Language large model-based space infrared target characteristic rapid prediction method

Abstract

The invention discloses a spatial infrared target characteristic rapid prediction method based on a large language model, and the method comprises the steps: carrying out the feature inversion of unknown target parameters through constructing a causal reasoning feature extraction frame and fusing the expert knowledge, information reports and other multi-source information; and establishing a target characteristic cross-domain migration prediction model, realizing cross-domain generation of characteristic images, and finally forming a complex space environment-oriented target characteristic rapid prediction system. According to the method, the problems of target model parameter missing, characteristic cross-time domain dynamic evolution, obvious multi-working-condition scene characteristic difference and the like can be effectively solved, and the infrared target characteristic prediction efficiency and accuracy are greatly improved.

Description

Technical Field

[0001] The present invention belongs to the field of space technology, and particularly relates to a method for rapidly predicting the characteristics of space infrared targets based on a large language model. Background Art

[0002] With the rapid development of space technology, tasks such as on-orbit maintenance and space debris cleaning in space are increasing day by day. These tasks usually require autonomous identification and characteristic analysis of non-cooperative targets. Data-driven autonomous identification methods need to pre-obtain real active and passive characteristic data samples of the target for template construction or training set generation. However, there are the following technical problems in practical applications:

[0003] (1) Difficulty in obtaining real data: It is extremely difficult to obtain real measured infrared data of space targets, especially infrared characteristic data under different environmental conditions, with high cost and long time consumption.

[0004] (2) Limitations of traditional physical modeling methods: The current mainstream methods for generating target characteristic data mainly rely on physical mechanism modeling, which requires accurate target parameters and environmental parameters, and the modeling process is complex, consuming a large amount of computing resources. These methods have obvious defects:

[0005] (1) Time lag: The modeling process takes a long time and is difficult to meet the requirements of rapid response;

[0006] (2) Weak cross-scene generalization ability: Re-modeling is required for new environmental conditions;

[0007] (3) Strong parameter dependence: It is difficult to accurately model unknown targets;

[0008] (4) Dynamic changes in target characteristics: The characteristics of space targets will change dynamically with factors such as time, orbital position, attitude, and solar irradiation angle. Traditional static modeling is difficult to accurately reflect this dynamic characteristic;

[0009] (5) Significant differences in multi-condition scenarios: The infrared characteristics under different environmental conditions show significant differences, and existing methods are difficult to efficiently achieve cross-scene characteristic prediction.

[0010] In response to the above technical problems, there are currently some solutions, such as image conversion methods and domain adaptation methods based on deep learning. However, these methods often require a large amount of labeled data and are difficult to effectively utilize domain expert knowledge and physical laws, and cannot meet the requirements of rapid algorithm update and iteration. Therefore, there is an urgent need to develop a new method that can integrate multi-source information and achieve rapid prediction of target infrared characteristics. Summary of the Invention

[0011] To overcome the deficiencies of the prior art, the present invention provides a method for rapidly predicting the characteristics of space infrared targets based on a large language model. By constructing a causal inference feature extraction framework, multi-source information such as expert knowledge and intelligence reports is fused to perform feature inversion on unknown target parameters. A cross-domain transfer prediction model for target characteristics is established to achieve cross-domain generation of characteristic images, and finally a rapid prediction system for target characteristics facing complex space environments is formed. The present invention can effectively solve problems such as the lack of target model parameters, the dynamic evolution of characteristics across time domains, and the significant differences in multi-condition scene features, greatly improving the efficiency and accuracy of infrared target characteristic prediction.

[0012] The technical solution adopted by the present invention to solve its technical problems is as follows:

[0013] Step 1: Multi-source data acquisition and preprocessing;

[0014] Step 2: Causal feature extraction and knowledge representation based on a large language model;

[0015] Step 3: Unsupervised domain adaptive image semantic segmentation;

[0016] Step 4: Infrared characteristic assignment and gradient generation based on a large language model;

[0017] Step 5: Model calibration and accuracy optimization based on measured data.

[0018] Preferably, the specific content of Step 1 is as follows:

[0019] Step 1-1: Source domain image data: Obtain known source domain images ;

[0020] Step 1-2: Target scene knowledge: Extract expert knowledge for the target scene to be predicted, i.e., the target domain ;

[0021] Step 1-3: Construction of a domain knowledge base: Integrate common sense of physical laws , expert knowledge and domain professional knowledge to form a comprehensive knowledge base ;

[0022] Step 1-4: After data preprocessing of the source domain images and expert knowledge obtain the preprocessed source domain images and structured knowledge representation .

[0023] Preferably, the source domain images are infrared target data or visible light image data.

[0024] Preferably, the target scenario includes the spatial environment, relevant literature, and report information.

[0025] Preferably, the data preprocessing includes image normalization, noise removal, and geometric correction.

[0026] Preferably, Step 2 is specifically as follows:

[0027] Step 2-1: Design of prompt engineering for open-source language models;

[0028] Adopt a structured causal reasoning prompt engineering design, including the following components:

[0029] a) Context setting module ;

[0030] b) Causal link orientation module ;

[0031] c) Structured output specification module ;

[0032] d) Uncertainty quantification module ;

[0033] e) Counterfactual thinking module ;

[0034] The construction of the complete prompt template is as follows:

[0035]

[0036] Among them, represents the complete prompt template;

[0037] Step 2-2: Causal feature extraction of unstructured knowledge;

[0038] Define the causal feature extraction function:

[0039]

[0040] Among them, represents the set of extracted causal features, represents the language model function for feature extraction;

[0041] Step 2-3: Causal chain analysis and construction of knowledge representation framework;

[0042] Based on the extracted causal features, construct a causal chain analysis model:

[0043]

[0044] Among them, represents the causal graph, Represents a set of nodes, i.e., each part or characteristic attribute corresponding to the target. Represents a set of edges, i.e., the causal relationships between parts. Represents a weight matrix, i.e., the strength of causal relationships.

[0045] Automatically construct a causal graph by analyzing the structured JSON data output by a language large model; the weight of an edge is calculated based on the strength and confidence of the causal relationship given by the model.

[0046] Step 2 - 4: Inference of target characteristics in a specific state.

[0047] Based on the constructed causal graph, use probability inference methods to calculate the temperature values of each part of the target and radiation values :

[0048]

[0049] Among them, represents the probability distribution of temperature values and radiation values under a specific state and the causal graph ; represents an inference function based on the language model.

[0050] Use the Monte Carlo sampling method to estimate the probability distribution of temperature values and radiation values, and calculate the confidence interval using the uncertainty range given by the language large model.

[0051] Output the temperature values of each part of the target and radiation values where

[0052] Preferably, the specific steps of step 3 are as follows:

[0053] Step 3 - 1: Design of the overall architecture of the segmentation model.

[0054] The segmentation model consists of the following four modules:

[0055] a) Feature encoder : Located at the very front of the segmentation model, responsible for extracting multi-scale features of the image; uses ResNet - 50 as the backbone network and is pre-trained and initialized; the feature encoder contains 5 stages, and each stage downsamples once to generate feature maps of different scales ;

[0056] b) Domain adaptation module : Located between the feature encoder and the feature decoder, it consists of multiple domain adversarial discriminators corresponding to features at different levels; through adversarial training, it reduces the difference in feature distributions between the source domain and the target domain; the domain adaptation module performs the following mapping:

[0057]

[0058] Among them, is the output feature of the last layer of the encoder, is the domain discriminator, is the feature after domain adaptation; represents the domain adaptation transformation function;

[0059] c) Feature decoder : Located in the second half of the segmentation model, it is responsible for mapping the features of the feature encoder and the domain adaptation module to the segmentation result; the feature decoder adopts a cascaded upsampling structure to fuse skip connection features of different scales:

[0060]

[0061] Among them, represents the output feature of the th layer of the decoder, represents the upsampling result of the previous layer's feature, represents the skip connection feature of the corresponding encoder layer, represents the feature decoding transformation function;

[0062] d) Segmentation head : Located at the very end of the segmentation model, it maps the features output by the feature decoder to the final segmentation result:

[0063]

[0064] Among them, represents the output segmentation mask, represents the output feature of the last layer of the feature decoder, represents the feature mapping function of the segmentation head;

[0065] Step 3-2: Training phase;

[0066] The training phase adopts an unsupervised domain adaptation strategy, specifically as follows:

[0067] a) Source domain supervised learning: Use the preprocessed source domain images for supervised training to optimize the segmentation loss:

[0068]

[0069] Among them, represents the cross-entropy loss function, represents the segmentation label of the source-domain image, represents the feature encoding transformation function;

[0070] b) Adversarial domain adaptation: Introduce a domain discriminator , and reduce the domain difference through adversarial training:

[0071]

[0072] Among them, represents the target-domain image, represents the natural logarithm function, represents the expectation operation on the target-domain image, represents the expectation operation on the source-domain image, represents the adversarial loss function;

[0073] c) Pseudo-label generation: Infer the target-domain image to generate high-confidence pseudo-labels:

[0074]

[0075] Among them, represents the generated pseudo-label, represents the confidence threshold, represents the threshold-based binarization function;

[0076] d) Self-training iteration: Iteratively optimize the model using the generated pseudo-labels:

[0077]

[0078] Among them, represents the self-training loss function;

[0079] e) Overall optimization objective:

[0080]

[0081] Among them, represents the regularization term, , , are all weight coefficients;

[0082] Step 3-3: Testing phase;

[0083] a) Input the image to be segmented ;

[0084] b) Extract features through the feature encoder: ;

[0085] c) Process the features through the domain adaptation module: ;

[0086] d) Generate multi-scale features through the feature decoder: ;

[0087] e) Generate the final segmentation result through the segmentation head: ;

[0088] f) Apply post-processing to optimize the segmentation boundary to obtain the final segmentation mask:

[0089]

[0090] Among them, the final segmentation mask contains semantic regions , and each region corresponds to a component of the target; represents the post-processing optimization function;

[0091] Preferably, step 4 is specifically:

[0092] Step 4-1: Region-level infrared characteristic assignment;

[0093] According to the temperature value and radiation value obtained in step 2, assign infrared characteristics to each semantic region segmented in step 3:

[0094]

[0095] Among them, represents the initial infrared characteristic value at pixel point , represents the probability that pixel point belongs to region ; respectively represent the temperature value and radiation value of the i-th region;

[0096] Step 4-2: Generation of temperature gradient based on the heat transfer model;

[0097] Use the conduction coefficient matrix between components given by the language large model to construct a temperature gradient formula:

[0098]

[0099] Among them, represents the infrared characteristic value considering the gradient, represents the influence coefficient of the -th region, represents the attenuation parameter, Represents a point to the area The shortest distance to the boundary;

[0100] Step 4-3: Target domain adaptability adjustment;

[0101] According to the environmental parameters of the target domain , perform adaptability adjustment on the infrared characteristics:

[0102]

[0103] Among them, Represents the infrared characteristic value after adapting to the target domain, Represents the global adjustment parameter, Represents the influence function of the environmental parameters.

[0104] Preferably, the specific steps of step 5 are as follows:

[0105] Step 5-1: Difference measurement and loss function construction;

[0106] Define the measured characteristics and the predicted characteristics The difference measurement between them:

[0107]

[0108] Among them, Represents the mean square error, Represents the structural similarity index, and Represents the weight coefficient, Represents the difference measurement loss function;

[0109] Step 5-2: Parameter optimization and calibration;

[0110] By minimizing the difference measurement, optimize the model parameters:

[0111]

[0112] Among them, Represents the optimized parameters;

[0113] Step 5-3: Final infrared characteristic generation;

[0114] Based on the optimized parameters, generate the final infrared characteristic prediction result :

[0115] .

[0116] The beneficial effects of the present invention are as follows:

[0117] 1. Because a causal reasoning feature extraction framework based on a large language model is adopted, the present invention can solve the problem of strong dependence on target parameters in traditional methods and achieve the effect of automatically extracting key parameters from unstructured knowledge. The large language model can understand and integrate physical laws, expert knowledge, and domain-specific knowledge, and extract prior information with causal significance from them, greatly reducing the dependence on precise parameters.

[0118] 2. Because an unsupervised domain adaptation semantic segmentation technology is adopted, the present invention can solve the problem of weak cross-scene generalization ability in traditional methods and achieve an effective migration effect from the source domain to the target domain. This method does not require labeled data in the target domain, greatly reducing the data acquisition cost.

[0119] 3. Because a large model-guided infrared feature assignment and gradient generation method is adopted, the present invention can solve the problem that traditional methods are difficult to express the heat conduction effect between components and achieve a more realistic infrared feature simulation effect. The temperature gradient generation based on the heat transfer model makes the temperature transition between different parts of the target more natural and conforms to physical laws.

[0120] 4. Because an adaptive model calibration mechanism based on measured data is adopted, the present invention can solve the problem of time lag in traditional methods and achieve the effect of rapid model update and iteration. Only a small amount of measured data is required to effectively calibrate the model, significantly improving the prediction accuracy.

[0121] In summary, the present invention breaks through the bottleneck of traditional infrared target feature prediction technology. Through intelligent analysis and prediction driven by a large language model, it greatly improves the efficiency, accuracy, and adaptability of infrared target feature prediction, providing reliable technical support for autonomous recognition of space targets. Brief Description of the Drawings

[0122] Figure 1 is the flowchart of the method of the present invention. Detailed Embodiments

[0123] The present invention will be further described below in conjunction with the drawings and embodiments.

[0124] As Figure 1 shown, the present invention provides a rapid infrared target feature prediction method based on a large language model. By constructing a causal reasoning feature extraction framework, it fuses multi-source information such as expert knowledge and intelligence reports to perform feature inversion on unknown target parameters; establishes a cross-domain migration prediction model for target features to achieve cross-domain generation of feature images, and finally forms a rapid target feature prediction system for complex space environments. The present invention can effectively solve problems such as missing target model parameters, dynamic evolution of features across time domains, and significant differences in multi-condition scene features, greatly improving the efficiency and accuracy of infrared target feature prediction.

[0125] Step 1: Multi-source data acquisition and preprocessing;

[0126] Step 1-1: Source domain image data: Obtain known source domain images , such as infrared target data in plain areas or visible light image data;

[0127] Step 1-2: Target scenario knowledge: For the target scenario to be predicted, i.e., the target domain, such as the space environment, collect relevant literature, reports, etc. and extract expert knowledge ;

[0128] Step 1-3: Domain knowledge base construction: Integrate common sense of physical laws , expert knowledge and domain-specific expertise to form a comprehensive knowledge base ;

[0129] Step 1-4: Preprocess the source domain images and expert knowledge through data preprocessing, including image normalization, noise removal, geometric correction, etc., to obtain preprocessed source domain images and structured knowledge representation .

[0130] Step 2: Causal feature extraction and knowledge representation based on large language models;

[0131] Step 2-1: Design of prompting engineering for open-source large language models;

[0132] Adopt structured causal reasoning prompting engineering design, including the following components:

[0133] a) Context setting module ;

[0134] b) Causal link orientation module ;

[0135] c) Structured output specification module ;

[0136] d) Uncertainty quantification module ;

[0137] e) Counterfactual thinking module ;

[0138] The complete prompting template is constructed as follows:

[0139]

[0140] Step 2-2: Causal feature extraction of unstructured knowledge;

[0141] Define the causal feature extraction function:

[0142]

[0143] Step 2-3: Causal chain analysis and construction of knowledge representation framework;

[0144] Based on the extracted causal features, construct a causal chain analysis model:

[0145]

[0146] By analyzing the structured JSON data output by the language large model, automatically construct a causal graph; the weight of the edge is calculated based on the causal relationship strength and confidence given by the model;

[0147] Step 2-4: Inference of target characteristics under specific states;

[0148] Based on the constructed causal graph, use the probabilistic inference method to calculate the temperature values of each part of the target under a specific state and radiation values :

[0149]

[0150] Among them, represents the probability distribution of temperature values and radiation values under a specific state and the causal graph conditions, represents the inference function based on the language model;

[0151] Adopt the Monte Carlo sampling method to estimate the probability distribution of temperature values and radiation values, and calculate the confidence interval using the uncertainty range given by the language large model;

[0152] Output the temperature values of each part of the target under a specific state and radiation values , where is the number of target parts.

[0153] Step 3: Unsupervised domain adaptation image semantic segmentation;

[0154] Step 3-1: Overall architecture design of the segmentation model;

[0155] The segmentation model consists of the following four modules:

[0156] a) Feature encoder : Located at the very front of the segmentation model, responsible for extracting multi-scale features of the image; ResNet-50 is used as the backbone network and pre-trained for initialization; the feature encoder consists of 5 stages, each stage downsamples once to produce feature maps of different scales ;

[0157] b) Domain Adaptation Module : Located between the feature encoder and the feature decoder, composed of multiple domain adversarial discriminators corresponding to features at different levels; through adversarial training, reduce the difference in feature distributions between the source domain and the target domain; the domain adaptation module performs the following mapping:

[0158]

[0159] c) Feature Decoder : Located in the second half of the segmentation model, responsible for mapping the features of the feature encoder and the domain adaptation module to the segmentation result; the feature decoder adopts a cascaded upsampling structure to fuse skip connection features of different scales:

[0160]

[0161] d) Segmentation Head : Located at the very end of the segmentation model, mapping the features output by the feature decoder to the final segmentation result:

[0162]

[0163] Step 3-2: Training Phase;

[0164] The training phase adopts an unsupervised domain adaptation strategy, specifically as follows:

[0165] a) Source Domain Supervised Learning: Use the preprocessed source domain images for supervised training to optimize the segmentation loss:

[0166]

[0167] b) Adversarial Domain Adaptation: Introduce a domain discriminator , and reduce the inter-domain difference through adversarial training:

[0168]

[0169] c) Pseudo-Label Generation: Infer the target domain images to generate high-confidence pseudo-labels:

[0170]

[0171] d) Self-Training Iteration: Use the generated pseudo-labels to iteratively optimize the model:

[0172]

[0173] e) Overall optimization objective:

[0174]

[0175] Step 3-3: Testing phase;

[0176] a) Input the image to be segmented ;

[0177] b) Extract features through the feature encoder: ;

[0178] c) Process the features through the domain adaptation module: ;

[0179] d) Generate multi-scale features through the feature decoder: ;

[0180] e) Generate the final segmentation result through the segmentation head: ;

[0181] f) Apply post-processing to optimize the segmentation boundary to obtain the final segmentation mask:

[0182]

[0183] Among them, the final segmentation mask contains semantic regions , and each region corresponds to a component of the target; represents the post-processing optimization function;

[0184] Step 4: Infrared property assignment and gradient generation based on the language large model;

[0185] Step 4-1: Region-level infrared property assignment;

[0186] According to the temperature values and radiation values obtained in Step 2, perform infrared property assignment to each semantic region segmented in Step 3:

[0187]

[0188] Step 4-2: Temperature gradient generation based on the heat transfer model;

[0189] Utilize the conduction coefficient matrix between components given by the language large model to construct the temperature gradient formula:

[0190]

[0191] Step 4-3: Target domain adaptability adjustment;

[0192] According to the environmental parameters of the target domain , perform adaptability adjustment on the infrared characteristics:

[0193]

[0194] Step 5: Model calibration and accuracy optimization based on measured data;

[0195] Step 5-1: Difference measurement and loss function construction;

[0196] Define the difference measurement between the measured characteristics and the predicted characteristics :

[0197]

[0198] Step 5-2: Parameter optimization and calibration;

[0199] Optimize the model parameters by minimizing the difference measurement:

[0200]

[0201] Step 5-3: Generation of final infrared characteristics;

[0202] Based on the optimized parameters, generate the final infrared characteristic prediction result :

[0203] .

[0204] Step 6: System integration and application deployment;

[0205] In this step, integrate the above steps into a complete infrared target characteristic rapid prediction system and perform actual application deployment.

[0206] 1) Process automation and interface design;

[0207] Design unified data input and result output interfaces to achieve automated execution of Steps 1 to 5.

[0208] 2) Construction of multi-scenario parameter configuration library;

[0209] Establish a parameter configuration library for different application scenarios to support rapid switching and deployment.

[0210] 3) Real-time prediction and result visualization;

[0211] Achieve real-time prediction and three-dimensional visualization display of infrared characteristics, and support interactive analysis and applications.

[0212] Test results:

[0213] Table 1 Performance comparison of different methods in predicting the infrared characteristics of spatial scene targets

[0214] Method Structural Similarity (SSIM) Peak Signal-to-Noise Ratio (PSNR / dB) Mean Absolute Error (MAE / °C) Computation Time (s) The method of the present invention 0.89 35.7 2.3 18 Traditional physical modeling method 0.74 29.3 5.8 432 Deep learning image conversion method 0.81 32.5 3.7 46 GAN based on domain adaptation 0.83 33.1 3.2 87

[0215] It can be seen from the test results that the method of the present invention is superior to the existing methods in key indicators such as structural similarity, peak signal-to-noise ratio and mean absolute error, and at the same time, the calculation time is greatly reduced, which fully proves the effectiveness and advancement of the method.

Claims

1. A fast prediction method for spatial infrared target characteristics based on a large language model, characterized in that, It includes the following steps: Step 1: Multi-source data acquisition and preprocessing; Step 2: Causal feature extraction and knowledge representation based on large language models; Step 3: Unsupervised domain adaptation image semantic segmentation; Step 4: Infrared feature assignment and gradient generation based on large language models; Step 5: Model calibration and accuracy optimization based on measured data.

2. A method for rapidly predicting the characteristics of spatial infrared targets based on a large language model according to claim 1, characterized in that, The specific content of Step 1 is as follows: Step 1-1: Source domain image data: Obtain known source domain images ; Step 1-2: Target scenario knowledge: Extract expert knowledge for the target scenario to be predicted, i.e., the target domain. ; Step 1-3: Domain knowledge base construction: Integrate common sense of physical laws , expert knowledge and domain-specific expertise to form a comprehensive knowledge base ; Step 1-4: For the source domain image and expert knowledge After data preprocessing, the preprocessed source domain image and structured knowledge representation .

3. A method for quickly predicting the characteristics of space infrared targets based on a large language model according to claim 2, characterized in that, The source domain image is infrared target data or visible light image data.

4. A rapid prediction method for spatial infrared target characteristics based on a large language model according to claim 2, characterized in that, The target scenarios include space environment, relevant literature, and report information.

5. A fast prediction method for spatial infrared target characteristics based on a large language model according to claim 2, characterized in that, The data preprocessing includes image normalization, noise removal, and geometric correction.

6. The rapid prediction method for spatial infrared target characteristics based on a large language model according to claim 2, characterized in that, The specific content of Step 2 is as follows: Step 2-1: Design of prompt engineering for open-source large language models; Structured causal reasoning prompt engineering design is adopted, including the following components: a) Context setting module ; b) Causal Linkage Oriented Module ; c) Structured output specification module ; d) Uncertainty quantification module ; e) Counterfactual thinking module ; The complete prompt template is constructed as follows: ； Among them, represents the complete prompt word template; Step 2-2: Causal feature extraction of unstructured knowledge; Define the causal feature extraction function: ； Among them, represents the set of extracted causal features, represents the language large model function for feature extraction; Step 2-3: Causal chain analysis and construction of knowledge representation framework; Based on the extracted causal features, construct a causal chain analysis model: ； Among them, represents a causal diagram, represents a node set, that is, each part or characteristic attribute corresponding to the target, represents an edge set, that is, the causal relationship between parts, represents a weight matrix, that is, the strength of the causal relationship; Automatically construct a causal graph by analyzing the structured JSON data output by the large language model; the weight of the edge is calculated based on the causal relationship strength and confidence given by the model; Step 2-4: Inference of target characteristics under specific states; Based on the constructed causal graph, using the probabilistic reasoning method, calculate the temperature values of each part of the target under a specific state and the radiation values under a specific state : ； Among them, represents the temperature value and radiation value probability distribution under a specific state and the causal diagram under the condition, represents the inference function based on the language model; Adopt the Monte Carlo sampling method to estimate the probability distributions of temperature values and radiation values, and calculate the confidence intervals using the uncertainty range given by the large language model; Output the temperature values of each part of the target under specific states and radiation values , where is the number of target parts 7. A method for quickly predicting the characteristics of space infrared targets based on a large language model according to claim 6, characterized in that, The specific content of Step 3 is as follows: Step 3-1: Overall architecture design of the segmentation model; The segmentation model consists of the following four modules Composition: a) Feature Encoder : Located at the very front of the segmentation model, responsible for extracting multi-scale features of the image; Use ResNet-50 as the backbone network and perform pre-training initialization; the feature encoder consists of 5 stages, each stage downsamples once to generate feature maps of different scales ; b) Domain Adaptation Module : Located between the feature encoder and the feature decoder, it consists of multiple domain adversarial discriminators corresponding to features at different levels; through adversarial training, it reduces the difference in feature distributions between the source domain and the target domain; the domain adaptation module performs the following mapping: ； Among them, is the output feature of the last layer of the encoder, is the domain discriminator, is the feature after domain adaptation; represents the domain adaptation transformation function; c) Feature decoder : Located in the second half of the segmentation model, responsible for mapping the features of the feature encoder and the domain adaptation module to the segmentation result; the feature decoder adopts a cascaded upsampling structure to fuse the skip connection features of different scales: ； Among them, represents the output feature of the -th layer of the decoder, represents the upsampling result of the feature of the previous layer, represents the skip connection feature corresponding to the encoder layer, represents the feature decoding transformation function; d) Splitting head : Located at the very end of the splitting model, it maps the features output by the feature decoder to the final splitting result: ； Among them, represents the output segmentation mask, represents the output features of the last layer of the feature decoder, represents the feature mapping function of the segmentation head; Step 3-2: Training stage; The unsupervised domain adaptation strategy is adopted in the training stage, specifically as follows: a) Source domain supervised learning: Use the preprocessed source domain images for supervised training to optimize the segmentation loss: ； Among them, represents the cross-entropy loss function, represents the segmentation label of the source domain image, represents the feature encoding transformation function; b) Adversarial domain adaptation: Introduce a domain discriminator , and reduce the inter-domain difference through adversarial training: ； Among them, represents the target domain image, represents the natural logarithm function, represents the expectation operation on the target domain image, represents the expectation operation on the source domain image, represents the adversarial loss function; c) Pseudo-label generation: Infer the target domain images to generate high-confidence pseudo-labels: ； Among them, represents the generated pseudo-label, represents the confidence threshold, represents the threshold-based binarization function; d) Self-training iteration: Use the generated pseudo-labels to iteratively optimize the model: ； Among them, represents the self-training loss function; e) Overall optimization objective: ； Among them, represents the regularization term, , , are all weight coefficients; Step 3-3: Testing stage; a) Input the image to be segmented ; b) Extract features through the feature encoder: ; c) Process the features through the domain adaptation module: ; d) Generating multi-scale features through the feature decoder: ; e) Generate the final segmentation result through the segmentation head: ; f) Apply post-processing to optimize the segmentation boundary to obtain the final segmentation mask: ； Among them, the final segmentation mask includes semantic regions , and each region corresponds to a component of the target; represents a post-processing optimization function.

8. A method for quickly predicting the characteristics of space infrared targets based on a large language model according to claim 7, characterized in that, The specific content of Step 4 is as follows: Step 4-1: Region-level infrared feature assignment; The temperature value obtained according to Step 2 and the radiation value , perform infrared characteristic assignment on each semantic region segmented in Step 3: ； Among them, represents the initial infrared characteristic value at the pixel point , and represents the probability that the pixel point belongs to the region . respectively represent the temperature value and radiation value of the i-th region; Step 4-2: Temperature gradient generation based on the heat transfer model; Conduction coefficient matrix between components given by the large language model , construct the temperature gradient formula: ； Among them, represents the infrared characteristic value after considering the gradual change, represents the influence coefficient of the th region, represents the attenuation parameter, represents the point to the shortest distance of the region Step 4-3: Target domain adaptability adjustment; According to the environmental parameters of the target domain , adaptively adjust the infrared characteristics: ； Among them, represents the infrared characteristic value after adapting to the target domain, represents the global adjustment parameter, represents the influence function of environmental parameters.

9. A method for rapidly predicting the characteristics of space infrared targets based on a large language model according to claim 8, characterized in that, The specific content of Step 5 is as follows: Step 5-1: Difference measurement and construction of loss function; Define measured characteristics and predicted characteristics The difference metric between: ； Among them, represents the mean square error, represents the structural similarity index, and represents the weight coefficient, represents the difference metric loss function; Step 5-2: Parameter optimization and calibration; Optimize the model parameters by minimizing the difference measurement: ； Among them, represents the optimized parameter; Step 5-3: Final infrared feature generation; Generate the final infrared characteristic prediction result based on the optimized parameters : 。

Images