Intelligent aided design method and device for heating system based on pre-training large model

Abstract

The embodiment of the invention discloses an intelligent aided design method and device for a heating system based on a pre-training large model, and relates to the field of engineering design. A user puts forward knowledge problems related to a heating system at a client side, such as module design parameter data, and performs related problem analysis on the heating system at a server side. The method comprises the following steps: firstly, acquiring heating system modules and system global knowledge data, constructing a training sample database, training a big language model, and analyzing answers of corresponding questions by using the trained big language model according to the to-be-answered heating system related knowledge questions provided by a client; related modules can be matched from a module database according to answers of question analysis, finally the answers of the corresponding questions and the modules are returned to a client, on this basis, engineers are assisted in design work of the heating system, the manual workload is reduced, meanwhile, high-accuracy module question and answer results are provided, and the working efficiency is improved.

Description

Technical Field

[0001] This invention relates to the field of engineering design technology for heating systems, and in particular to an intelligent assisted design method and device for heating systems based on a pre-trained large model. Background Technology

[0002] Traditional engineering design processes face numerous challenges. On the one hand, product design typically requires assembling numerous components into a complex whole, and manually designing the assembly relationships of these components is not only time-consuming but also prone to errors. On the other hand, with increasing product complexity, engineering drawing design work consumes a significant amount of time. To further improve the efficiency of designers, many AI-assisted design tools based on Large Language Models (LLMs) have been launched. How to combine large language models with engineering design work to achieve knowledge-based question answering in engineering design based on large models, thereby optimizing design efficiency and assembly accuracy, has become an important research direction. This not only requires solving technical challenges but also validating its feasibility and effectiveness in practical applications, thereby promoting the further development of engineering design technology.

[0003] While large-scale models have demonstrated powerful natural language processing capabilities and transformed vertical application paradigms across multiple fields, models like SMoE and ChatGLM have not yet been widely adopted in the industrial sector. Therefore, models trained on general corpora such as Bard cannot capture the specialized expressions of specific industrial domains. Furthermore, the scarcity of domain-specific data and the difficulty in obtaining high-quality labeled data for specialized fields, coupled with the model's reliance on general corpora for training, can lead to biases in the understanding of technical terms, such as providing incorrect or outdated standard data for heating systems.

[0004] These deviations, occurring during the engineering design process, can lead to very serious consequences, resulting in inaccuracies and unreliability issues in the application of large models to specific industrial sectors. Consequently, AI-assisted tools, such as Deepseek and Grox, which began widespread adoption several years ago, while experiencing rapid popularity and attempts to apply them to engineering design, still require review by experienced designers due to inaccuracies and unreliability in their results. This has also caused design deviations with serious consequences, leading designers to gradually abandon these AI tools in heating system design.

[0005] Therefore, how to assist engineers in designing heating systems, reducing manual workload while providing highly accurate Q&A results, has become a research topic. Summary of the Invention

[0006] The embodiments of the present invention provide an intelligent assisted design method and device for heating systems based on pre-trained large models, which can assist engineers in designing heating systems, reduce human workload, and provide highly accurate question-and-answer results.

[0007] To achieve the above objectives, the embodiments of the present invention adopt the following technical solutions:

[0008] A smart-aided design method for heating systems based on pre-trained large models includes:

[0009] S1. The server receives knowledge data and establishes a training sample database, wherein the knowledge data includes: module knowledge data of the heating system and global knowledge data of the heating system, and the training sample database is used to train a large language model.

[0010] S2. Receive the heating system problem to be processed uploaded by the questioning terminal, and use the large language model to analyze it to obtain the analysis result;

[0011] S3. Based on the analysis results, extract the corresponding heating system module from the heating system module database, and then feed back the extracted heating system module and answer information to the question terminal.

[0012] Specifically, S1 includes: forming the problem content corresponding to the module knowledge data and the global knowledge data of the heating system. The training data is in the form of a question, Q, and A, where Q represents the question and A represents the answer to the question. The sample database also includes operating data of the heating system, including collected temperature difference and heat load data. For example, the knowledge data uploaded by the client includes module knowledge data within the heating system, system-wide knowledge data, and the corresponding questions. The training data is composed of a question-and-answer format. The dataset in the training sample database includes temperature difference data, heat load data, corresponding module knowledge data of the heating system, and system-wide knowledge. During the training phase, question Q and answer A are used as input to the large language model, and the training data for the large language model is presented as text pairs.

[0013] The DCI-LLM model application framework used in this embodiment is divided into three parts: data processing, hybrid fine-tuning, and automatic design of complex systems. Complex system knowledge data undergoes data processing and extraction, and is divided into training and testing sets. The component library corresponding to the module knowledge will be used in subsequent mapping. The hybrid fine-tuning of DCI-LLM, also known as 'local-cooperative' fine-tuning, includes internal module knowledge training ('local training') and global system knowledge training ('cooperative training'). The 'local-cooperative' fine-tuning of DCI-LLM corresponds to the process of a mechanical engineer learning the overall system composition principle, i.e., first learning local module knowledge and then learning the system composition principle. We divide the model fine-tuning process into two stages. In the first stage, we use detailed knowledge of the system's internal modules and generate system module knowledge question-and-answer data using an API interface to perform LoRA (Low-Rank Adaptation) fine-tuning on Qwen2.5-7B. The model after the first stage of fine-tuning can answer basic questions about the system's internal modules; we name it DCI-LLM-T. However, it is prone to illusions when answering questions involving the global system. The second stage of fine-tuning processes the system's global knowledge data. Using limited global knowledge, a Freeze test is performed on the model fine-tuned in the first stage. The resulting DCI-LLM, while ensuring the basic questions within the question-answering system's modules, possesses the ability to answer system-related questions. Based on DCI-LLM, this framework can generate a final engineering drawing based on the answers to user-asked system questions. This final engineering drawing only requires engineers to verify its completeness and correctness.

[0014] Complex system knowledge data consists of global system knowledge and module knowledge. Global knowledge comprises partial local module knowledge and system composition knowledge. This paper uses Qwen2.5 as the baseline model and, based on the compositional characteristics of complex system knowledge data, fine-tunes the backbone model using a 'local-cooperative' approach to mine domain knowledge, capture the global compositional features and coupling relationships of the system, and, combined with post-processing programs, automatically draws complex system engineering diagrams to assist engineers in drafting. 'Local fine-tuning' builds upon module knowledge, enabling the model to learn module knowledge under corresponding operating conditions. 'Cooperative fine-tuning', based on the model's mastery of module knowledge, leverages global knowledge to guide the model in mining the coupling relationships between modules within the system, thereby learning the system's compositional structure and other knowledge under corresponding operating conditions. The DCI-LLM fine-tuning and inference process is as follows:

[0015] a) Knowledge distillation is performed using the knowledge of the internal modular components involved in the complex system. Data processing is performed on professional documents to make them meet the data requirements and formats for fine-tuning. The parameters of the source model are fine-tuned with a certain amount of training data. In the first stage, the professional knowledge data of each module in the system are used to fine-tune the model, so that the model learns the specific knowledge of each module under various working conditions.

[0016] (b) By fine-tuning the model using global system knowledge to minimize model illusions, the model learns the overall global knowledge of the system under various operating conditions, ultimately training a large model to assist in the engineering drawing of complex systems. Based on this, engineers only need to ask relevant questions to obtain Q&A on specific module knowledge under various operating conditions and Q&A on global system knowledge, and obtain the final system diagram.

[0017] The training of the large language model includes: a LoRA (Low-Rank Adaptation) fine-tuning process using the modular knowledge data of the heating system and a Freeze fine-tuning process using the global knowledge data of the heating system. The large language model comprises two training phases: a LoRA fine-tuning process based on modular knowledge and a Freeze fine-tuning process based on global system knowledge. In the LoRA fine-tuning process based on local modular knowledge, the large language model is trained using the model's local modular knowledge question-answering pairs. In the Freeze fine-tuning process based on global system knowledge, the large language model obtained from the LoRA fine-tuning based on local modular knowledge is named HSQLM-T (Heating-system Question Language Model-T). Building upon the local modular knowledge of HSQLM-T, it is guided to learn and master global system knowledge.

[0018] The LoRA fine-tuning process utilizing the modular knowledge data of the heating system includes: freezing all layers of the large language model and training for the newly added adapter parameters, wherein... , , A and B are the incremental weights used to approximate the fine-tuning of all parameters. Given two matrices, matrix A is initialized using Gaussian initialization, and matrix B is initialized with zero initialization. The number of fine-tuning parameters is... Reduce to d represents the d-dimensional parameter, r is the rank, and a is the hyperparameter. Indicates r d-dimensional matrix d r-dimensional matrix Indicates the model input, This represents the model output; during the training process for the newly added adapter parameters, gradient updates matrices A and B; weight merging is performed during inference deployment to eliminate inference latency.

[0019] The Freeze fine-tuning process utilizing the global knowledge data of the heating system includes: freezing the base layer of the large language model and updating the parameters of the last 5 layers of the large language model, wherein the frozen parameters are... , where i represents the index of the parameter, k represents the boundary index between the frozen layer and the trainable layer, i≤k; during the parameter update process, the gradient descent of the frozen parameters is zero;

[0020] The parameters involved in fine-tuning are j represents the index involved in training and updating, j>k, and t is the iteration number. Let L be the learning rate, and L be the loss function.

[0021] The intelligent assisted design method for heating systems based on a pre-trained large model provided in this invention offers several advantages. It allows for rapid construction of a training sample database; training the large language model using the sample database is quick and easy; and the problem analysis phase significantly reduces workload compared to manual calculations. It efficiently and accurately obtains heating system knowledge answers corresponding to the problem, and then uses these answers to locate the corresponding modules. These modules and their answers are then provided to engineers to assist them in completing the heating system design. By analyzing the heating system knowledge questions raised by engineers, the method can obtain answers to the knowledge related to heating system issues under specific operating conditions, thus assisting engineers in drawing up heating system diagrams. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 A schematic diagram of the hardware environment provided for an embodiment of the present invention;

[0024] Figure 2 A flowchart of the method provided in an embodiment of the present invention;

[0025] Figure 3 Flowchart for processing knowledge data in heating system case studies;

[0026] Figure 4 Example diagram showing details of data extraction and transformation;

[0027] Figure 5 This is a schematic diagram of the heating system module composition provided in an embodiment of the present invention;

[0028] Figure 6 This is a schematic diagram of a heating system fine-tuning training provided in an embodiment of the present invention;

[0029] Figure 7 This is a schematic diagram of a specific application interface provided in an embodiment of the present invention. Detailed Implementation

[0030] To enable those skilled in the art to better understand the technical solutions of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Embodiments of the present invention will be described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention. Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in the specification of the present invention means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or couplings. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items. It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the meaning consistent with their meaning in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless defined as herein.

[0031] This invention provides an intelligent design method for heating systems based on pre-trained large models. The method is applied to a heating system such as... Figure 1The client-server computing resource allocation system shown in the diagram involves the client providing the server with module knowledge data and global system knowledge data of the heating system. The server constructs a training sample database from this data and uses it to train a large language model. The client then inputs the heating system problem to be analyzed, and the server feeds this problem into the trained large language model to analyze the corresponding solution. The server then returns the solution and the relevant heating system modules to the client. The process of the large language model training method and apparatus based on heating system knowledge data is as follows: Figure 2 As shown, it includes:

[0032] The client submits questions and answers related to the heating system to the server, and the large model on the server performs rapid analysis and prediction of heating system problems.

[0033] The server accepts knowledge data uploaded by the client, including knowledge of heating system modules and global system knowledge.

[0034] A training sample database is constructed; a large language model is trained using the sample database; the trained large language model is used to analyze and predict the heating system problem to be processed. The object of analysis is the heating system knowledge problem provided by the questioning terminal. The questioning terminal can be a client that uploads knowledge data or a third-party terminal device with a client program installed.

[0035] The analysis results are used to filter heating system modules, and the corresponding modules and responses are returned to the client to assist engineers in drawing diagrams. The training sample database includes: module knowledge of the heating system and global knowledge of the heating system. The data processing flow and final format of the heating system are as follows: Figure 3 , Figure 4 As shown in the figure, these datasets include design parameters and operating conditions of internal modules of the heating system, as well as global knowledge of the heating system. The global knowledge consists of module knowledge and macro-level knowledge of the entire system. Question-and-answer data was constructed around the case knowledge base, with training data in a question-and-answer format. A total of 1770 question-and-answer data entries related to heating system case knowledge were compiled. Approximately 1070 of these entries relate to heating system modules, and approximately 700 relate to global knowledge of the heating system.

[0036] For highly precise details such as the inner diameter of the water pump and the location of the water tank, such data is of little reference value for HSQLM. Therefore, when processing case data, it is necessary to remove such overly detailed data that does not significantly affect the parameters required for the overall system composition. This article unifies the query portion of the heating system case knowledge into a question "Q" and uses the corresponding query results as the answer "A" for processing. The module composition of the heating system is as follows: Figure 5 As shown. The method for fine-tuning training a large language model is as follows: Figure 6 As shown. Specifically, heating system module knowledge data refers to the design parameters and installation locations of each module in the heating system under a certain temperature difference and heat load condition (i.e., operating condition). The overall system knowledge consists of heating system module knowledge, an overview of the system, and system functions.

[0037] The knowledge data is used as the model input, that is, the original data is represented as: , among which, among which, For the question, This is an answer to the relevant questions. In this embodiment, the training of the large language model using the sample database includes two training processes: LoRA fine-tuning based on heating system module knowledge and Freeze fine-tuning based on global heating system knowledge. Targeting the compositional characteristics of complex system knowledge data, the backbone model is fine-tuned using a 'local-cooperative' approach to perform domain knowledge mining, capture the global compositional features and coupling relationships of the system, and combine this with post-processing programs to automatically draw complex system engineering diagrams, thereby assisting engineers in drawing. 'Local fine-tuning' involves the model learning module knowledge under corresponding operating conditions based on module knowledge; 'cooperative fine-tuning' involves the model, based on its mastery of module knowledge, using global knowledge to guide the model in mining the coupling relationships between modules within the system, thereby learning the system's compositional structure and other knowledge under corresponding operating conditions.

[0038] In the LoRA fine-tuning based on heating system module knowledge, LoRA freezes all parts of the pre-trained model and trains only on the newly added low-rank matrix parameters (i.e., adapter parameters).

[0039] The training logic of LoRA can be explained through the following key parts: In LoRA, matrices are used... sum matrix Approximate expression of incremental weights in full parameter fine-tuning . ,in , At this point, the number of model fine-tuning parameters is from Reduce to .

[0040] For the matrix respectively Gaussian initialization is used for the matrix. Zero initialization is used, where Rank-based zero initialization ensures that model training starts from zero, preventing noise, while Gaussian initialization provides direction for random exploration early in model training. For hyperparameters... Initialize it to the settings of the model in the first experiment. , here It is a larger value that can encompass new knowledge while minimizing the impact of new knowledge. Approximating full parameter fine-tuning .

[0041] Based on this, when Smaller low-rank matrices provide a more concise representation of new knowledge but may miss information. The impact of new knowledge during forward propagation on the model is relatively large, amplifying its influence. When The larger the value, the richer the information contained in the low-rank matrix, and the more closely it relates to full parameter fine-tuning. Similar. Training Dynamics: Gradient updates only optimize the matrix. and Inference deployment: Weight merging, inference delay removal This indicates the model weights after fine-tuning. This represents the initial weight matrix of the model.

[0042] After the first stage, HSQLM-T has gained a relatively good grasp of the knowledge of the internal modules of the heating system, enabling it to answer some questions from engineers regarding module knowledge during engineering drawing, and demonstrating reasonable accuracy. To preserve HSQLM-T's performance in the question-answering task related to the internal modules of the heating system and to mitigate the illusion of a large model answering global system knowledge, this paper chooses to freeze the model's base layers in the second stage of fine-tuning, allowing updates only to the parameters of the last five layers. Regarding the frozen parameters... (i≤k): ,i represents the parameter index, and k represents the boundary index separating the frozen layer and the trainable layer. For the parameters involved in fine-tuning... (j>k): ,j represents the indexes involved in training and updating, which are updated using gradient descent. For the number of iterations, For learning rate, It is the loss function. During fine-tuning, for each frozen parameter... Its gradient descent is zero, that is Therefore, these parameters will not be updated during backpropagation.

[0043] By introducing Freeze tuning in the second stage, the powerful representation learning ability of the pre-trained model can be effectively preserved. Secondly, only a small number of parameters need to be fine-tuned to achieve adaptation to specific tasks, demonstrating excellent applicability in scenarios with limited training data or computational resources. This embodiment addresses the limited scale of global knowledge data in heating systems by employing a Freeze tuning strategy to effectively refine the model. The HSQLM model after Freeze tuning not only maintains its excellent performance in knowledge question answering tasks within heating system modules but also significantly improves its ability to handle global knowledge-related questions in heating systems.

[0044] For a specific example, this embodiment designs a fine-tuning of the heating system. It treats the knowledge data of the heating system modules and the global knowledge data of the heating system as problems within a large language model. and answers , in question and answer The training data for the model is in the form of '. This embodiment utilizes knowledge distillation based on the internal modular components involved in the heating system, processes the professional documents to make them conform to the data requirements and format for fine-tuning, and fine-tunes the parameters of the source model with a certain amount of training data. In the first stage, the model is fine-tuned using professional data from each module within the system, enabling the model to learn the specific knowledge of each module under various operating conditions.

[0045] By leveraging the system-wide knowledge of the heating system to fine-tune the model and minimize model illusions, the model learns the overall system knowledge under various operating conditions. This ultimately trains a large language model to assist in engineering drawing of the heating system. Through a two-stage training approach, the final fine-tuned model can provide question-and-answer sessions on specific module knowledge under various operating conditions, as well as on global system knowledge.

[0046] Figure 7 The final HSQLM Q&A usage was demonstrated. Engineers asked questions to HSQLM, HSQLM analyzed and provided answers, and the engineers then created engineering drawings based on the module numbers and answers they received.

[0047] An intelligent assisted design device for a heating system based on a pre-trained large model, comprising:

[0048] The data management module is used to receive knowledge data and establish a training sample database. The knowledge data includes module knowledge data of the heating system and global knowledge data of the heating system. The training sample database is used to train a large language model.

[0049] The analysis module is used to receive heating system problems to be processed uploaded by the questioning terminal, and to analyze them using the large language model to obtain analysis results;

[0050] The response module is used to extract the corresponding heating system module from the heating system module database based on the analysis results, and then feed back the extracted heating system module and the answer information to the question terminal.

[0051] The data management module is specifically used to compose [a set of data] based on the corresponding question content of the module knowledge data and the global knowledge data of the heating system. Training data in the form of For the question, This is to answer related questions; the sample database also includes operating data of the heating system, which includes collected temperature difference data and heat load data of the heating system.

[0052] The training of the large language model includes: a LoRA fine-tuning process using the module knowledge data of the heating system, and a Freeze fine-tuning process using the global knowledge data of the heating system.

[0053] The embodiments of the present invention provide an intelligent design method and device for heating systems based on a pre-trained large model. It can analyze heating system knowledge questions raised by engineers to obtain answers to the knowledge related to heating system issues under certain working conditions, and assist engineers in drawing system diagrams for heating systems.

[0054] The main advantages of this embodiment are: fast construction of training sample database; short training time and easy implementation of large language model through sample database; significant reduction in workload compared to manual calculation and design in the problem analysis stage; efficient and accurate acquisition of heating system knowledge answers corresponding to the problem; and based on the obtained answers, finding the corresponding module, providing the corresponding module and its answer to the engineer to assist in completing the design of the heating system.

[0055] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on its differences from other embodiments. In particular, the device embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. The above descriptions are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A smart-aided design method for a heating system based on a pre-trained large model, characterized in that, include: S1. The server receives knowledge data and establishes a training sample database, wherein the knowledge data includes: module knowledge data of the heating system and global knowledge data of the heating system, and the training sample database is used to train a large language model. S2. Receive the heating system problem to be processed uploaded by the questioning terminal, and use the large language model to analyze it to obtain the analysis result; S3. Based on the analysis results, extract the corresponding heating system module from the heating system module database, and then feed back the extracted heating system module and answer information to the question terminal.

2. The method according to claim 1, characterized in that, S1 includes: Based on the module knowledge data and the global knowledge data of the heating system, the corresponding question content is composed of... Training data in the form of For the question, For answers to related questions; The sample database also includes operating data of the heating system, which includes collected temperature difference data and heat load data of the heating system.

3. The method according to claim 1, characterized in that, The process of training a large language model includes: The LoRA fine-tuning process utilizes the module knowledge data of the heating system.

4. The method according to claim 2 or 3, characterized in that, The training of large language models also includes: The Freeze fine-tuning process utilizes the global knowledge data of the heating system.

5. The method according to claim 4, characterized in that, The LoRA fine-tuning process utilizing the module knowledge data of the heating system includes: Freeze all layers of the large language model and train it for the newly added adapter parameters, where, , , A and B are the incremental weights used to approximate the fine-tuning of all parameters. Given two matrices, matrix A is initialized using Gaussian initialization, and matrix B is initialized with zero initialization. The number of fine-tuning parameters is... Reduce to d represents the d-dimensional parameter, r is the rank, and a is the hyperparameter. Indicates r d-dimensional matrix d r-dimensional matrix Indicates the model input, This represents the model output; during the training process for the newly added adapter parameters, gradient update matrices A and B are performed; and weight merging is performed during inference deployment.

6. The method according to claim 5, characterized in that, The Freeze fine-tuning process utilizing the global knowledge data of the heating system includes: The base layer of the large language model is frozen, and the parameters of the last 5 layers of the large language model are updated. The frozen parameters are: , where i represents the index of the parameter, k represents the boundary index between the frozen layer and the trainable layer, i≤k; during the parameter update process, the gradient descent of the frozen parameters is zero; The parameters involved in fine-tuning are j represents the index involved in training and updating, j>k, and t is the iteration number. Let L be the learning rate, and L be the loss function.

7. An intelligent auxiliary design device for a heating system based on a pre-trained large model, characterized in that, include: The data management module is used to receive knowledge data and establish a training sample database. The knowledge data includes module knowledge data of the heating system and global knowledge data of the heating system. The training sample database is used to train a large language model. The analysis module is used to receive heating system problems to be processed uploaded by the questioning terminal, and to analyze them using the large language model to obtain analysis results; The response module is used to extract the corresponding heating system module from the heating system module database based on the analysis results, and then feed back the extracted heating system module and the answer information to the question terminal.

8. The apparatus according to claim 7, characterized in that, The data management module is specifically used to compose [a set of data] based on the corresponding question content of the module knowledge data and the global knowledge data of the heating system. Training data in the form of For the question, This is to answer related questions; the sample database also includes operating data of the heating system, which includes collected temperature difference data and heat load data of the heating system.

9. The apparatus according to claim 7, characterized in that, The process of training a large language model includes: The LoRA fine-tuning process utilizing the module knowledge data of the heating system, and the Freeze fine-tuning process utilizing the global knowledge data of the heating system.

Images