One-stop adaptive architecture and method for constructing multi-mode large model in credible electromagnetic radar field

Abstract

The invention discloses a one-stop adaptive architecture and method for constructing a trusted electromagnetic radar field multi-mode large model, and the method comprises the steps: obtaining unsupervised text data of an electromagnetic radar field, and carrying out the incremental pre-training of a preset base language model, so as to construct a fact base model; training is carried out by using supervised question and answer data in the field of electromagnetic radars through an efficient parameter fine tuning method, and alignment of instruction following and dialogue interaction capability is realized; performing multi-modal fine tuning by using the multi-modal data containing the radar image or the electromagnetic signal; generating a preliminary answer by using the multi-modal fine-tuned model; and starting an auto-reflection module to check the logic consistency and the fact accuracy of the preliminary answer so as to generate a final checked or corrected answer. According to the method, the credibility construction process of the model is decomposed into clear and progressive stages, and the problem that knowledge depth, dialogue ability and multi-modal understanding are difficult to consider in a single fine tuning method is solved.

Description

Technical Field

[0001] This invention relates to the field of large language model vertical domain application technology, and in particular to a one-stop adaptation architecture and method for constructing a reliable electromagnetic radar multimodal large model. Background Technology

[0002] In recent years, Large Language Models (LLMs), represented by general pre-trained models, have demonstrated powerful versatility in natural language understanding and generation, and have made breakthroughs in many general scenarios. How to effectively transfer and apply this powerful versatility to highly specialized and cutting-edge vertical fields such as national defense, aerospace, meteorology, and autonomous driving, and build domain-specific large models, has become a key challenge for the industrialization of artificial intelligence technology.

[0003] However, when general-purpose large language models are directly applied to knowledge-intensive and highly accuracy-critical fields like electromagnetic radar, they face three major technical bottlenecks: First, insufficient knowledge depth; general-purpose models lack a deep understanding of the vast amount of specialized terminology, complex formulas, and underlying principles within the field. Second, factual errors and "illusions"; the models are prone to generating erroneous information that contradicts the facts of the field, making their output unreliable. Third, a lack of understanding of multimodal information; research and applications in the electromagnetic radar field heavily rely on the analysis of key images and signals such as radar system diagrams, antenna patterns, and electromagnetic signal spectra, while pure text models cannot handle this non-textual information. Currently, the industry often uses a single fine-tuning method to adapt the model, but these methods often address one aspect while neglecting another, making it difficult to solve all the above problems simultaneously. Furthermore, the adaptation process heavily relies on manual experience for optimization, resulting in high costs and low efficiency, ultimately leading to models with insufficient credibility and difficulty in fulfilling important responsibilities in actual scientific research and production. Therefore, there is an urgent need for a systematic architecture that can comprehensively utilize multi-source heterogeneous data within the domain to fully enhance the model's knowledge depth, factual accuracy, and multi-dimensional perception capabilities in the professional field, so as to build a truly credible and usable domain-specific large model. Summary of the Invention

[0004] To address the aforementioned technical challenges, this invention provides a one-stop adaptation architecture and method for constructing a reliable multimodal large-scale model in the electromagnetic radar field. It aims to systematically adapt multi-source heterogeneous data, including unsupervised text, supervised question-and-answer sessions, and image-text signals, to electromagnetic radar through a layered enhancement strategy that combines incremental pre-training, efficient parameter fine-tuning, multimodal fine-tuning, and self-reflective verification. This fundamentally solves problems such as insufficient knowledge, factual illusions, and lack of multimodal understanding in general models within specialized fields, ultimately outputting a highly reliable large-scale model specifically for the electromagnetic radar field.

[0005] The technical solution proposed in this invention is as follows:

[0006] A one-stop adaptation architecture and method for constructing a reliable multimodal large model in the field of electromagnetic radar includes:

[0007] Step 1: Obtain unsupervised text data in the field of electromagnetic radar and perform incremental pre-training on the pre-set base language model to construct a fact base model infused with domain factual knowledge;

[0008] Step 2: Based on the aforementioned fact-based model, supervised question-and-answer data from the electromagnetic radar field is used to train the model through an efficient parameter fine-tuning method, so as to align the model's adherence to instructions and its ability to engage in dialogue.

[0009] Step 3: Based on the model trained in the previous steps, perform multimodal fine-tuning using multimodal data containing radar images or electromagnetic signals to enable the model to have multidimensional perception capabilities to understand domain-specific non-textual information.

[0010] Step 4: In the reasoning stage, after receiving the user's input question, a preliminary answer is generated using the multimodal fine-tuned model; and a self-reflection module is activated to check the logical consistency and factual accuracy of the preliminary answer in order to generate a final verified or corrected answer.

[0011] As a preferred option, step 1 specifically involves:

[0012] S1.1: Construct unsupervised text data, including collecting raw text materials from academic papers, technical manuals, industry standards and textbooks, and cleaning, deduplicating and segmenting them to form a corpus in the field of electromagnetic radar;

[0013] S1.2: Incremental pre-training of the base language model using a corpus in the field of electromagnetic radar.

[0014] As a preferred option, step 2 specifically involves:

[0015] S2.1: Constructing supervised question-and-answer data: Constructing professional questions in the field of electromagnetic radar and their corresponding standard answers into a command-response pair format;

[0016] S2.2: Fine-tuning the fact base model using efficient parameter fine-tuning methods, including low-rank adaptation, prefix fine-tuning, and cue fine-tuning.

[0017] Preferably, the multimodal fine-tuning in step 3 is implemented based on the encoder-adapter-language model architecture, specifically as follows:

[0018] S3.1: For radar images, visual features are extracted using a visual encoder; for electromagnetic signals, signal features are extracted using a pre-trained signal processing model.

[0019] S3.2: Input the visual features or signal features into a multilayer perceptron that acts as an adapter, and the adapter maps the dimensions of the input features to dimensions that match the word embedding space of the base language model, so as to achieve cross-modal feature alignment;

[0020] S3.3: The feature vector processed by the adapter is concatenated or fused with the corresponding text embedding vector and used as input to fine-tune the base language model.

[0021] As a preferred option, step 4 specifically involves:

[0022] S4.1: Combine the user-input question and the model-generated preliminary answer into a self-criticism prompt word;

[0023] S4.2: Input the self-criticism prompts into the language model and guide the model to output a structured evaluation result containing criticisms, accuracy scores, and improvement suggestions;

[0024] S4.3: If the structured evaluation results indicate that improvement is needed, the original question, preliminary answer, and criticisms are combined into an improvement suggestion word;

[0025] S4.4: Input the improved prompt words into the language model, guide the model to re-reason based on the criticisms, and generate an improved answer as the final output.

[0026] Beneficial effects:

[0027] This invention decomposes the complex domain adaptation process into clear and progressive stages by constructing a layered enhancement architecture that includes a fact base, capability alignment, multi-dimensional perception, and final verification. This enables one-stop adaptation from multi-source heterogeneous data to a trusted domain model, significantly reducing the adaptation threshold and high human resource costs of building domain-specific large models.

[0028] This invention injects domain-specific knowledge through incremental pre-training, then uses supervised data for efficient parameter fine-tuning to train the model's instruction compliance and professional dialogue capabilities. Furthermore, multimodal fine-tuning extends its understanding to radar images and electromagnetic signals. This progressive training process not only equips the model with a deep knowledge base but also teaches it how to accurately and professionally interact with humans. It solves the problem that single fine-tuning methods struggle to balance knowledge depth, dialogue ability, and multimodal understanding, making the model a truly interactive and applicable domain expert.

[0029] This invention deploys a self-reflective module during the inference stage to perform final logical and factual verification of the model output, forming a dynamic self-evaluation and correction closed loop. This mechanism can proactively suppress and correct "illusions" and factual errors that may occur when the model generates content. Compared with output methods that rely solely on the model's own memory, this ensures the high reliability and professionalism of the final output results. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of a one-stop adaptation architecture for constructing a large multi-model model in the field of trusted electromagnetic radar, provided by an embodiment of the present invention.

[0031] Figure 2 This is a flowchart illustrating a method for constructing a large multimodal model in the field of trusted electromagnetic radar, provided as an embodiment of the present invention.

[0032] Figure 3 This is a schematic diagram of the "encoder-adapter-language model" architecture used inside the multimodal fine-tuning module in this embodiment of the invention.

[0033] Figure 4 This is a flowchart illustrating the workflow of the self-reflection module performing a two-step verification and correction process of "criticism-improvement" in an embodiment of the present invention. Detailed Implementation

[0034] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0035] This embodiment discloses a one-stop adaptation architecture and method for constructing a reliable multi-model large model in the electromagnetic radar domain. It employs a layered enhancement strategy: First, a fact base layer: acquiring massive amounts of unsupervised text data within the electromagnetic radar domain, and injecting domain factual knowledge into the base language model through incremental pre-training to construct a fact base; Second, a capability alignment layer: based on this fact base, utilizing supervised question-and-answer data from the electromagnetic radar domain, and through efficient parameter fine-tuning, training the model's instruction compliance and dialogue interaction capabilities to achieve capability alignment; Third, a multi-dimensional perception layer: further utilizing key multi-modal data including radar images and electromagnetic signals, performing multi-modal fine-tuning to enable the model to understand domain-specific images and signals, constructing a multi-dimensional perception layer; Finally, a final verification layer: during the inference stage, deploying a self-reflective module to check the logical consistency and factual accuracy of the model's initial answers, and performing final verification and correction of the output results.

[0036] Combination Figures 1 to 4 This paper presents a one-stop adaptation architecture and method for building multiple and large-scale models in the field of trusted electromagnetic radar, including:

[0037] like Figure 1 As shown, this architecture, from left to right, illustrates the complete process of data input, hierarchical enhancement strategies, and final model output for a reliable multimodal large model in the electromagnetic radar domain. The core processing architecture includes a fact base layer, a capability alignment layer, a multidimensional perception layer, and a final verification layer.

[0038] like Figure 2 As shown in the figure, this invention provides a flowchart of a method for constructing a multimodal large model in the field of trusted electromagnetic radar. The method mainly includes: Step 1, constructing a fact-based model using unsupervised text data and incremental pre-training; Step 2, achieving capability alignment through supervised question-answering data and efficient parameter fine-tuning; Step 3, constructing a multi-dimensional perception layer using multimodal data and multimodal fine-tuning; Step 4, performing final verification through prompt word design and self-reflection. The above four steps correspond to... Figure 1 The functions of each core module.

[0039] The following section details step 1 of this method: injecting massive amounts of highly reliable factual knowledge about electromagnetic radar into a general base language model, fundamentally solving the problem of missing knowledge in the model's domain. This includes the following two sub-steps:

[0040] Step 1: Acquisition and Processing of Unsupervised Text Data in the Electromagnetic Radar (EMR) Field. The goal is to construct a high-quality, structured corpus specifically for the EMR field. Data sources include literature, books, GJB equipment standards, open-source intelligence from the internet, etc. These raw data are collected and processed through cleaning, deduplication, and segmentation. Finally, all text fragments are combined to form the final EMR-specific corpus, denoted as […]. .

[0041] Step 2: Incremental pre-training based on the domain corpus. The goal is to utilize the corpus built in the previous step. For a general, pre-trained base language model Secondary pre-training was conducted to adapt it to the knowledge system and language paradigm of the electromagnetic radar field.

[0042] The training process is described as follows:

[0043] 1. Base Model Definition: Select a pre-trained base language model. This typically refers to large-scale models based on the Transformer architecture (such as the GPT series, LLaMA series, etc.). This model consists of its parameter set. Defined.

[0044] 2. Training Objective: Incremental pre-training adopts the standard autoregressive language modeling objective, namely "Next Token Prediction." For the corpus... any text sequence The goal of training is to maximize the joint probability of the entire sequence, which can be decomposed into the product of the conditional probabilities at each time step. The corresponding loss function is... Negative Log-Likelihood:

[0045] ;

[0046] in, The model predicts the next word given the preceding context. The probability, These are the parameters to be optimized in the model.

[0047] 3. Parameter Update: Use stochastic gradient descent (SGD) or its variants (such as AdamW) to minimize the loss function. To achieve this, the model's parameters are updated. The initial parameters for this process are... After sufficient training, a new set of model parameters was obtained. : .

[0048] 4. Fact Base Model Generation: After training, a new set of parameters is generated. The defined model is the fact-based model in this invention. Compared to the universal base model The model has internalized the technical terms, basic principles and factual knowledge of the electromagnetic radar field as part of its parameters, laying a solid foundation for its subsequent capability alignment and multi-dimensional perception.

[0049] By executing step 1, this invention completes the construction of a "fact base" that is knowledgeable in the field but not yet adept at dialogue interaction, providing the most critical knowledge source for the entire one-stop adaptation process.

[0050] The following section details step 2 of this method: through supervised fine-tuning, the model is trained to understand and follow human instructions, enabling it to express its internalized knowledge in a professional and accurate question-and-answer or dialogue format, thus completing the crucial transformation from a "knowledge base" to a "domain expert." This includes the following two sub-steps:

[0051] Step 1: Constructing supervised question-answering data in the domain. The goal is to create a high-quality instruction fine-tuning dataset to guide the model's behavior. This dataset... It consists of a series of instruction-response pairs:

[0052] ;

[0053] in: It is an instruction, that is, a question or task description posed to the model. The construction of instructions aims to cover typical query scenarios in the field of electromagnetic radar; Is with instructions The corresponding response. This response is written or reviewed by domain experts to ensure its professionalism, accuracy, logic, and completeness, serving as the "standard answer" for model learning.

[0054] These data pairs can be obtained through methods such as manual writing, extraction from the end-of-chapter exercises of professional textbooks, or summarizing and rewriting domain documents using a more powerful "teacher model" and then having them verified by experts.

[0055] Step 2: Capability-aligned training based on Parameter-Efficient Fine-Tuning (PEFT). Its goal is to efficiently train the model's ability to follow instructions with low computational and storage costs, while maintaining the stability of the knowledge gained in Step 1. Compared to Full Fine-tuning, the PEFT method only updates a small portion of the model's parameters, effectively preventing "catastrophic forgetting" and facilitating model storage and deployment.

[0056] The training process is described as follows:

[0057] 1. Training objective: For the dataset Each instruction-response pair To piece them together into a complete text sequence The goal of training is to maximize the probability that the model will generate a response R given an instruction I. In practice, this is typically achieved using an autoregressive loss function, but the loss is calculated only for the response R, forcing the model to learn to "answer" rather than "recite" the entire sequence. Loss function Defined as:

[0058] ;

[0059] in, It is the t-th token in sequence S. It is the length of the instruction section. It comes from the fact-based model Most of the parameters are frozen, and These are the few new parameters introduced by the PEFT method that require training; the number of parameters is | |≪| |

[0060] 2. Preferred Implementation: Low-Rank Adaptation (LoRA): As defined in claim 3, this embodiment preferably employs the Low-Rank Adaptation (LoRA) method. The core idea of ​​LoRA is that for a certain pre-trained weight matrix in the model (usually in the self-attention module of the Transformer), Do not update directly during fine-tuning Instead, a "bypass" is added next to it, using two low-rank matrices. and The product of BA is used to approximate the weight update. Here, the rank *r* is a hyperparameter much smaller than *d* and *k*. During forward propagation, the original computation... Modified to During training, the original weights The system remains frozen, with only the parameters of matrices A and B being updated. Therefore, That is, the set of parameters for matrices A and B in all LoRA layers.

[0061] 3. Other implementation methods: In addition to LoRA, the present invention can also use other PEFT methods, such as prefix-tuning or hint-tuning, which can achieve the purpose of efficient fine-tuning and model alignment.

[0062] 4. Capability Alignment Model Generation: After training converges, a capability alignment model is obtained. The model is based on a fixed factual foundation. It is defined together with a set of pre-trained, lightweight PEFT parameters (such as the A and B matrices of LoRA). This model now has the ability to perform professional and fluent question answering in the field of electromagnetic radar, laying an interactive foundation for subsequent multimodal understanding.

[0063] The following section details step 3 of this method, which, through multimodal fine-tuning, endows the model with the ability to understand key visual and signal information, upgrading it from a pure text model to a large multimodal model capable of joint understanding of text, images, and signals. This includes the following two sub-steps:

[0064] Step 1: Constructing domain-specific multimodal data. This dataset... It consists of a series of pairs (non-textual information, textual description):

[0065] ;

[0066] in: It is a non-textual information sample. In this invention, it mainly includes two types: radar images and electromagnetic signals. Is and non-text information A corresponding text description. This description could be a detailed interpretation of the image or signal content, a related question-and-answer pair, or a summary title.

[0067] Step 2: Multimodal fine-tuning based on the "encoder-adapter-LLM" architecture. The detailed process is as follows: Figure 3 As shown, the description is divided into the following three points:

[0068] 1. Feature encoding (corresponding to) Figure 3 (Encoder part in radar image) Deep visual features are extracted using a pre-trained vision encoder. This encoder can be, but is not limited to, the image encoding part of a Vision Transformer (ViT) or CLIP. The encoding process can be formalized as follows:

[0069] ;

[0070] in, It is the feature vector obtained after encoding. For electromagnetic signals... The signal features are extracted using a pre-trained signal processing model (Signal Encoder). The encoding process can be formalized as follows:

[0071] ;

[0072] in, It is the feature vector obtained after encoding.

[0073] 2. Feature space alignment (corresponding) Figure 3 The Adapter section in the code: Due to the feature dimension of the encoder output ( or Word embedding space dimension of Large Language Model (LLM) Typically, these features don't match, and the features themselves lack linguistic semantics. Therefore, an adapter is needed as a bridge. This embodiment preferably uses a simple multilayer perceptron (MLP) as the adapter. This MLP receives the features output by the encoder. or This is then projected onto the word embedding space of the LLM. This process can be formalized as:

[0074] ;

[0075] in, It is a "soft prompt" or feature sequence that has been processed by the adapter, aligned with the LLM word embedding dimension, and can be understood by LLM.

[0076] 3. Multimodal fusion and end-to-end fine-tuning (corresponding to) Figure 3 The LLM section in the text): will be combined with non-text information Paired text description Transform into text embedding vectors through the word embedding layer of LLM. .

[0077] Modal features after adapter processing With text embedding vectors The input sequences can be spliced ​​or merged to form a unified multimodal input sequence. For example, simply splicing them together: The fused input sequence Input into the large language model (i.e., the one obtained in step 2) The goal of training is to enable the model to autoregressively generate correct text descriptions based on the fused context. .

[0078] In the initial training phase, only the adapter parameters are updated, while the encoder and large language model parameters are frozen. The goal of this stage is to quickly complete modality alignment. After alignment, efficient parameter fine-tuning methods such as LoRA are used to jointly fine-tune some parameters of the large language model and the adapter to improve end-to-end performance.

[0079] By performing step 3, a reliable multimodal domain large model is finally obtained. This model not only inherits the domain knowledge and instruction-following capabilities injected in previous steps, but more importantly, it can now understand non-textual information such as radar images and electromagnetic signals, achieving a comprehensive understanding of multi-dimensional information within the domain, thus preparing it to solve complex real-world problems.

[0080] The following section details step 4 of this method, which adds a self-reflection process to the model's output. Through an iterative "criticism-and-refine" loop, it proactively identifies and corrects potential logical fallacies or factual errors, thereby maximizing the reliability and accuracy of the final answer. The detailed workflow of the self-reflection mechanism in this embodiment is as follows: Figure 4 As shown, its specific implementation logic is as follows:

[0081] Step 1: Generate an initial answer. Upon receiving the user's question... Then, the system first calls the trained model. Perform a standard reasoning exercise to generate a preliminary, unverified answer, denoted as... The problem It can be a structured object containing different types of questions (such as multiple choice questions, true / false questions) and metadata (such as options): This preliminary answer is the result generated by the model based on its internalized knowledge, and will serve as the starting point for subsequent self-reflection.

[0082] Step Two: The Self-Criticism Stage (Critique). This stage corresponds to... Figure 4 The self-criticism phase aims to guide the model to examine and evaluate the generated initial answers. First, a self-criticism prompt is constructed: the system uses the user's original question... and the initial answer generated by the model To programmatically construct a structured set of "self-criticism prompts" The template for this prompt is as follows:

[0083] You will be given a pair of questions and responses. Your task is to objectively evaluate the accuracy of the responses and provide feedback.

[0084] Question: {formatted_question}

[0085] Reply: {A_initial}

[0086] Please analyze this reply carefully, reasoning step by step. Your output should follow this format:

[0087] Criticism: (Your detailed criticism and analysis)

[0088] Accuracy rating: (1-10, 1 being the lowest and 10 being the highest)

[0089] Does it need improvement? (Yes / No)

[0090] Here, `{formatted_question}` is the formatted question text according to the question type (e.g., for a multiple-choice question, it would include the question and all options). Next, structured evaluation results are generated: the constructed self-criticism prompts are used. Input into the same model again In the middle. Due to the strong guidance of the prompt words, the model will then output a structured evaluation text containing criticisms, accuracy scores, and improvement suggestions, denoted as . .

[0091] Step 3: Improve the decision-making stage. The system evaluates the text returned by the model. The analysis is performed. This analysis process is based on a series of pre-defined rules with clear priorities:

[0092] Rule 1 (Explicit Positive Judgment): First, use a regular expression to match a predefined list of positive patterns (e.g., ["Improvement Needed: Yes", "Error Exists", "Inaccurate"]). If a match is found, improvement is deemed necessary, and the decision-making process ends.

[0093] Rule 2 (Score Threshold Judgment): If Rule 1 is not matched, the "accuracy score" value is extracted from the text. If the value is lower than or equal to a preset threshold (e.g., 7 points), it is determined that improvement is needed, and the decision-making process ends.

[0094] Rule 3 (Explicit Negation): If none of the above rules are triggered, a regular expression is used to match a predefined list of negative patterns (e.g., ["Needs improvement: No", "Answer correct", "No problem"]). If a match is found, no improvement is required, the entire self-reflection process terminates, and the current answer is accepted.

[0095] Rule 4 (Default Judgment): If none of the above rules provide a clear conclusion, the default judgment is "needs improvement" to ensure a more cautious approach to any uncertain answers.

[0096] Step 4: Response Refinement Stage. If the conclusion of the decision refinement stage is "yes," then this step is executed, corresponding to... Figure 4 The answer refinement stage involves several steps. First, a refinement prompt is constructed: the system presents the original question... Answers that need improvement and the complete evaluation text generated in the second step These are combined to form an "improvement prompt word". The prompt template is as follows:

[0097] Please improve your answer based on the following criticisms:

[0098] Original reply: {A_initial}

[0099] Criticism: {E_eval}

[0100] Now please answer this question again:

[0101] {formatted_question}

[0102] Require:

[0103] 1. Carefully consider the issues raised in the criticisms.

[0104] 2. ... (Other specific requirements)

[0105] {answer_format_instruction}

[0106] The `{answer_format_instruction}` parameter provides explicit answer format requirements based on the question type (e.g., multiple-choice questions require the output "The answer is X"). Next, the final answer is generated: the improved prompts are then used. The third input to the model In this case, the model will focus on the pointed-out error and re-reason based on the criticisms and improvement suggestions, thereby generating a higher-quality, corrected answer. .

[0107] Step 5: Iteration and Termination. The aforementioned "criticism-decision-correction" cycle can be executed for a preset fixed number of rounds (e.g., 2 rounds), or terminated early when the improvement decision stage of a certain round determines "no improvement is needed." In each iteration, the newly generated improved answer will become the "original answer" for the next round of reflection. Finally, the system outputs the best answer obtained in the last round of the cycle. Simultaneously, all intermediate products throughout the process (such as the initial answer, criticisms, and improved answers for each round) can be recorded as reflection history for subsequent analysis and auditing.

[0108] By executing step 4, this invention establishes a reliable quality control closed loop at the model's output. This self-reflective mechanism effectively utilizes the model's powerful reasoning ability for self-supervision, greatly enhancing the "credibility" of the output results and giving it real practical value in demanding professional fields.

[0109] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A one-stop adaptation architecture and method for constructing a reliable multimodal large model in the field of electromagnetic radar, employing a layered enhancement strategy, characterized in that: Includes the following steps: Step 1: Obtain unsupervised text data in the field of electromagnetic radar and perform incremental pre-training on the pre-set base language model to construct a fact base model infused with domain factual knowledge; Step 2: Based on the aforementioned fact-based model, supervised question-and-answer data from the electromagnetic radar field is used to train the model through an efficient parameter fine-tuning method, so as to align the model's adherence to instructions and its ability to engage in dialogue. Step 3: Based on the model trained in the previous steps, perform multimodal fine-tuning using multimodal data containing radar images or electromagnetic signals to enable the model to have multidimensional perception capabilities to understand domain-specific non-textual information. Step 4: In the reasoning stage, after receiving the user's input question, a preliminary answer is generated using the multimodal fine-tuned model; The system then initiates a self-reflection module to check the logical consistency and factual accuracy of the preliminary answer in order to generate a final verified or corrected answer.

2. The method according to claim 1, characterized in that, Step 1 specifically involves: S1.1: Construct unsupervised text data, including collecting raw text materials from academic papers, technical manuals, industry standards and textbooks, and cleaning, deduplicating and segmenting them to form a corpus in the field of electromagnetic radar; S1.2: Incremental pre-training of the base language model using a corpus in the field of electromagnetic radar.

3. The method according to claim 1, characterized in that, Step 2 specifically involves: S2.1: Constructing supervised question-and-answer data: Constructing professional questions in the field of electromagnetic radar and their corresponding standard answers into a command-response pair format; S2.2: Fine-tuning the fact base model using efficient parameter fine-tuning methods, including low-rank adaptation, prefix fine-tuning, and cue fine-tuning.

4. The method according to claim 1, characterized in that, The multimodal fine-tuning in step 3 is implemented based on the encoder-adapter-language model architecture, specifically as follows: S3.1: For radar images, visual features are extracted using a visual encoder; for electromagnetic signals, signal features are extracted using a pre-trained signal processing model. S3.2: Input the visual features or signal features into a multilayer perceptron that acts as an adapter, and the adapter maps the dimensions of the input features to dimensions that match the word embedding space of the base language model, so as to achieve cross-modal feature alignment; S3.3: The feature vector processed by the adapter is concatenated or fused with the corresponding text embedding vector and used as input to fine-tune the base language model.

5. The method according to claim 1, characterized in that, Step 4 specifically involves: S4.1: Combine the user-input question and the model-generated preliminary answer into a self-criticism prompt word; S4.2: Input the self-criticism prompts into the language model and guide the model to output a structured evaluation result containing criticisms, accuracy scores, and improvement suggestions; S4.3: If the structured evaluation results indicate that improvement is needed, the original question, preliminary answer, and criticisms are combined into an improvement prompt word; S4.4: Input the improved prompt words into the language model, guide the model to re-reason based on the criticisms, and generate an improved answer as the final output.

Images