeneous data and embed task labels to obtain a labeled input sequence; the task labels are used to identify different decision-making tasks in lunar space.

[0021] The encoding module is configured to use a shared encoder of the Transformer model to obtain a shared representation based on the labeled input sequence; each shared representation includes at least the task label of the shared representation.

[0022] The matching module is configured to acquire a shared representation of the target, including the target task label, from the self-shared latent space; the target task label corresponds to the target near-lunar space decision task.

[0023] The decoding module is configured to use an adaptive decoder of the Transformer model to decode the target shared representation to obtain the output sequence;

[0024] The decision module is configured to use the target task head in the multi-task head of the Transformer model to determine the target near-lunar space decision based on the output sequence;

[0025] Among them, the target mission head corresponds to the lunar space decision-making task; the shared encoder, adaptive decoder and each linear layer in the target mission head are Bayesian linear layers.

[0026] A third aspect of this application provides an electronic device including 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 steps of the above-described method.

[0027] A fourth aspect of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described method.

[0028] The beneficial effects of the embodiments in this application compared with the prior art are:

[0029] This application's embodiments capture common features across tasks using a shared encoder, while leveraging task embedding tokens and an adaptive decoder for task-specific processing. This design allows the model to flexibly adapt to the specific output requirements of each task based on the shared feature extractor, thereby promoting knowledge transfer and collaborative optimization.

[0030] This application's embodiments are capable of processing diverse heterogeneous data collected by satellites and sensors in near-lunar space and on the lunar surface, including flight status, operational status, communication data (such as text, images, and videos), Earth-Moon space situational awareness data, and space environment monitoring data. To efficiently process this data, the model transforms different data points into tokenized embeddings. For example, numerical data is normalized and projected into a vector space, categorical data is one-hot encoded, and text data is processed through word tokenization and embedding layers. These modality-specific embeddings are then concatenated and combined with positional embeddings to form a "unified input token sequence" for processing by the Transformer's self-attention mechanism. This approach enables the model to understand complex cross-data type dependencies.

[0031] This application uses an improved Bayesian Neural Network (BNN) to model its weight distribution based on a Gaussian mixture prior, rather than a single Gaussian prior. The Gaussian Mixture Model (GMM) offers greater flexibility, enabling the modeling of multimodal or complex weight distributions, thereby capturing more subtle uncertainties than a single Gaussian prior.

[0032] This application integrates the improved BNN into the Transformer architecture, constructing a Bayesian Transformer layer. The projection matrices Q, K, and V in the Transformer's self-attention module are replaced by Bayesian linear layers. This means that the attention weights themselves have uncertainty, allowing the model to quantify the uncertainty it focuses on at different parts of the input sequence. The Feedforward Neural Network (FFN) is also constructed using Bayesian linear layers, thereby modeling uncertainty during nonlinear transformations and feature extraction. This integration gives the entire model probabilistic properties at the parameter level, enabling end-to-end uncertainty propagation and quantification.

[0033] This application's embodiments, through the integration of an improved BNN, can provide explicit confidence metrics for each prediction. This capability enables the model to apply dynamic confidence thresholds. When the model's confidence in a decision is higher than a preset threshold, the decision can be marked as "high confidence" and executed autonomously. Conversely, when the confidence is lower than the threshold, it is marked as "low confidence," triggering a manual intervention process for expert review and decision-making. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0035] Figure 1 This is a flowchart illustrating a near-lunar space decision-making method based on a multi-task unified neural network model provided in an embodiment of this application.

[0036] Figure 2 This is a flowchart illustrating the method for preprocessing and task tag embedding of multi-source heterogeneous data provided in the embodiments of this application.

[0037] Figure 3 This is a system architecture diagram for implementing the near-lunar space decision-making method based on a multi-task unified neural network model provided in the embodiments of this application.

[0038] Figure 4 This is a schematic diagram of the structure of the multi-task unified Transformer model.

[0039] Figure 5 This is a schematic diagram of the Monte Carlo inference process provided in the embodiments of this application.

[0040] Figure 6 These are ROC curves of different models when performing situational awareness information transmission.

[0041] Figure 7 These are ROC curves for different models when performing lunar base station selection.

[0042] Figure 8 These are ROC curves for different models when selecting charging stations for transport vehicles.

[0043] Figure 9 This is a graph showing the ROC curves of different models when performing impactor defense selection.

[0044] Figure 10 This is a comparison chart of the overall performance of different models for the four tasks.

[0045] Figure 11 This is a schematic diagram of a near-lunar space decision-making device based on a multi-task unified neural network model provided in an embodiment of this application.

[0046] Figure 12 This is a schematic diagram of the electronic device provided in the embodiments of this application. Detailed Implementation

[0047] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0048] The following will describe in detail, with reference to the accompanying drawings, a near-lunar space decision-making method and apparatus based on a multi-task unified neural network model according to embodiments of this application.

[0049] As mentioned above, methods based on single-task deep learning models and methods based on deterministic multi-task learning frameworks have various shortcomings. These shortcomings together indicate that existing technologies are unable to meet the comprehensive requirements of high reliability, high accuracy, and uncertainty quantification for near-lunar space decision-making.

[0050] In view of this, embodiments of this application provide a method that adds task tags to the input sequence by performing task tag embedding processing on the acquired multi-source heterogeneous data; encodes the input sequence using a Transformer model to obtain a shared representation; obtains a target shared representation including the target task tag from the shared latent space; decodes the target shared representation using an adaptive decoder to obtain an output sequence; and uses the target task head to determine the target near-lunar space decision based on the output sequence. This method integrates efficient multi-task processing and accurate uncertainty quantification, thereby improving the efficiency and reliability of near-lunar space decision-making.

[0051] Figure 1 This is a flowchart illustrating a near-lunar space decision-making method based on a multi-task unified neural network model provided in an embodiment of this application. Figure 1 As shown, the method includes the following steps:

[0052] In step S101, multi-source heterogeneous data is acquired.

[0053] Among them, multi-source heterogeneous data includes at least near-lunar space data and lunar surface data.

[0054] In step S102, the multi-source heterogeneous data is preprocessed and task label embedding is performed to obtain the labeled input sequence.

[0055] Among them, mission markers are used to identify different decision-making missions in lunar space.

[0056] In step S103, the shared encoder of the Transformer model obtains a shared representation based on the labeled input sequence.

[0057] Each shared representation includes at least the task tag of that shared representation.

[0058] In step S104, the self-shared latent space is used to obtain a target shared representation including the target task label.

[0059] Among them, the target mission marker corresponds to the target's near-lunar space decision-making mission.

[0060] In step S105, the target shared representation is decoded using the adaptive decoder of the Transformer model to obtain the output sequence.

[0061] In step S106, the target mission head in the multi-mission head of the Transformer model determines the target near-lunar space decision based on the output sequence.

[0062] Among them, the target mission head corresponds to the lunar space decision-making task; the shared encoder, adaptive decoder and each linear layer in the target mission head are Bayesian linear layers.

[0063] In some embodiments of this application, the method may be executed by a server or by a terminal device with certain processing capabilities.

[0064] In some embodiments of this application, multi-source heterogeneous data can be acquired, which includes at least near-lunar space data and lunar surface data. In one example, the multi-source heterogeneous data may include flight status, operational status, relay communication data such as text, images and videos, Earth-Moon space situational awareness data, space environment monitoring data, and scientific research data.

[0065] In some embodiments of this application, the multi-source heterogeneous data can be preprocessed and task-labeled embedding processed to obtain a labeled input sequence. The task labels are used to identify different near-lunar space decision-making tasks.

[0066] In some embodiments of this application, a shared encoder of the Transformer model can be used to obtain a shared representation based on the labeled input sequence, and the shared representation can be stored in a shared latent space. Each shared representation includes at least the task label for that shared representation.

[0067] In some implementations, a shared representation of the target, including the target mission marker, can be obtained from the shared potential space. The target mission marker corresponds to the target near-lunar space decision-making task.

[0068] The target shared representation can be decoded using the adaptive decoder of the Transformer model to obtain the output sequence. Furthermore, the target task head in the multi-task head of the Transformer model can determine the target's near-lunar spatial decision based on the output sequence.

[0069] Among them, the target mission head corresponds to the lunar space decision-making task; the shared encoder, adaptive decoder and each linear layer in the target mission head are Bayesian linear layers.

[0070] According to the technical solution provided in the embodiments of this application, task labels are added to the input sequence by performing task label embedding processing on the acquired multi-source heterogeneous data; the input sequence is encoded using the Transformer model to obtain a shared representation; the target shared representation including the target task label is obtained from the shared latent space; the target shared representation is decoded using an adaptive decoder to obtain an output sequence; and the target task head is used to determine the target near-lunar space decision based on the output sequence. This can integrate efficient multi-task processing and accurate uncertainty quantification, thereby improving the efficiency and reliability of near-lunar space decision-making.

[0071] Figure 2 This is a flowchart illustrating the method for preprocessing and task label embedding of multi-source heterogeneous data provided in an embodiment of this application. Figure 2 As shown, the method includes the following steps:

[0072] In step S201, the multi-source heterogeneous data is preprocessed.

[0073] In step S202, the preprocessed multi-source heterogeneous data is tokenized to obtain an initial sequence.

[0074] In step S203, task tags are embedded into each initial sequence to obtain the tagged input sequence.

[0075] In some embodiments of this application, when preprocessing and embedding task tags into multi-source heterogeneous data, the multi-source heterogeneous data can be preprocessed first, then tokenized to obtain an initial sequence. Finally, task tags are embedded into each initial sequence to obtain a labeled input sequence.

[0076] In other words, to effectively process this diverse data using a Transformer-based architecture, each distinct data point—whether a scalar value, text report, or coordinate tuple—is transformed into a tokenized embedding. Specifically, numerical data is normalized and linearly projected into a dense vector space; categorical data is encoded before embedding; and textual data, such as communication data, is processed through word tokenization and embedding layers. These modality-specific embeddings are then concatenated and combined with positional embeddings to form a “unified input token sequence” for processing by the Transformer’s self-attention mechanism. This approach integrates different types of data into a unified feature space, enabling the Transformer to capture their complex interdependencies, thereby achieving a more comprehensive situational understanding and providing comprehensive information support for subsequent task node selection.

[0077] In some embodiments of this application, the shared encoder comprises L identical layers, each of which performs a position feedforward network after performing multi-head self-attention; L is a positive integer.

[0078] Multi-head self-attention determines the weight distribution based on a Gaussian mixture model, where the Gaussian mixture model is... ,in, For weight distribution, For random variables, This represents the normal distribution. The number of Gaussian components in the mixture model. This is the sequence number of the mixed part. Represents the mixing coefficient. and Let represent the mean and variance of the Gaussian components, respectively.

[0079] In some embodiments of this application, the weights of the Bayesian linear layer are calculated using the formula [formula missing]. Determined; among them, For weight values, The average of the weights. The variance of the weights, Let be a random variable sampled from a standard normal distribution.

[0080] Furthermore, during each forward propagation, the bias raw sampling of the Bayesian linear layer adopts the formula... Determined; among them, These are the biased raw sampled values. The mean of the biased raw sampled values. This represents the variance of the biased raw sample values.

[0081] The output of the Bayesian linear layer is ;in, The output of the Bayesian linear layer, For the input of the Bayesian linear layer, This is the weight matrix.

[0082] In some embodiments of this application, Monte Carlo sampling is used in the inference phase of the Transformer model to estimate prediction inaccuracies.

[0083] Monte Carlo sampling includes:

[0084] For a given input Model execution Each forward pass involves adjusting the new network weights. Posterior sampling Different Monte Carlo predictions ;in, The values ​​are predicted, and MT-UT is the model designation. Uncertainty decomposition is performed on each Monte Carlo prediction to obtain the expected calibration error and the maximum calibration error.

[0085] In some embodiments of this application, the Transformer model is trained iteratively using an adaptive task-consistent dynamic loss weighting mechanism.

[0086] The adaptive task-consistent dynamic loss weighting mechanism includes: calculating the loss function for each decision task; and using the formula... Calculate the prediction weights for each loss function; where, For the first The loss function at time... The predicted weights, It is an update function. Based on the The loss function at time... The current weight, the first The loss gradient of the i-th task and the i-th task The historical performance of each task is determined; the total loss function of the model is obtained by weighted summation of each loss function based on the prediction weights.

[0087] In some embodiments of this application, the determined target near-lunar space decision includes at least a predicted confidence level. If the predicted confidence level is determined to be lower than the dynamic confidence level threshold, the decision result can be discarded or a prompt message can be generated.

[0088] The dynamic confidence threshold is determined as follows: obtain the prediction results of historical tasks within a preset time range; calculate the variance of the prediction results of historical tasks; determine the dynamic confidence threshold based on the variance; the dynamic confidence threshold is negatively correlated with the equation.

[0089] Figure 3This is a system architecture diagram for implementing the near-lunar space decision-making method based on a multi-task unified neural network model provided in the embodiments of this application. For example... Figure 3 As shown, the system includes a data preprocessing and unified representation module and a multi-task unified Transformer model.

[0090] The data preprocessing and unified representation module includes a multi-source heterogeneous data input module, a data preprocessing module, a tokenization and embedding initialization sequence X module, and a unified input sequence X'. The multi-source heterogeneous data input preprocessing module preprocesses the data, which is then processed by the tokenization and embedding initialization sequence X module to obtain the initialization sequence X. The unified input sequence X' embeds a task token, i.e., a task label T, into the initialization sequence X to obtain the unified input sequence X'.

[0091] The multi-task unified Transformer model comprises a shared encoder, an adaptive decoder, a Monte Carlo inference unit, and a confidence threshold determination unit. The shared encoder is implemented using a Bayesian network. The adaptive decoder can output sequences corresponding to different tasks, including situational awareness information transmission, lunar base station planning, transport vehicle charging planning, and impactor defense task allocation. The Monte Carlo inference module enables uncertainty quantification. When the confidence threshold determination result is high, task decisions can be executed automatically; when the confidence threshold is low, human intervention can be requested.

[0092] Figure 4 This is a schematic diagram of the structure of a multi-task unified Transformer model. For example... Figure 4 As shown, the multi-task unified Transformer model aims to process heterogeneous data from multiple sources while performing interdependent assignment tasks. Its core innovation lies in facilitating shared representations while maintaining task-specific processing to achieve collaborative learning. The architecture includes a shared encoder, an adaptive multi-task decoder, and a specialized task-specific output head. Its input is a unified sequence of labels. This represents embedded data points, such as state readings. To enable task-specific conditioned reflexes and unified processing, task embedding tags are introduced. ,in , , and These correspond to one of the four evaluation tasks. Therefore, the effective input sequence of the Transformer encoder becomes... .

[0093] The shared Transformer encoder contains L identical layers for processing the combined sequence. Each layer performs multi-head self-attention, followed by a positional feedforward network. The self-attention mechanism is the core of the Transformer model; it computes a weighted sum of the input values, with weights determined by the similarity between the query and the key. ,in To scale point attention, , and They are obtained by performing different linear transformations on the encoded input features H. , , , Here is the neural network parameter matrix; softmax is the normalized exponential function. This represents the dimension of the key vector.

[0094] Multi-head attention mechanisms allow a model to focus on information from different locations within multiple representation subspaces. For Attention head, output calculation as follows ,in, ; , and Let be the specific weight matrix of the i-th attention head. It is a projection matrix that integrates the outputs of all attention heads. Concat is a string concatenation function, and h is the total number of attention heads.

[0095] in, This represents the learnable parameter matrix. After self-attention, each layer uses positional FFN: The encoder employs residual connections and layer normalization in both sublayers, which enhances training stability and performance.

[0096] The shared decoder transforms various input labels into a rich, context-aware, unified latent representation. The shared representation Z essentially captures the correlations and dependencies between patterns and tasks, thus facilitating knowledge transfer. From the shared latent space, outputs corresponding to task embedding labels in Z are extracted and fed into a task-specific decoder or output layer. For each task j, a dedicated output head processes the corresponding task representation z. j This modular design allows the model to flexibly adapt to the specific output requirements of each task, while sharing a common feature extractor, thus addressing various problems in intelligent decision-making. For assignment tasks, each output head acts as a classifier, used to select among multiple task nodes.

[0097] In high-risk tasks, particularly in decision-making, it is crucial to communicate prediction confidence and potential error, which traditional deterministic neural networks cannot provide because they can only output point estimates. Bayesian neural networks (BNNs) overcome this limitation by modeling weights as probability distributions rather than fixed values, thereby quantifying uncertainty.

[0098] Standard BNNs employ a simple Gaussian prior for weight distribution. However, the complex environment and high-dimensional data characteristics near the lunar surface can lead to a highly non-Gaussian posterior distribution of network weights. To address this challenge, the improved BNN in this model uses a Gaussian mixture prior as its weight distribution. Gaussian mixture models (GMMs) offer greater flexibility in modeling multimodal or complex weight distributions, thereby capturing more subtle uncertainties compared to a single Gaussian prior.

[0099] For the weight parameters, Gaussian mixture model weights can be used instead of single Gaussian model weights: ,in, For weight distribution, For random variables, This represents the normal distribution. The number of Gaussian components in the mixture model. The number of mixed parts, Represents the mixing coefficient. and Let represent the mean and variance of the Gaussian components, respectively.

[0100] Hybrid prioritization enables models to capture more complex multimodal posterior distributions, enhancing their ability to model weights with true uncertainty. This is particularly valuable for handling potential multimodal weight distributions arising from heterogeneous data sources and complex task features.

[0101] To implement Bayesian inference in neural networks, variational inference (VI) can be used. In the model implementation, each linear layer in the Transformer (including the linear transformation and feedforward network) is replaced by a Bayesian linear layer. For such a Bayesian linear layer, the weights and biases are raw modeled as independent Gaussian distributions, where σ is guaranteed to be positive. During each forward pass, the weights and biases are raw sampled from their posterior distributions using parameterization techniques: , ,in For weight values, The average of the weights. The variance of the weights, Let be a random variable sampled from a standard normal distribution. These are the biased raw sampled values. The mean of the biased raw sampled values. This represents the variance of the biased raw sample values.

[0102] in, Following a standard normal distribution, the output of the linear layer is calculated as follows: ,in, The output of the Bayesian linear layer, For the input of the Bayesian linear layer, This is the weight matrix.

[0103] The multi-task unified neural network model provided in this application integrates the improved Bayesian neural network into the Transformer architecture to construct a Bayesian Transformer layer.

[0104] The projection matrices Q, K, V, and output projection matrix W in the Transformer's self-attention module. Q W K W V W O It is replaced by a Bayesian linear layer. This means that the attention weights themselves have uncertainty, allowing the model to quantify the uncertainty of its focus on different parts of the input sequence. For example, the attention weights are now randomized, and their inherent uncertainty can be propagated through sampling. This allows the model to reflect ambiguous or conflicting input information through an uncertain initial attention mechanism.

[0105] Within each Transformer layer, a positional FFN is constructed using a Bayesian linear layer. This structure enables the modeling of uncertainties in the nonlinear transformation of the FFN and the feature extraction process. By incorporating Bayesian elements into these core components, the entire model acquires probabilistic properties at the parameter level, which facilitates end-to-end uncertainty propagation and quantification.

[0106] In the inference phase, Monte Carlo (MC) sampling can be used to estimate prediction uncertainty. Figure 5 This is a schematic diagram of the Monte Carlo inference process provided in an embodiment of this application. For example... Figure 5 As shown, for a given input Model execution Each forward pass involves adjusting the new network weights. Posterior sampling, which will produce Different predictions From these Monte Carlo examples, the final predicted Expected Calibration Error (ECE) and Maximum Calibration Error (MCE) can be estimated.

[0107] Here, ECE refers to the expected value that measures the difference between the model's predicted probability and the actual accuracy. The predicted probabilities are divided into buckets, and the absolute difference between the average predicted probability and the accuracy within each bucket is calculated. Then, a weighted average is performed based on the sample size. Where B is the number of buckets, n i is the number of samples in bucket i, acc(i) is the accuracy of bucket i, and conf(i) is the average predicted probability (confidence level) of bucket i.

[0108] MCE represents the maximum absolute difference between the predicted probability and accuracy across all buckets. .

[0109] To optimize model performance in multi-task learning, the Adaptive Task-Consistent Dynamic Loss Weighting (ATCDLW) mechanism can be introduced. While multi-task learning promotes knowledge sharing, it often faces the problem of negative transfer, where learning from one task may interfere with learning from another. ATCDLW aims to dynamically adjust the weights of the loss for each task to achieve balanced learning and collaborative optimization across tasks. Traditional MTL loss functions are typically simple weighted sums of the losses from each task: ,in w is the loss function for the j-th task. j These are its corresponding static weights. The core of ATCDLW lies in making w j These parameters become dynamically learnable and are adjusted based on the training progress, gradient conflicts, and uncertainties of the task.

[0110] The ATCDLW mechanism dynamically adjusts weights by monitoring the convergence speed and gradient conflict of each task. Specifically, the mechanism introduces a learnable parameter λ for each task j. j The final weight wj of the task loss can be calculated, for example, based on the instantaneous rate of change of the task loss or the magnitude of its gradient relative to other tasks. Here, f is an update function that considers the current weights, the loss gradient of task j, and its historical performance. In this way, the model can automatically allocate more learning resources to tasks that are difficult to converge or whose gradients conflict significantly with those of other tasks, thereby effectively mitigating negative transfer and ensuring that all tasks receive sufficient and coordinated training. This ultimately improves the overall generalization ability and robustness of the model, which is crucial for learning how to intelligently schedule tasks across multiple task nodes.

[0111] The improved BNN in the model not only quantifies uncertainty, but more importantly, it applies these uncertainty measures to dynamic confidence threshold adjustments, significantly reducing the misjudgment rate in critical decisions. In high-risk scenarios, high accuracy is crucial, but identifying and avoiding erroneous predictions is equally critical.

[0112] For each task prediction, BNN provides the mean and variance of the prediction distribution. The model uses the prediction variance as an inverse indicator of confidence: the larger the variance, the lower the confidence. A task-specific or global dynamic confidence threshold is defined. When the model's prediction confidence (e.g., the mean of the predicted probability reaches a certain high value, while the prediction variance is below a certain threshold) falls below the preset dynamic threshold, the model does not directly give a final judgment, but instead adopts a strategy of "abstaining" or "requesting manual intervention".

[0113] By dynamically adjusting these thresholds, the accuracy of predictions can be flexibly balanced with the risk of misjudgment based on the criticality of the task and the risk tolerance. In critical scenarios such as impactor defense task allocation, even with high model accuracy, a single incorrect judgment (e.g., selecting an inappropriate impactor node) can lead to significant losses. In such cases, stricter confidence thresholds can be set, prioritizing more human intervention over low-confidence automated decisions. This significantly improves the safety and reliability of the model in practical operation.

[0114] The technical solution of this application provides a novel model that can simultaneously solve the multi-task learning problem and accurately quantify and predict uncertainties. Specifically, it addresses four interrelated tasks commonly encountered in near-lunar space: communication information transmission, lunar surface base station relay selection, transport vehicle charging path planning, and impactor defense task allocation. The specific objective is to intelligently select the most suitable node from four types of facilities—communication and navigation satellites, lunar surface base stations, charging stations, and impactors—to execute specific tasks based on real-time conditions, in order to cope with the complex and dynamic environment.

[0115] Situational awareness information transmission: When threats such as meteorites and space debris are identified, the model needs to intelligently select the most suitable communication and navigation satellites to efficiently and reliably transmit the situational awareness information sent by the situational awareness satellites to the lunar surface base station.

[0116] Lunar base station relay selection: Based on the current state of communication satellites, lunar base stations, and space environment, the model needs to select a suitable lunar base station to issue instructions to the lunar transport vehicle.

[0117] Transport vehicle charging route planning: After receiving instructions from the base, the model needs to decide on the best route for the lunar rover, including selecting a suitable charging station to ensure that the transport vehicle has sufficient power.

[0118] Impactor defense task allocation: When a meteorite strikes, the model will intelligently select the most suitable impactor to perform the defense task based on an accurate assessment of the meteorite threat, in order to maximize the success rate of the defense while minimizing resource consumption and collateral damage.

[0119] These tasks are highly interdependent, collectively forming a holistic and interconnected scenario decision-making task. The entire model operates as a scheduling loop: from the moment a situational awareness satellite detects a meteorite threat, the MT-UT model initiates the scheduling decision-making process. First, it selects a communication and navigation satellite for information transmission; then, it selects a lunar surface base station for command; next, the vehicle selects a charging station for charging and avoidance; and finally, it selects the impactor for defense. Each step is an optimal scheduling decision made by the model based on the current state, data, and learned patterns of all nodes. This inherent interdependence underscores the necessity of multi-task learning methods. Core objectives include:

[0120] First, this application overcomes the problems of parameter redundancy and knowledge silos in existing single-task models when handling multiple tasks. The embodiments of this application construct a unified multi-task learning architecture, enabling different decision tasks to share key feature representations, thereby reducing unnecessary model parameters and improving computational efficiency. Simultaneously, through knowledge sharing between tasks, the embodiments of this application aim to enhance the model's generalization ability when facing tasks with limited data, allowing it to acquire useful information from other related tasks and improve overall performance.

[0121] Second, this application addresses the limitation of existing deterministic neural networks in quantifying prediction uncertainty. For high-risk, low-fault-tolerance applications such as near-lunar space decision-making, this application aims to provide a technical solution capable of accurately assessing prediction confidence. By introducing a probabilistic inference mechanism into the model, this application provides an uncertainty range for each decision output, allowing decision-makers to assess risk based on the degree of uncertainty, thereby avoiding erroneous decisions caused by blindly relying on point estimation results.

[0122] Third, this application aims to overcome the shortcomings of traditional Bayesian neural networks and multi-task learning frameworks in handling complex task correlations and data distributions. The embodiments of this application aim to propose a more advanced probabilistic modeling method that can accurately capture the complex weight distributions caused by multi-source heterogeneous data from near-lunar space, thereby more realistically reflecting the model's predictive uncertainty. Simultaneously, the embodiments of this application aim to design a task-adaptive mechanism that enables the model to dynamically balance the relationship between shared representations and task specificity, thereby better handling interdependent, strongly coupled tasks and achieving optimized multi-task collaborative performance.

[0123] To verify the technical effects of the embodiments of this application, the following experiment was designed:

[0124] The experiment used lunar satellite data provided by the Deep Space Exploration Laboratory and publicly available lunar surface data from NASA, and was labeled for four core assigned tasks. The training, validation, and test sets were divided in an 8:1:1 ratio. The comparison models include:

[0125] Traditional machine learning methods: using ensemble learning models such as random forests;

[0126] Single-Task Transformer: A Transformer model is trained independently for each task, without knowledge sharing between tasks;

[0127] Deterministic Neural Network (NN): A multi-task feedforward neural network that does not contain Bayesian layers or uncertainty quantization;

[0128] MT-UT+BNN (Proposed Model): The model proposed in this application integrates a multi-task unified Transformer architecture and an improved Bayesian neural network.

[0129] The experimental results are as follows:

[0130] Four different models were tested, and the accuracy, recall and F1 score of the models on the comprehensive task are shown in Table 1.

[0131] Table 1 Model Performance Comparison

[0132]

[0133] Table 1 shows that the MT-UT BNN model outperforms all baseline methods in terms of precision and recall across all four assignment tasks. Compared to traditional machine learning methods, deep learning models demonstrate stronger feature learning capabilities, while the MT-UT BNN model achieves outstanding results with an F1 score of 0.87. Furthermore, based on precision and recall, ROC curves for different models on the four tasks can be plotted, such as... Figure 6-9 As shown.

[0134] Figure 6 These are ROC curves of different models when performing situational awareness information transmission. Figure 7 These are ROC curves for different models when performing lunar base station selection. Figure 8 These are ROC curves for different models when selecting charging stations for transport vehicles. Figure 9 This is a graph showing the ROC curves of different models when performing impactor defense selection.

[0135] Figure 6-9 The receiver operating characteristic (ROC) curves of the improved model and other models are presented for performing four tasks. The area under the curve (AUC) value is used as a key indicator to evaluate the overall discrimination ability of the model. A value close to 1 indicates that the model performs well and has a stronger ability to distinguish between positive and negative samples. Based on the performance of each model on different tasks, the relative difficulty of the tasks can be inferred.

[0136] Situational Awareness Messaging: All models performed relatively poorly on this task, indicating that it may be the most challenging task, involving complex contextual understanding and information processing.

[0137] Transporter Charging Scheduling: Medium difficulty, with all models performing at an average level.

[0138] Lunar Base Station Scheduling: Relatively easy, most models perform well on this task.

[0139] Impactor Defense Task Assignment: Although the task is complex, advanced models (especially MT-UT+BNN) perform well on this task, indicating that this type of structured decision-making task is more suitable for advanced neural network architectures.

[0140] Therefore, when considering the results, it is necessary to take into account the impact of task characteristics on model performance. Different tasks have different requirements for model capabilities; for example, situational awareness tasks are challenging for all models. The combination of multi-task learning and uncertainty modeling enables MT-UT+BNN to perform well on various tasks, especially in complex defense task allocation. In addition to the ROC curves of different models under the four tasks, the comprehensive performance of the four models for the four tasks can also be obtained, such as... Figure 10 As shown.

[0141] Figure 10 This is a comparison chart of the overall performance of different models for the four tasks. For example... Figure 10As shown, model complexity is positively correlated with performance. From traditional machine learning to MT-UT+BNN, as model complexity increases, performance also improves accordingly, with the MT-UT BNN model performing the best. The performance improvement compared to single-task Transformers and deterministic NNs mainly stems from its unified multi-task architecture. Specifically, MT-UT effectively captures common features across tasks through its shared encoder, while achieving task-specific processing through task embedding tokens and adaptive decoders, thereby promoting excellent knowledge transfer and collaborative optimization. Furthermore, compared to DNNs, MT-UT BNNs also show a lower false positive rate. The results are shown in Table 2.

[0142] Table 2 Comparison of MT-UT BNN and DNN in uncertainty quantification

[0143]

[0144] Table 2 illustrates the substantial benefits of the improved BNN integration. Deterministic neural networks cannot directly quantify uncertainty, cannot compute expected calibration error (ECE) and maximum calibration error (MCE), and tend to have a higher misclassification rate. In contrast, the MT-UT BNN model exhibits lower ECE and MCE values ​​across all tasks, indicating that its prediction confidence is calibrated very well. The results show that when the MT-UT BNN model selects a high-confidence representation for a specific task node, the probability of that selection being correct is indeed very high; conversely, when it represents a low-confidence representation, the probability of error is higher.

[0145] Even without applying a confidence threshold, the inherent misclassification rate of MT-UT BNN is already lower than that of deterministic neural networks, thanks to its superior prediction accuracy. When the confidence threshold is increased to 80% and 90%, the misclassification rate drops to 4.1% and 2.9%, respectively. The results show that the model can identify and avoid nearly 90% of node selection predictions, thereby reducing errors in high-risk decision-making.

[0146] All of the above-mentioned optional technical solutions can be combined in any way to form the optional embodiments of this application, and will not be described in detail here.

[0147] The following are embodiments of the apparatus described in this application, which can be used to execute the embodiments of the method described in this application. For details not disclosed in the apparatus embodiments of this application, please refer to the embodiments of the method described in this application.

[0148] Figure 11 This is a schematic diagram of a near-lunar space decision-making device based on a multi-task unified neural network model provided in an embodiment of this application. Figure 11 As shown, the device includes:

[0149] The acquisition module 1101 is configured to acquire multi-source heterogeneous data; the multi-source heterogeneous data includes at least near-lunar space data and lunar surface data.

[0150] The preprocessing module 1102 is configured to preprocess multi-source heterogeneous data and embed task labels to obtain a labeled input sequence; the task labels are used to identify different decision-making tasks in lunar space.

[0151] Encoding module 1103 is configured to use a shared encoder of the Transformer model to obtain a shared representation based on the labeled input sequence; each shared representation includes at least the task label of the shared representation.

[0152] Matching module 1104 is configured to acquire a target shared representation including the target task label in the self-shared latent space; the target task label corresponds to the target near-lunar space decision task.

[0153] Decoding module 1105 is configured to use an adaptive decoder of the Transformer model to decode the target shared representation to obtain an output sequence.

[0154] Decision module 1106 is configured to use the target task head in the multi-task head of the Transformer model to determine the target near-lunar space decision based on the output sequence.

[0155] Among them, the target mission head corresponds to the lunar space decision-making task; the shared encoder, adaptive decoder and each linear layer in the target mission head are Bayesian linear layers.

[0156] According to the technical solution provided in the embodiments of this application, task labels are added to the input sequence by performing task label embedding processing on the acquired multi-source heterogeneous data; the input sequence is encoded using the Transformer model to obtain a shared representation; the target shared representation including the target task label is obtained from the shared latent space; the target shared representation is decoded using an adaptive decoder to obtain an output sequence; and the target task head is used to determine the target near-lunar space decision based on the output sequence. This can integrate efficient multi-task processing and accurate uncertainty quantification, thereby improving the efficiency and reliability of near-lunar space decision-making.

[0157] In some implementations, preprocessing and task tag embedding of multi-source heterogeneous data includes: preprocessing the multi-source heterogeneous data; tokenizing the preprocessed multi-source heterogeneous data to obtain an initial sequence; and embedding task tags into each initial sequence to obtain a tagged input sequence.

[0158] In some implementations, the shared encoder comprises L identical layers, each of which performs a position feedforward network after multi-head self-attention; L is a positive integer; the multi-head self-attention determines the weight distribution based on a Gaussian mixture model, where the Gaussian mixture model is... ,in, For weight distribution, For random variables, This represents a normal distribution. The number of Gaussian components in the mixture model. This is the sequence number of the mixed part. Represents the mixing coefficient. and Let represent the mean and variance of the Gaussian components, respectively.

[0159] In some implementations, the weights of the Bayesian linear layer are determined using a formula during each forward pass. Determined; among them, For weight values, The average of the weights. The variance of the weights, Let be random variables sampled from a standard normal distribution; and in each forward pass, the biased raw sampling of the Bayesian linear layer uses the formula Determined; among them, These are the biased raw sampled values. The mean of the biased raw sampled values. The variance of the biased raw sampled values ​​is given; the output of the Bayesian linear layer is... ;in, The output of the Bayesian linear layer, For the input of the Bayesian linear layer, This is the weight matrix.

[0160] In some implementations, the inference phase of the Transformer model uses Monte Carlo sampling to estimate prediction inaccuracies; Monte Carlo sampling includes: for a given input Model execution Each forward pass involves adjusting the new network weights. Posterior sampling Different Monte Carlo predictions ;in, The predicted value is MT-UT, which is the model name. Uncertainty decomposition is performed on each Monte Carlo prediction to obtain the expected calibration error and the maximum calibration error.

[0161] In some implementations, the Transformer model uses an adaptive task-consistent dynamic loss weighting mechanism for iterative training; this mechanism includes: calculating the loss function for each decision task; and using the formula... Calculate the prediction weights for each loss function; where, For the first The loss function at time... The predicted weights, It is an update function. Based on the The loss function at time... The current weight, the first The loss gradient of the i-th task and the i-th task The historical performance of each task is determined; the total loss function of the model is obtained by weighted summation of each loss function based on the prediction weights.

[0162] In some implementations, the determined target near-lunar space decision includes at least a prediction confidence level; in response to determining that the prediction confidence level is lower than a dynamic confidence threshold, the current decision result is discarded or a prompt message is generated; wherein, the dynamic confidence threshold is determined in the following way: obtaining historical task prediction results within a preset time range; calculating the variance of historical task prediction results; determining the dynamic confidence threshold based on the variance; the dynamic confidence threshold is negatively correlated with the equation.

[0163] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0164] Figure 12 This is a schematic diagram of the electronic device provided in an embodiment of this application. For example... Figure 12 As shown, the electronic device 12 of this embodiment includes: a processor 1201, a memory 1202, and a computer program 1203 stored in the memory 1202 and executable on the processor 1201. When the processor 1201 executes the computer program 1203, it implements the steps in the various method embodiments described above. Alternatively, when the processor 1201 executes the computer program 1203, it implements the functions of each module / unit in the various device embodiments described above.

[0165] Electronic device 12 can be a desktop computer, laptop, handheld computer, cloud server, or other electronic device. Electronic device 12 may include, but is not limited to, processor 1201 and memory 1202. Those skilled in the art will understand that... Figure 12 This is merely an example of electronic device 12 and does not constitute a limitation on electronic device 12. It may include more or fewer components than shown, or different components.

[0166] The processor 1201 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0167] The memory 1202 can be an internal storage unit of the electronic device 12, such as a hard disk or RAM of the electronic device 12. The memory 1202 can also be an external storage device of the electronic device 12, such as a plug-in hard disk, Smart Media Card (SMC), Secure Digital (SD) card, FlashCard, etc., equipped on the electronic device 12. The memory 1202 can also include both internal and external storage units of the electronic device 12. The memory 1202 is used to store computer programs and other programs and data required by the electronic device.

[0168] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0169] If an integrated module / unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program may include computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. A computer-readable medium may include: any entity or device capable of carrying computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0170] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A near-lunar space decision-making method based on a multi-task unified neural network model, characterized in that, include: Acquire multi-source heterogeneous data; the multi-source heterogeneous data includes at least near-lunar space data and lunar surface data; The multi-source heterogeneous data is preprocessed and task-labeled embedding is performed to obtain the labeled input sequence; The task markers are used to identify different near-lunar space decision-making tasks; A shared encoder using the Transformer model obtains a shared representation based on the labeled input sequence; each shared representation includes at least the task label of this shared representation; The self-shared potential space acquisition includes a target-shared representation with target task labels; the target task labels correspond to the target near-lunar space decision-making task. The target shared representation is decoded using an adaptive decoder based on the Transformer model to obtain the output sequence; The target task head in the multi-task head of the Transformer model determines the target's near-lunar space decision based on the output sequence; The target mission head corresponds to the target near-lunar space decision-making mission; the shared encoder, the adaptive decoder, and each linear layer in the target mission head are all Bayesian linear layers.

2. The method of claim 1, wherein, The preprocessing and task labeling embedding of the multi-source heterogeneous data includes: The multi-source heterogeneous data is preprocessed; The preprocessed multi-source heterogeneous data is tokenized to obtain the initial sequence; The task tags are embedded into each initial sequence to obtain the labeled input sequence.

3. The method of claim 1, wherein, The shared encoder contains L identical layers, each of which performs a position feedforward network after multi-head self-attention; L is a positive integer; The multi-head self-attention method determines the weight distribution based on a Gaussian mixture model, which is: ,in, For weight distribution, For random variables, This represents the normal distribution. The number of Gaussian components in the mixture model. The sequence number represents the mixed portion. Represents the mixing coefficient. and Let represent the mean and variance of the Gaussian components, respectively.

4. The method of claim 1, wherein, During each forward pass, the weights of the Bayesian linear layer are calculated using the formula... Determined; among them, For weight values, The mean of the weights. The variance of the weights, Let be a random variable sampled from a standard normal distribution; Furthermore, during each forward propagation, the biased raw sampling of the Bayesian linear layer adopts the formula... Determined; among them, These are biased raw sampled values. The mean of the biased raw sampled values. The variance of the biased raw sample values; The output of the Bayesian linear layer is ; wherein, is the output of the Bayesian linear layer, is the input of the Bayesian linear layer, is a weight matrix.

5. The method of claim 1, wherein, The inference phase of the Transformer model uses Monte Carlo sampling to estimate prediction inaccuracies; The Monte Carlo sampling includes: For a given input , the model performs forward passes, each with new network weights and posterior sampling yields different Monte Carlo predictions ; where, is the prediction, and MT-UT is the model's moniker; Uncertainty decomposition calculations are performed on each Monte Carlo prediction to obtain the expected calibration error and the maximum calibration error.

6. The method of claim 1, wherein, The Transformer model is trained iteratively using an adaptive task-consistent dynamic loss weighting mechanism. The adaptive task-consistent dynamic loss weighting mechanism includes: Calculate the loss function for each decision task; Use formula Calculate the prediction weights for each loss function; where, For the first The loss function at time... The predicted weights, It is an update function. Based on the The loss function at time... The current weight, the first The loss gradient of the i-th task and the i-th task The historical performance of each task is determined; The total loss function of the model is obtained by weighted summation of each loss function based on the prediction weights.

7. The method of claim 1, wherein, Target-specific near-lunar space decision-making should include at least the prediction confidence level; In response to determining that the predicted confidence level is lower than the dynamic confidence threshold, the current decision result is discarded or a prompt message is generated; The dynamic confidence threshold is determined in the following manner: Obtain historical task prediction results within a preset time range; Calculate the variance of the historical task prediction results; The dynamic confidence threshold is determined based on the variance; the dynamic confidence threshold is negatively correlated with the variance.

8. A near-lunar space decision-making device based on a multi-task unified neural network model, characterized in that, include: The acquisition module is configured to acquire heterogeneous data from multiple sources. The multi-source heterogeneous data includes at least near-lunar space data and lunar surface data; The preprocessing module is configured to preprocess the multi-source heterogeneous data and perform task label embedding processing to obtain a labeled input sequence; The task markers are used to identify different near-lunar space decision-making tasks; The encoding module is configured to use a shared encoder of the Transformer model to obtain a shared representation based on the labeled input sequence; each shared representation includes at least the task label of the shared representation. The matching module is configured to acquire a target shared representation, including target task labels, from a self-shared latent space; the target task labels correspond to the target near-lunar space decision-making task. The decoding module is configured to decode the target shared representation using an adaptive decoder of the Transformer model to obtain an output sequence; The decision module is configured to use the target task head in the multi-task head of the Transformer model to determine the target near-lunar space decision based on the output sequence; The target mission head corresponds to the target near-lunar space decision-making mission; the shared encoder, the adaptive decoder, and each linear layer in the target mission head are all Bayesian linear layers.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1-9. When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 7.

Images