Industrial large model platform and its system

Abstract

An embodiment of the present application provides an industrial large model platform and its system. The industrial large model platform is used to train industrial large models. The industrial large model platform includes: an infrastructure layer, a base layer, a model adaptation layer, an interaction layer, and an application layer; the infrastructure layer is used to build the basic resources for training industrial large models; the base layer is used to perform industrial multi-modal pre-training to improve the processing ability of industrial large models on multi-modal data; the model adaptation layer is used to perform industrial mechanism embedded fine-tuning to improve the processing ability of industrial large models on specific tasks and the generalization ability in unknown field scenarios; the interaction layer is used to improve the interaction ability of industrial large models; the application layer is used to demonstrate the functions of industrial large models. The industrial large model platform and its system according to the embodiments of the present application solve the problems of poor generalization ability, low training efficiency, and lack of understanding of industrial mechanisms in existing industrial large models.

Description

Technical Field

[0001] This application relates to the field of computer applications, and in particular, to an industrial large model platform and its system. Background Art

[0002] With the advent of the Internet of Things era, industrial large models are built by establishing complex mathematical models and computer simulations to simulate, predict, and control the operation and behavior of industrial systems.

[0003] However, existing industrial large models have problems such as poor generalization ability, low training efficiency, and lack of understanding of industrial mechanisms. Among them, poor generalization ability means that existing industrial large models can only be applied to a single industrial scenario, lacking flexibility and adaptability; low training efficiency means that existing industrial large models need to be trained separately, which takes a long time and consumes a large amount of resources; lack of understanding of industrial mechanisms means that existing industrial large models lack in-depth understanding of the specific mechanisms in the industrial field, resulting in inaccurate and unreliable model output results.

[0004] Therefore, the poor generalization ability, low training efficiency, and lack of understanding of industrial mechanisms of existing industrial large models are problems that this application urgently needs to solve. Summary of the Invention

[0005] This application provides an industrial large model platform and its system to solve the problems of poor generalization ability, low training efficiency, and lack of understanding of industrial mechanisms of existing industrial large models.

[0006] In the first aspect of the embodiments of this application, an industrial large model platform is provided. The industrial large model platform is used to train an industrial large model, and the industrial large model platform includes: an infrastructure layer, a base layer, a model adaptation layer, an interaction layer, and an application layer;

[0007] The infrastructure layer is used to build the basic resources for training the industrial large model;

[0008] The base layer is used to perform industrial multi-modal pre-training on a pre-set industrial base large model through a pre-set multi-modal data set to improve the processing ability of the industrial large model on multi-modal data; among them, the industrial base large model is the base of the industrial large model, and the industrial base large model has the general solution ability for industrial tasks;

[0009] The model adaptation layer is used to perform industrial mechanism embedded fine-tuning on the pre-trained industrial base large model through a pre-set specific data set to obtain an industrial large model. The specific data set includes a task-oriented large model and an industry domain large model. The task-oriented large model is used to perform multi-task instruction fine-tuning on the pre-trained industrial base large model to improve the processing ability of the industrial large model in specific tasks. The industry domain large model is used to perform industry domain knowledge embedding and adapter fine-tuning on the industrial base large model after multi-task instruction fine-tuning to improve the generalization ability of the industrial large model in unknown domain scenarios.

[0010] The interaction layer is used to improve the interaction ability of the industrial large model through industrial efficient reasoning. The industrial large model is used to interact with technicians and cyber-physical systems respectively.

[0011] The application layer is used to demonstrate the functions of the industrial large model.

[0012] Optionally, for the industrial large model platform as described above, the task-oriented large model includes: intelligent question answering model, scenario recognition model, process decision-making model, terminal control model, content generation model, and scientific discovery model.

[0013] Then the functions of the industrial large model include: intelligent question answering, scenario recognition, process decision-making, terminal control, content generation, and scientific discovery.

[0014] Optionally, for the industrial large model platform as described above, the functions of the industrial large model are used to support applications in research and development design, production manufacturing, test and verification, operation and management, and operation and maintenance services.

[0015] Optionally, for the industrial large model platform as described above, the model adaptation layer is also used to perform scenario knowledge internalization and reinforcement self-training on the industrial base large model after industry domain knowledge embedding and adapter fine-tuning through pre-set limited labeled data to improve the processing ability of the industrial large model in industrial sub-scenarios.

[0016] Optionally, for the industrial large model platform as described above, the base layer is pre-set with tool chains for industrial multi-modal pre-training, industrial mechanism embedded fine-tuning, and industrial efficient reasoning respectively, as well as multiple industrial template large models. The industrial base large model is one of the multiple industrial template large models.

[0017] Optionally, for the industrial large model platform as described above, the basic resources include: industrial data, computing resources, and industrial knowledge.

[0018] Industrial data is used to provide training data for the industrial large model.

[0019] Computing resources are used to provide computing power support for the industrial large model.

[0020] Industrial knowledge is used to provide deep logic for the industrial large model.

[0021] Optionally, for the industrial large model platform as shown above, the industrial large model platform further includes: a pre-training module supported by the base layer;

[0022] The pre-training module is used to perform industrial multi-modal pre-training on the industrial base large model through at least one of multiple pre-training methods; among them, the multiple pre-training methods include: distributed training method, transfer learning training method, self-supervised learning method, and multi-task training method;

[0023] The pre-training module is also used to set the industrial data called by the industrial base large model, as well as the model parameters of the industrial base large model.

[0024] Optionally, for the industrial large model platform as shown above, the industrial large model platform further includes: an adaptation and fine-tuning module supported by the model adaptation layer;

[0025] The adaptation and fine-tuning module is used to perform industrial mechanism embedding and fine-tuning on the pre-trained industrial base large model through one of multiple fine-tuning training methods; among them, the multiple fine-tuning training methods include: industrial knowledge fusion and industrial mechanism embedding method, supervised fine-tuning method, direct preference optimization method, and human feedback reinforcement learning method.

[0026] Optionally, for the industrial large model platform as shown above, the industrial large model platform further includes: an interaction and reasoning module supported by the interaction layer;

[0027] The interaction and reasoning module is used to improve the interaction ability of the industrial large model through multiple interaction reasoning optimization methods; among them, the multiple interaction reasoning optimization methods include: inference acceleration, industrial knowledge base retrieval and enhanced generation, model orchestration, and plug-in mechanism.

[0028] Optionally, for the industrial large model platform as shown above, the industrial large model platform further includes: an evaluation module supported by the application layer;

[0029] The evaluation module is used to evaluate the function of the industrial large model through at least one of multiple evaluation methods; among them, the multiple evaluation methods include: automatic evaluation and manual evaluation.

[0030] The second aspect of the embodiments of the present application provides an industrial large model platform system, which includes: an industrial large model, and a cyber-physical system that interacts with the industrial large model;

[0031] The industrial large model is a model trained through the industrial large model platform according to any one of the first aspect.

[0032] An industrial large model platform and its system provided by an embodiment of the present application include an infrastructure layer, a base layer, a model adaptation layer, an interaction layer, and an application layer; the infrastructure layer is used to build the basic resources for training the industrial large model; the base layer is used for industrial multi-modal pre-training to improve the processing ability of the industrial large model on multi-modal data; the model adaptation layer is used for industrial mechanism embedded fine-tuning to improve the processing ability of the industrial large model on specific tasks and the generalization ability on unknown domain scenarios; the interaction layer is used to improve the interaction ability of the industrial large model; the application layer is used to demonstrate the functions of the industrial large model. The following technical effects are achieved: the model adaptation layer improves the generalization ability of the industrial large model on unknown domain scenarios through the large model in the industry field, and solves the problem of poor generalization ability of the existing industrial large models; the model adaptation layer improves the processing ability of the industrial large model on specific tasks through the industrial base large model, and solves the problem that the existing industrial large models lack the understanding of industrial mechanisms; the interaction layer improves the interaction ability of the industrial large model through industrial efficient reasoning, and solves the problem of low training efficiency of the existing industrial large models. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the technical solutions in the present application or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0034] Figure 1 It is an application scenario diagram of the industrial large model platform provided by an embodiment of the present application;

[0035] Figure 2 Structural schematic of the industrial large model platform provided by an embodiment of the present application Figure 1 ;

[0036] Figure 3 Structural schematic of the industrial large model platform provided by an embodiment of the present application Figure 2 。

[0037] Reference Signs:

[0038] 100 - Technical personnel;

[0039] 200 - Industrial large model platform; 210 - Infrastructure layer; 220 - Base layer; 221 - Pre-training module; 230 - Model adaptation layer; 231 - Adaptation fine-tuning module; 240 - Interaction layer; 241 - Interaction reasoning module; 250 - Application layer; 251 - Evaluation module;

[0040] 310 - Cyber system; 320 - Physical system. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0041] Exemplary embodiments will be described in detail herein, and examples thereof are shown in the accompanying drawings. When the following description refers to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present application. On the contrary, they are merely examples of devices and methods consistent with some aspects of the present application as detailed in the appended claims.

[0042] The technical solutions of the present application will be described in detail below with specific embodiments. These specific embodiments below can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. The embodiments of the present application will be described below with reference to the accompanying drawings.

[0043] In the embodiments of the present application, terms such as "first" and "second" are used to distinguish identical or similar items with basically the same functions and effects. Those skilled in the art can understand that the terms "first" and "second" do not limit the quantity and execution order, and the terms "first" and "second" do not necessarily mean different. It should be noted that in the embodiments of the present application, words such as "exemplary" or "for example" are used to represent examples, illustrations or explanations. Any embodiment or design solution described as "exemplary" or "for example" in the present application should not be construed as being more preferred or having more advantages than other embodiments or design solutions. Rather, the use of words such as "exemplary" or "for example" is intended to present related concepts in a specific manner. In the embodiments of the present application, "at least one" means one or more, and "a plurality" means two or more.

[0044] It should be noted that "when... " in the embodiments of the present application can be at the instant when a certain situation occurs, or within a period of time after a certain situation occurs. The embodiments of the present application do not make specific limitations on this. In addition, an industrial large model platform provided in the embodiments of the present application is only an example, and the industrial large model platform may also include more or less content.

[0045] For the convenience of clearly describing the technical solutions of the embodiments of the present application, some terms and technologies involved in the embodiments of the present application are briefly introduced below:

[0046] Internet of Things (IoT): It refers to a network that connects any item to the Internet through information sensing devices, such as devices based on Radio Frequency Identification (RFID), Global Navigation Satellite System (GNSS), infrared sensing, or laser scanning, and conducts information exchange and communication according to an agreed protocol to achieve intelligent identification, positioning, tracking, monitoring, and management. The application fields of the IoT are extensive, including multiple fields such as aerospace, ships, automobiles, chemical engineering, metallurgy, and steel.

[0047] Generalization ability of industrial models: It refers to the applicability and accuracy of industrial large models in unknown field scenarios. An industrial large model with good generalization ability can not only accurately simulate, predict, and control industrial scenarios already seen in the training dataset but also make reasonable predictions and judgments on unknown field scenarios. This ability helps industrial large models maintain high efficiency and reliability in complex and changing industrial environments.

[0048] Industrial mechanism: It refers to the professional knowledge such as principles, theorems, and laws in the production process of the industrial field. Industrial mechanisms include thermodynamic equations and chemical reaction kinetics equations, etc.

[0049] Understanding of the industrial mechanism of industrial models: It refers to the understanding and grasp of the interaction and influence mechanism among the internal components of an industrial system during the construction and operation of an industrial large model. Through in-depth understanding, the industrial large model can more accurately simulate and predict the behavior of the industrial system, improve the accuracy and reliability of the model's industrial mechanism. At the same time, this understanding also helps to discover problems and bottlenecks in the industrial system and provides strong support for optimizing and improving the industrial system.

[0050] To clearly understand the technical solution of this application, the solutions of the prior art will be introduced in detail first.

[0051] An industrial large model refers to a model constructed by using complex mathematical models and computer simulations that can simulate, predict, and control the operation of an industrial system. Industrial large models usually involve knowledge in multiple disciplinary fields and aim to achieve precise grasp of the overall performance and behavior of industrial systems through highly integrated and comprehensive analysis methods.

[0052] However, existing industrial large models have problems such as poor generalization ability, low training efficiency, and lack of understanding of industrial mechanisms. Among them, poor generalization ability means that existing industrial large models can only be applied to a single industrial scenario, lacking flexibility and adaptability; low training efficiency means that existing industrial large models need to be trained separately, which takes a long time and consumes a large amount of resources; lack of understanding of industrial mechanisms means that existing industrial large models lack in-depth understanding of the specific mechanisms in the industrial field, resulting in inaccurate and unreliable model outputs.

[0053] In the research, it is found that for the problem of poor generalization ability, the generalization ability of industrial large models can be improved through pre-set large models and diverse training methods, enabling them to better adapt to different industrial application scenarios; for the problem of low training efficiency, through efficient model training means, fast training of industrial large models can be achieved on the platform, improving training efficiency and reducing time and costs; for the problem of lack of understanding of industrial mechanisms, by introducing the method of fine-tuning with mechanism embedding in the industrial knowledge base, the understanding of industrial mechanisms by industrial large models can be enhanced, improving the accuracy and effectiveness of the models in practical applications.

[0054] Based on the above creative findings, the industrial large model platform and its system of this application are proposed. The industrial large model platform is a one-stop large model development and service operation platform, covering multiple key links such as model training, model fine-tuning, model inference, and model evaluation. In the model training link, it not only provides dataset management and model fine-tuning functions, but also pre-sets multiple large models for users to choose from, supporting diverse training methods to meet different needs; in the model inference link, the industrial large model platform improves the inference speed and the quality of model outputs through various interactive inference optimization methods; in the model evaluation link, the industrial large model platform provides two ways of automatic evaluation and manual evaluation to ensure a comprehensive and accurate understanding of the model's performance.

[0055] Next, the application scenarios of the industrial large model platform provided by the embodiments of this application will be introduced. Figure 1 This is the application scenario diagram of the industrial large model platform provided by the embodiments of this application. It should be noted that Figure 1 The shown is only an example of the scenario where the embodiments of this application can be applied, to help those skilled in the art understand the technical content of this application, but it does not mean that the embodiments of this application cannot be used in other devices, systems, environments or scenarios.

[0056] As Figure 1 shown, this application scenario includes three interactive objects: technician 100, industrial large model platform 200, and Cyber-Physical Systems (CPS). The three form an organic whole to interact and assist in the industrial production process.

[0057] The technical personnel 100 can be various types of specialized personnel, including R & D personnel, engineering personnel, production operators, and operation and management personnel, etc. The technical personnel 100 are used to train the industrial large model through the industrial large model platform 200 according to the different goals and plans of specific industrial tasks and interact with the industrial large model.

[0058] The cyber-physical system is a complex system that includes computing, networking, and physical entities. It realizes the interaction between the cyber space and the physical world through the organic integration and in-depth cooperation of 3C (Computing, Communication, Control) technologies. In the cyber-physical system, the cyber system 310 (Cyber System) and the physical system 320 (Physical System) are two important components and are the objects for the industrial large model to perceive, act, and control.

[0059] The cyber system 310 refers to the technologies and systems related to computing, communication, and control, covering multiple aspects such as computer networks, information systems, and control algorithms. It is the core part of the cyber-physical system responsible for data processing, information transmission, and decision-making. The cyber system 310 can include memories, computing units, databases, industrial software, and industrial networks, etc.

[0060] The physical system 320 refers to the systems that exist in the real world and have physical properties and dynamic behaviors. The physical system 320 can include shift levers, engine governors, robots, machine tools, and automated guided vehicles, etc.

[0061] The embodiments of the present application will be introduced below with reference to the accompanying drawings of the specification. Figure 2 The structural schematic of the industrial large model platform provided by the embodiments of the present application Figure 1 The industrial large model platform provided by the embodiments of the present application is used to train the industrial large model. As Figure 2 shown, the industrial large model platform 200 includes: an infrastructure layer 210, a base layer 220, a model adaptation layer 230, an interaction layer 240, and an application layer 250.

[0062] The infrastructure layer 210 is used to build the basic resources for training the industrial large model.

[0063] Specifically, the infrastructure layer 210 is a key layer in the training of industrial large models. In industrial scenarios, especially in scenarios that require processing massive amounts of data and performing complex computing tasks, such as in multiple fields including aerospace, shipbuilding, automotive, chemical, metallurgy, and steel, building industrial large models has extremely high requirements for storage resources, computing resources, and data processing capabilities. The infrastructure layer 210 is designed and deployed to meet these requirements. The infrastructure layer 210 ensures the smooth progress of the training process and provides a solid foundation for the ultimate application of industrial large models.

[0064] The base layer 220 is used to perform industrial multi-modal pre-training on a pre-set industrial base large model through a pre-set multi-modal data set, so as to improve the processing ability of the industrial large model on multi-modal data; among them, the industrial base large model is the base of the industrial large model, and the industrial base large model has the general solution ability for industrial tasks.

[0065] Specifically, the base layer 220 is the core support in the training of industrial large models. The pre-training of the industrial base large model involves using a large-scale multi-modal data set that has nothing to do with downstream tasks for the initial training of the industrial large model, that is, industrial multi-modal pre-training, so that the industrial large model has the general processing ability to understand and process multi-modal data in industrial scenarios and is improved. Among them, the multi-modal data set can be pre-set when building the industrial large model platform, or can be pre-set by the user when training the industrial large model according to the industrial large model platform.

[0066] At least one large model can be pre-set in the base layer 220. The industrial base large model can be selected by the user from these large models by himself, or can be selected by the user from these large models through a certain selection algorithm. The industrial base large model has the general solution ability for industrial tasks, providing a basis for the more refined model adaptation and scenario application of the industrial large model.

[0067] Furthermore, the base layer 220 is used to perform industrial multi-modal pre-training on a pre-set industrial base large model or a user-built large model through a pre-set multi-modal data set and / or a user-built data set, so as to improve the processing ability of the industrial large model on multi-modal data.

[0068] Specifically, the base layer 220 can perform industrial multi-modal pre-training through the above-mentioned multi-modal data set, or a data set built by the user himself, that is, a user-built data set, or a data set including all the data of the multi-modal data set and the user-built data set. It should be noted that when performing industrial multi-modal pre-training through the user-built data set, the improvement degree of the processing ability of the industrial large model on multi-modal data is affected by the data value of the user-built data set.

[0069] The training object for industrial multi-modal pre-training can be the above-mentioned industrial base large model or a large model built by the user; among them, the large model built by the user can be a large model built by the user himself or a large model built by the user based on the industrial base large model. It should be noted that when using the large model built by the user as the training object, the degree of improvement in the processing ability of the pre-trained large model built by the user on multi-modal data is affected by the modeling degree of the large model built by the user.

[0070] The model adaptation layer 230 is used to perform in-industry mechanism embedding fine-tuning on the pre-trained industrial base large model through a preset specific data set to obtain an industrial large model; among them, the specific data set includes a task-oriented large model and an industry domain large model; the task-oriented large model is used to perform multi-task instruction fine-tuning on the pre-trained industrial base large model to improve the processing ability of the industrial large model on specific tasks; the industry domain large model is used to perform industry domain knowledge embedding and adapter fine-tuning on the industrial base large model after multi-task instruction fine-tuning to improve the generalization ability of the industrial large model in unknown domain scenarios.

[0071] Specifically, after pre-training, the industrial large model can obtain a certain degree of general-purpose generalization ability to handle industrial tasks. However, in industrial multi-scenario tasks, the industrial large model usually lacks professional knowledge and mechanisms in the industrial field, resulting in difficulty in accurately understanding and processing industrial problems and poor credibility of the output results. Therefore, multi-task instruction fine-tuning is required.

[0072] The model adaptation layer 230 is the main part when training the industrial large model. The secondary training of the industrial base large model, that is, in-industry mechanism embedding fine-tuning, involves using a small-scale specific data set to perform industrial task adaptation and industry domain adaptation on the pre-trained industrial base large model, enabling the industrial large model to have the processing ability on specific tasks and the generalization ability in unknown domain scenarios and improving them.

[0073] Industrial task adaptation means, based on the pre-trained industrial base large model, performing multi-task instruction fine-tuning through the task-oriented large model, so that while retaining the general problem-solving ability of the industrial large model in industrial tasks, it has the processing ability on specific tasks and improves it.

[0074] Industry domain adaptation means, based on the industrial base large model after multi-task instruction fine-tuning, performing industry domain knowledge embedding and adapter fine-tuning through the industry domain large model, so that the industrial large model has the generalization ability in unknown domain scenarios and improves it. Among them, the industry domain includes discrete manufacturing fields such as aerospace, shipbuilding, and automotive, as well as process manufacturing fields such as chemical industry, metallurgy, and steel.

[0075] The commonly used instruction fine-tuning method for industrial large models is to fine-tune the pre-trained industrial base large model on a paired set consisting of human instructions and expected outputs. In order to integrate industrial knowledge into industrial large models, the industrial large models in the embodiments of this application, compared with industrial large models in the prior art, also require industrial knowledge fusion and industrial mechanism embedding methods, that is: First, relevant industrial knowledge needs to be collected from multiple sources, including empirical knowledge such as process flows and technical specifications. Second, these industrial knowledge are converted into forms that industrial large models can understand and process, for example, creating triple data in a knowledge graph (including head entities, relationships, and tail entities), and using knowledge graph embedding technology to convert knowledge into embedded vector features for training. Finally, to enable industrial large models to integrate industrial knowledge, the knowledge graph can be integrated into the training objective and instruction fine-tuning of the knowledge graph can be performed.

[0076] In order to embed the equations of industrial mechanisms into industrial large models, the industrial mechanisms can be characterized as feature information recognizable by neural networks and embedded into the network architecture. Second, at the model output stage, guide the model to output in line with industrial physical laws. For example, add mechanism equations to the loss function to punish output results that violate physical laws. In addition, the output of the model can be checked and constrained to conform to industrial mechanisms according to the rules in the mechanism knowledge base, and simulation tools or humans can also be used to verify the credibility and accuracy of the model output.

[0077] In order to make industrial large models have stronger adaptability when applied to various industries, further industry knowledge embedding fine-tuning is required, so that industrial large models can be more proficient in the professional knowledge of a certain industry, so as to perform excellently in various industries. In the fine-tuning process of industry domain large models, using adapter fine-tuning can enhance the adaptability of industrial large models to different industries while retaining the fine-tuning of task-oriented large models. Next, detailed industry datasets need to be prepared, and the model performance is fine-tuned on these industry-specific datasets to improve the performance of industrial large models in industry-specific problems.

[0078] Furthermore, the model adaptation layer 230 is used to perform industrial mechanism embedding fine-tuning on the pre-trained industrial base large model or the user-built large model through pre-set specific datasets and / or user-built datasets to obtain an industrial large model.

[0079] The implementation principle and technical effect of the industrial mechanism embedding fine-tuning provided in this embodiment are similar to those of the industrial multi-modal pre-training in the above embodiment, and will not be elaborated here in this embodiment.

[0080] The interaction layer 240 is used to improve the interaction ability of the industrial large model through industrial efficient reasoning; among them, the industrial large model is used to interact with technicians and cyber-physical systems respectively.

[0081] Specifically, in the interaction layer 240, technicians, industrial large models, and cyber-physical systems are the three interaction objects that carry out interactive collaboration.

[0082] Industrial efficient reasoning enables industrial large models to quickly and accurately process data and make decisions in complex industrial environments. Industrial efficient reasoning can be achieved based on methods such as model compression, hardware acceleration, and enhanced generation of industrial retrieval.

[0083] Furthermore, the interaction layer 240 trains the industrial large model into an agent, and technicians, agents, and cyber-physical systems carry out interactive collaboration. Among them, the agent integrates several basic capabilities of the industrial large model and enhances memory, planning, action, and perception to interact with the outside world. Compared with the industrial large model, the agent has the further ability to actively perceive and control the industrial environment and can spontaneously remember, observe, and influence the external environment.

[0084] The application layer 250 is used to demonstrate the functions of the industrial large model.

[0085] Specifically, the application layer 250 is the manifestation form of the industrial large model on the server side. The functions of the industrial large model enable the industrial large model to not only answer complex queries, but also understand and analyze complex industrial environments, make scientific process decisions, directly control cyber-physical systems, and automatically generate technical documents and simulation designs.

[0086] The industrial large model platform of the embodiments of the present application includes an infrastructure layer, a base layer, a model adaptation layer, an interaction layer, and an application layer; the infrastructure layer is used to build the basic resources for training the industrial large model; the base layer is used to perform industrial multi-modal pre-training to improve the processing ability of the industrial large model on multi-modal data; the model adaptation layer is used to perform industrial mechanism embedded fine-tuning to improve the processing ability of the industrial large model on specific tasks and the generalization ability on unknown domain scenarios; the interaction layer is used to improve the interaction ability of the industrial large model; the application layer is used to demonstrate the functions of the industrial large model. The following technical effects are achieved: the model adaptation layer improves the generalization ability of the industrial large model on unknown domain scenarios through the industry domain large model, and solves the problem of poor generalization ability of the existing industrial large models; the model adaptation layer improves the processing ability of the industrial large model on specific tasks through the industrial base large model, and solves the problem that the existing industrial large models lack the understanding of industrial mechanisms; the interaction layer improves the interaction ability of the industrial large model through industrial efficient reasoning, and solves the problem of low training efficiency of the existing industrial large models.

[0087] In a possible design, the task-oriented large model includes: an intelligent question answering model, a scenario recognition model, a process decision model, a terminal control model, a content generation model, and a scientific discovery model;

[0088] The functions of the industrial large model include: intelligent Q&A, scenario awareness, process decision-making, terminal control, content generation, and scientific discovery.

[0089] Specifically, the intelligent Q&A function means that the industrial large model can understand and answer complex industrial-related questions, providing instant information support. With extremely strong text understanding and generation capabilities, combined with specific knowledge in the industrial field and related technologies, the industrial large model can provide Q&A services for the entire process tasks at any time and place.

[0090] The scenario awareness function means that the industrial large model can identify and understand various scenarios and states in the industrial environment, providing a basis for further analysis and decision-making. The industrial large model has powerful perception capabilities for information such as vision and sensing, and incorporates it into the decision-making process to optimize the analysis and decision-making effects.

[0091] The process decision-making function means that the industrial large model can make optimized decisions in complex industrial processes, improving production efficiency and quality. Based on its own powerful logical reasoning and calculation capabilities, the industrial large model can provide reasonable solutions to various optimization problems abstracted from industrial processes.

[0092] The terminal control function means that the industrial large model can directly control industrial equipment to achieve automated operation. Embodied intelligence is one of the important capabilities of the industrial large model. Based on the obtained decision results and formed processing plans, the industrial large model can directly control production equipment to complete industrial tasks.

[0093] The content generation function means that the industrial large model can create content such as technical documents, design plans, and reports, improving work efficiency. Based on its huge network architecture and trained with a large amount of data, the industrial large model has extremely strong content generation capabilities and is widely used in applications such as code generation, image generation, and time series data generation.

[0094] The scientific discovery function means that the industrial large model has advanced correlation capabilities, thus promoting the development and application of new materials, new processes, new products, and new models. For example, through large-scale interdisciplinary knowledge integration, the industrial large model can identify the coupling mechanisms of complex products in multiple fields such as machinery, electricity, hydraulics, heat, gas, and magnetism, clarify the physical and chemical principles in new product design, and provide support for innovative process design plans.

[0095] The technical effect of this solution in this embodiment is: Through the task-oriented large model, the industrial large model has the functions of intelligent Q&A, scenario awareness, process decision-making, terminal control, content generation, and scientific discovery.

[0096] In a possible design, the functions of the industrial large model are used to support applications in R & D design, production manufacturing, test and verification, operation management, and operation and maintenance services.

[0097] Specifically, the industrial large model is functionally oriented towards the entire life cycle of industrial manufacturing and is used to support applications in five industrial links: R & D design, production manufacturing, test and verification, operation management, and operation and maintenance services.

[0098] R & D design involves the formation of new product concepts, principle design, prototype production, simulation design to determine the final product specifications and performance, so as to ensure that the product can meet market demands and technical standards. In the R & D design stage, the industrial large model mainly provides process assistance for design guidance, solution generation, and auxiliary design. The functions that can be achieved in R & D design include intelligent Q&A assistants and industrial content generation.

[0099] For intelligent Q&A assistants, the industrial large model has a comprehensive and professional industrial knowledge base, which can serve as an industrial knowledge Q&A assistant to provide corresponding suggestions on the processes, parameter types, and design guidelines of various product designs, and conduct rationality demonstration and inspection of the design solutions by combining industrial knowledge and mechanism understanding, simplifying the design process.

[0100] For industrial content generation, based on the formation of the design solution, the industrial large model uses its own content generation ability to generate corresponding engineering documents for the reasonable design results, such as Computer Aided Design (CAD) files or Computer Aided Engineering (CAE) files, etc., forming seamless connection with other industrial software to prepare for the subsequent production stage. At the same time, based on its own learning of industrial domain knowledge and mechanisms, the industrial large model can spontaneously and innovatively explore and generate new structures, new integration solutions, and new concepts, providing innovative ideas for designers. In addition, the industrial large model can design corresponding simulation systems to conduct simulation analysis on the design solutions, verify the rationality of parameter and structure designs, and the formed simulation analysis results are used to iteratively optimize the design solutions.

[0101] Production manufacturing is the process of transforming the R & D - designed products into actual items, including links such as raw material procurement, production planning, production line optimization, processing and assembly, and quality inspection. In the production manufacturing stage, the industrial large model mainly provides process assistance for product process generation, production equipment control, production plan arrangement, and production defect identification. The functions that can be achieved in production manufacturing include operation process guidance and production process control.

[0102] For operation process guidance, the industrial large model has the ability to possess professional processing technology knowledge, can guide the process design process, and interact with personnel to achieve operation process guidance, personnel training, and status feedback interaction, etc. At the same time, the industrial large model monitors the production process, and personnel can inquire about the production status from the industrial large model in real - time to control the production rhythm.

[0103] For production process control, terminal control is the key application of industrial large models in the manufacturing stage. Terminal control mainly includes two parts: production equipment motion control and production process control. Based on perception and decision-making capabilities, industrial large models form intelligent control solutions for equipment, such as robot arm movement speed, acceleration and direction, etc., to improve the performance, efficiency and flexibility of the motion system to better complete specific manufacturing tasks.

[0104] Experimental testing is to test and evaluate the performance, reliability and safety of manufactured products to ensure that the products meet the design requirements and industry standards and can meet the needs of users. In the experimental testing phase, the industrial large model mainly assists in the process of test plan design, test equipment control, test result analysis and test report generation. The functions that can be achieved by experimental testing include test environment recognition and test report generation.

[0105] In terms of test environment cognition, experimental testing involves various changing test environments. In this process, the industrial large model perceives environmental changes through various sensor data, further responds to the environment, and analyzes the corresponding test results. At the same time, the industrial large model also monitors the tested products in real time, perceives changes in product performance, and identifies product status characteristics.

[0106] For test report generation, the industrial big model receives detailed test requirements and design parameters, can generate specific test plans, and further generate test files executable by the equipment. After the test is completed, it generates a corresponding test analysis report to assist R&D personnel in product optimization design.

[0107] Operation management is the process of planning, organizing, coordinating, controlling and optimizing various aspects such as production, sales, finance and human resources in order to achieve its business goals. In the operation management stage, the industrial big model mainly assists in resource scheduling, production planning, document generation and report analysis. The functions that can be realized by operation management include operation status Q&A and operation report generation.

[0108] As for questions and answers about operational situations, the industrial big model is integrated into the entire production process. Managers can understand the production situation at any time through the industrial big model. From raw material supply to production line processing to product sales, they can interact with the industrial big model in the form of questions and answers. At the same time, it also includes auxiliary personnel to retrieve various industrial documents, which greatly facilitates managers' control over the entire process.

[0109] For the generation of operation reports, the industrial large model combines perception and decision-making capabilities, can generate specific scheduling plans, and can generate plans according to industrial documents or in line with the requirements of other industrial software. In addition, managers can use the industrial large model to generate production management reports. The industrial large model can generate report content with rich information according to the specific needs of managers, improving management efficiency.

[0110] Operation and maintenance services refer to the technical support and after-sales services provided by manufacturing enterprises to users after the products are sold, including installation, commissioning, repair, maintenance, and upgrade, etc., to ensure that the products can operate normally and stably; at the same time, it also includes the continuous monitoring of the internal manufacturing equipment and product quality of the enterprise. In the operation and maintenance service stage, the industrial large model mainly assists in processes such as fault diagnosis, after-sales service, and intelligent customer service. The functions that can be realized in operation and management include intelligent after-sales service and intelligent fault-tolerant maintenance.

[0111] For intelligent after-sales service, the industrial large model monitors the operating status of equipment and products, provides fault warnings and maintenance Q&A, and can interact with customers as an intelligent after-sales customer service, sorting out various email summaries and reducing labor costs.

[0112] For intelligent fault-tolerant maintenance, when the equipment fails or its performance degrades, the industrial large model can perform corresponding operation fault-tolerant control to ensure that the equipment can still maintain its specified functions when a fault occurs or continue to work at the cost of sacrificing performance. Based on intelligent detection and diagnosis of perception and decision-making, corresponding control measures are taken to deal with system failures, ensuring the safety and reliability of equipment operation, such as adjusting equipment operation parameters, slowing down the production rhythm, switching to standby systems, and giving timely fault warnings, etc., providing time for maintenance and reducing the accident rate.

[0113] The technical effect of this solution in this embodiment is: to specifically apply the functions of the industrial large model.

[0114] In a possible design, the model adaptation layer 230 is further used to perform scenario knowledge internalization and enhanced self-training on the industrial base large model after industry domain knowledge embedding and adapter fine-tuning through preset limited labeled data, so as to improve the processing ability of the industrial large model in industrial sub-scenarios.

[0115] Specifically, scenario knowledge internalization and enhanced self-training aims to address the problem of limited labeled data in some sub-scenarios of industrial industry knowledge. Through preliminary training on small-scale labeled data, the model is used to generate pseudo-labels to label large-scale unlabeled data, and a reinforcement learning reward mechanism is introduced to optimize the quality of pseudo-labels, thereby improving the model performance. In the case of limited labeled data, scenario knowledge internalization and enhanced self-training can support the fine-tuning of scenario large models for solving industrial sub-scenario problems, so as to improve the processing ability of the industrial large model in industrial sub-scenarios.

[0116] In this embodiment, the technical effect of this solution is: by internalizing scenario knowledge and strengthening self-training, the processing ability of the industrial large model in industrial sub-scenarios is improved.

[0117] In a possible design, the base layer 220 is preconfigured with toolchains for industrial multi-modal pre-training, industrial mechanism embedded fine-tuning, and industrial efficient inference respectively, as well as multiple industrial template large models; among them, the industrial base large model is one of the multiple industrial template large models.

[0118] Specifically, the base layer 220 is preconfigured with a toolchain for industrial multi-modal pre-training, so that it can be called at any time when the base layer 220 pre-trains the preconfigured industrial base large model; the base layer 220 is also preconfigured with a toolchain for industrial mechanism embedded fine-tuning, so that it can be called at any time when the model adaptation layer 230 performs secondary training on the pre-trained industrial base large model; the base layer 220 is also preconfigured with a toolchain for industrial efficient inference, so that it can be called at any time when the interaction layer 240 improves the interaction ability of the industrial large model. In addition, the base layer 220 is also preconfigured with multiple industrial template large models, and technicians can select an industrial template large model from these industrial template large models as the industrial base large model according to requirements.

[0119] In this embodiment, the technical effect of this solution is: by preconfiguring multiple toolchains and multiple industrial template large models in the base layer, the problem of platform support for industrial large models is solved.

[0120] In a possible design, the basic resources include: industrial data, computing resources, and industrial knowledge;

[0121] Industrial data is used to provide training data for the industrial large model;

[0122] Computing resources are used to provide computing power support for the industrial large model;

[0123] Industrial knowledge is used to provide in-depth logic for the industrial large model.

[0124] Specifically, industrial data includes Computer-Aided X (CAX) files, industrial time-series data, machine instructions, industrial documents, and multimodal data (such as images, videos, and audio). Industrial data is the basis for model training and operation. Computing resources include cloud-edge-end computing power and storage for industrial large model training and inference, as well as chips designed specifically for Artificial Intelligence (AI) operations. Industrial knowledge includes general industrial knowledge and enterprise private knowledge, covering structured documents such as industry specifications, operation manuals, machine operation principles, and maintenance experience, as well as dedicated knowledge graphs, providing in-depth logic for decision-making and analysis of industrial large models.

[0125] In this embodiment, the technical effect of the solution is that by pre-setting industrial data, computing resources, and industrial knowledge in the infrastructure layer, the problem of basic resources for industrial large models is solved.

[0126] Figure 3 Structural schematic of the industrial large model platform provided by the embodiment of the present application Figure 2 As Figure 3 shown, the industrial large model platform provided in this embodiment is a further refinement based on the industrial large model platform provided in the previous embodiment of the present application. Then, the industrial large model platform 200 provided in this embodiment further includes: a pre-training module 221 supported by a base layer 220;

[0127] The pre-training module 221 is used to perform industrial multimodal pre-training on the industrial base large model through at least one of multiple pre-training methods. Among them, the multiple pre-training methods include: distributed training method, transfer learning training method, self-supervised learning method, and multi-task training method.

[0128] Specifically, in the pre-training module 221, diverse pre-training methods are supported, including distributed training method, transfer learning training method, self-supervised learning method, and multi-task training method, etc. They can be combined arbitrarily to meet the training needs of different technicians and scenarios, ensuring that technicians can select appropriate pre-training methods according to their own needs.

[0129] The pre-training module 221 is also used to set the industrial data called by the industrial base large model, as well as the model parameters of the industrial base large model.

[0130] Specifically, the execution steps of the pre-training module 221 include data preprocessing, model construction, and model training, etc.

[0131] For data preprocessing, the pre-training module 221 supports unified management of industrial data in the infrastructure layer 210. Technicians can view the basic information of industrial data, including version number (ID), data volume, import logs, etc. Technicians can customize industrial data, set the data name, select the data type, and set the storage location of the data. When customizing data, technicians can combine industrial data with functions such as data annotation, data cleaning, and data augmentation to build industrial data suitable for the generative large model scenario. Among them, data cleaning includes operations such as converting traditional Chinese to simplified Chinese, converting uppercase to lowercase, and deleting abnormal characters in industrial data. For the domain-related corpus of technicians, distribution statistics and quality inspection of text data can be performed. This function can output the overall distribution and quality assessment results of the domain corpus, including the distribution statistics of the domain type and task type of the data, as well as preliminary data quality inspection. In addition, to help technicians quickly get started and verify the training effect, the pre-training module 221 can preset some industrial data specific to the industrial industry. Technicians can directly use these preset industrial data for training and testing to accelerate the model development process.

[0132] During the process of model construction and training, technicians can also set the model parameters of the industrial large model. The pre-training module 221 supports a variety of optimization algorithms and parallel computing architectures to ensure high efficiency and accuracy in large-scale data processing. In addition, technicians can also monitor the training progress in real time through visualization tools, analyze the performance of the model, and perform necessary tuning operations. After the industrial large model is trained, the pre-training module 221 provides a one-click deployment function, and technicians can quickly apply the trained industrial large model to the production environment to achieve intelligent decision-making, process optimization, or other specific business requirements. At the same time, the pre-training module 221 also supports online update and iterative optimization of the industrial large model to ensure its high efficiency and accuracy in constantly changing industrial scenarios.

[0133] The technical effect of this solution in this embodiment is: through the pre-training module supported by the base layer, a variety of training methods are used to pre-train the industrial base large model.

[0134] In a possible design, the industrial large model platform 200 further includes: an adaptation and fine-tuning module 231 supported by the model adaptation layer 230;

[0135] The adaptation and fine-tuning module 231 is used to perform industrial mechanism embedded fine-tuning on the pre-trained industrial base large model through one of a variety of fine-tuning training methods; among them, the variety of fine-tuning training methods include: industrial knowledge fusion and industrial mechanism embedding method, supervised fine-tuning method, direct preference optimization method, and human feedback reinforcement learning method.

[0136] Specifically, after the pre-training module 221 completes the pre-training of the industrial base large model, the adaptation and fine-tuning module 231 will fine-tune the model to enhance its performance. In the adaptation and fine-tuning module 231, diverse fine-tuning training methods are supported, including industrial knowledge fusion and industrial mechanism embedding methods, supervised fine-tuning methods, direct preference optimization methods, and human feedback reinforcement learning methods, etc. Technicians can select a suitable fine-tuning training method according to their own needs.

[0137] The industrial knowledge fusion and industrial mechanism embedding method creates triple data in the knowledge graph and uses knowledge graph embedding technology to convert knowledge into embedded vector features for training, enabling the industrial large model to fuse industrial knowledge. Secondly, in order to embed the industrial mechanism model into the industrial large model, the industrial mechanism can be characterized as feature information recognizable by the neural network and embedded into the network architecture. At the model output stage, guide the model to output in line with industrial physical laws. For example, add a mechanism equation to the loss function to punish the output results that violate physical laws.

[0138] Supervised fine-tuning uses a pre-trained neural network model and retrains it on a small amount of supervised data to adapt to specific task requirements. Based on the pre-trained industrial base large model, the process of additional training is usually to make the industrial large model better adapt to a specific field or task.

[0139] The direct preference optimization method precisely controls the output of the industrial large model by directly optimizing the output of the language model without the need for reinforcement learning. This method can accurately understand and learn the preferences of technicians and achieve more prominent effects.

[0140] Human feedback reinforcement learning means that by incorporating human feedback into the training process, the industrial large model can quickly master human experience. The preferences of humans are used as reward signals to guide the training of the model, thereby enhancing the model's understanding and satisfaction of human intentions. The success of human feedback reinforcement learning depends on the quality of the feedback provided by humans. The adaptation and fine-tuning module 231 has developed effective and scalable methods to collect and process such feedback. The adaptation and fine-tuning module 231 accesses multi-round conversations and ranking and prompt corpus datasets as human feedback for reinforcement learning (RLHF) training from human feedback to further improve the intelligence and efficiency of the industrial large model.

[0141] The technical effect of this solution in this embodiment is: through the adaptation and fine-tuning module supported by the model adaptation layer, using efficient model training means, the fast training of the industrial large model is realized.

[0142] In a possible design, the industrial large model platform 200 further includes: an interactive reasoning module 241 supported by the interaction layer 240;

[0143] The interactive reasoning module 241 is used to improve the interaction ability of the industrial large model through a variety of interactive reasoning optimization methods; among them, the variety of interactive reasoning optimization methods include: inference acceleration, industrial knowledge base retrieval enhanced generation, model orchestration, and plug-in mechanism.

[0144] Specifically, in the interactive reasoning module 241, the fine-tuned industrial large model can infer the conclusions expected by technicians or predict future events based on the input information and existing knowledge. The interactive reasoning module 241 uses a variety of interactive reasoning optimization methods such as inference acceleration, industrial knowledge base retrieval enhanced generation (Retrieval-Augmented Generation, RAG), model orchestration, and plug-in mechanism to enhance the capabilities of the industrial large model in specific industrial fields and optimize the performance of knowledge-intensive tasks.

[0145] In terms of inference acceleration, a variety of techniques such as model compression, pruning, and quantization can be adopted to reduce the computing and storage requirements of the industrial large model, thereby improving the inference speed. At the same time, high-performance hardware such as Graphic Processing Unit (GPU) and Tensor Processing Unit (TPU) is supported to speed up the inference process. These hardware acceleration technologies enable the industrial large model to efficiently process large amounts of data and meet industrial application scenarios with high real-time requirements, such as real-time monitoring of production lines and equipment fault detection. Inference acceleration can also reduce energy consumption and operating costs, making the industrial large model more available and popular in industrial sites with limited resources.

[0146] In terms of RAG, the interactive reasoning module 241 combines retrieval and generation technologies, which is particularly suitable for knowledge-intensive tasks in industrial scenarios. RAG first constructs an industrial knowledge base covering product design, process flow, and equipment maintenance, etc.; then, using efficient retrieval algorithms such as best match and dense retrieval, relevant information is found from the knowledge base. Based on the retrieved results, the industrial large model further generates more accurate and targeted answers. In the field of industrial manufacturing, RAG can be used to quickly provide technical support, generate maintenance manuals, and optimize production processes. For example, in equipment maintenance, technicians can quickly find fault solutions through RAG and automatically generate detailed repair steps, improving maintenance efficiency and accuracy.

[0147] Model orchestration involves model selection, task assignment, and result integration to achieve complex industrial tasks. Model selection chooses appropriate large models or small models according to specific task requirements; task assignment decomposes complex tasks into subtasks and assigns them to the most suitable models; result integration summarizes the outputs of sub-models through methods such as weighted average or logistic regression to generate the final solution. For example, in an intelligent factory, large models are responsible for production planning and scheduling, while small models handle operation monitoring and quality inspection. Through model orchestration, the overall production process is optimized, improving system efficiency.

[0148] In terms of the plugin mechanism, the interactive reasoning module 241 supports the installation and use of customized plugins to extend the functions of industrial large models. These plugins can be used to add new reasoning algorithms, support more data types or formats, provide visualization interfaces, etc. The interactive reasoning module 241 also provides a rich plugin resource library, such as knowledge base plugins and web page parsing plugins. Technicians can search, select, and install corresponding plugins as needed. In addition, technicians can integrate internal services into service orchestration through custom plugins. The plugin mechanism enhances the scalability and customization capabilities of the interactive reasoning module 241, making it show flexibility and adaptability in more application scenarios and meeting the wide-ranging needs of technicians.

[0149] The technical effect of this solution in this embodiment is that through the interactive reasoning module supported by the interactive layer, the processing speed and resource efficiency of industrial large models are improved, the application scenarios are expanded, and the adaptability and problem-solving ability of industrial large models for complex industrial tasks are enhanced.

[0150] In a possible design, the industrial large model platform 200 further includes: an evaluation module 251 supported by the application layer 250;

[0151] The evaluation module 251 is used to evaluate the functions of industrial large models through at least one of multiple evaluation methods; among them, the multiple evaluation methods include: automatic evaluation and manual evaluation.

[0152] Specifically, the evaluation module 251 includes a series of evaluation indicators and methods for performance evaluation of trained industrial large models. The evaluation module 251 provides two evaluation methods: automatic evaluation and manual evaluation, aiming to comprehensively measure the performance in multiple aspects such as semantic understanding, knowledge reasoning, professional ability, application ability, instruction following, robustness, bias, hallucination, and security; in addition, it also evaluates industrial-specific tasks, such as product design, process design, and quality control.

[0153] Automatic evaluation refers to the objective evaluation of the outputs generated by industrial large models through computer algorithms and statistical methods. This method is widely used in various tasks and can quickly and accurately evaluate the model performance. Automatic evaluation usually relies on predefined evaluation metrics and algorithms, mainly by calculating metrics such as accuracy, recall, and perplexity to quantify the performance of industrial large models in specific tasks. For industrial specific tasks, automatic evaluation also involves the application effects and practicality in the actual production environment. These tasks cover all aspects of industrial production, from product design to after-sales service, including but not limited to product design, process design, quality control, marketing management, after-sales service, planning and scheduling, production operations, warehousing and logistics, equipment management, safety and environmental protection, energy management, and supply chain, etc., in the manufacturing application capabilities. Such evaluation not only focuses on technical indicators but also pays attention to the performance and value of industrial large models in actual applications. Through evaluation, a comprehensive understanding of the applicability and benefits of industrial large models in the industrial environment can be obtained, thus providing a basis for the improvement and optimization of industrial large models.

[0154] Manual evaluation refers to the subjective evaluation of the outputs generated by industrial large models by evaluators. To ensure the consistency and accuracy of the evaluation, a set of evaluation metrics and scoring criteria are usually designed. For example, the fluency, accuracy, relevance, and innovativeness of the text generated by the model can all be used as evaluation dimensions. The evaluators score the performance of the industrial large model according to these criteria and provide detailed feedback and improvement suggestions.

[0155] The evaluation module 251 comprehensively measures the performance of industrial large models in aspects such as semantic understanding, knowledge reasoning, and industrial applications by combining automatic evaluation and manual evaluation. Automatic evaluation provides efficient and objective quantitative analysis, while manual evaluation supplements the detailed considerations of industrial large models in actual applications through the subjective feedback of evaluators. The two complement each other, providing comprehensive data support and practical references for the optimization and improvement of industrial large models, ensuring that industrial large models can better adapt to complex industrial application scenarios.

[0156] The technical effect of this solution in this embodiment is: through the evaluation module supported by the application layer, the performance of the trained industrial large model is evaluated.

[0157] This application embodiment also provides an industrial large model platform system, which includes: an industrial large model, and a cyber-physical system that interacts with the industrial large model;

[0158] The industrial large model is a model trained through the industrial large model platform provided in the above embodiments of this application.

[0159] The implementation principle and technical effect of the industrial large model platform system provided in this embodiment are similar to those of the industrial large model platform in the above embodiments, and will not be elaborated here in this embodiment.

[0160] So far, the technical solutions of the present application have been described in conjunction with the preferred embodiments shown in the accompanying drawings. However, it is easy for those skilled in the art to understand that the protection scope of the present application is obviously not limited to these specific embodiments. The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the various embodiments of the present application.

Claims

1. An industrial large model platform, characterized in that, The industrial large model platform is used to train industrial large models, and the industrial large model platform includes: an infrastructure layer, a base layer, a model adaptation layer, an interaction layer, and an application layer; The infrastructure layer is used to build the basic resources for training the industrial large model; the basic resources include: industrial data, computing resources, and industrial knowledge; the industrial data includes computer-aided design software files, industrial time-series data, machine instructions, industrial documents, and multi-modal data; The base layer is pre-equipped with toolchains for industrial multi-modal pre-training, industrial mechanism embedded fine-tuning, and industrial efficient reasoning respectively, as well as multiple industrial template large models; the base layer is used to perform industrial multi-modal pre-training on the pre-equipped industrial base large model through the pre-equipped multi-modal dataset to improve the processing ability of the industrial large model on multi-modal data; among them, the industrial base large model is one of the multiple industrial template large models and has the general problem-solving ability for industrial tasks; The model adaptation layer is used to perform industrial mechanism embedded fine-tuning on the pre-trained industrial base large model through the pre-equipped task-oriented large model and industry domain large model by means of industrial knowledge fusion and industrial mechanism embedding methods; among them, the task-oriented large model is used to perform multi-task instruction fine-tuning on the pre-trained industrial base large model to improve the processing ability of the industrial large model on specific tasks; the industry domain large model is used to perform industry domain knowledge embedding and adapter fine-tuning on the industrial base large model after multi-task instruction fine-tuning to improve the generalization ability of the industrial large model in unknown domain scenarios; among them, the industrial mechanism embedding method includes: representing the industrial mechanism as feature information recognizable by a neural network and embedding it into the network architecture of the industrial base large model so that the output result of the industrial large model conforms to industrial physical laws; The model adaptation layer is also used to perform scenario knowledge internalization and reinforcement self-training on the industrial base large model after industry domain knowledge embedding and adapter fine-tuning through the pre-equipped limited labeled data to improve the processing ability of the industrial large model in industrial sub-scenarios; The interaction layer is used to improve the interaction ability of the industrial large model through industrial efficient reasoning; among them, the industrial large model is used to interact with technicians and cyber-physical systems respectively; The application layer is used to present the functions of the industrial large model.

2. The industrial large model platform according to claim 1, wherein The task-oriented large model includes: an intelligent question answering model, a scenario recognition model, a process decision-making model, a terminal control model, a content generation model, and a scientific discovery model; Then the functions of the industrial large model include: intelligent question answering, scenario recognition, process decision-making, terminal control, content generation, and scientific discovery.

3. The industrial large model platform according to claim 2, wherein The functions of the industrial large model are used to support applications in research and development design, production manufacturing, test and testing, operation and management, and operation and maintenance services.

4. The industrial large model platform according to claim 1, wherein The industrial data is used to provide training data for the industrial large model; The computing resources are used to provide computing power support for the industrial large model; The industrial knowledge is used to provide in-depth logic for the industrial large model.

5. The industrial large model platform according to claim 4, characterized in that, The industrial large model platform further includes: a pre-training module supported by the base layer; The pre-training module is used to perform industrial multi-modal pre-training on the industrial base large model through at least one of a variety of pre-training methods; wherein, the variety of pre-training methods include: distributed training method, transfer learning training method, self-supervised learning method, and multi-task training method; The pre-training module is further used to set the industrial data called by the industrial base large model, as well as the model parameters of the industrial base large model.

6. The industrial large model platform according to claim 1, characterized in that, The industrial knowledge fusion method includes: After converting the industrial knowledge into embedded vector features by using knowledge graph embedding technology, integrating the embedded vector features into the model training process, and fine-tuning the pre-trained industrial base large model through instructions guided by the knowledge graph.

7. The industrial large model platform according to claim 1, wherein, The industrial large model platform further includes: an interactive reasoning module supported by the interaction layer; The interactive reasoning module is used to improve the interaction ability of the industrial large model through a variety of interactive reasoning optimization methods; wherein, the variety of interactive reasoning optimization methods include: reasoning acceleration, industrial knowledge base retrieval enhanced generation, model orchestration, and plug-in mechanism.

8. The industrial large model platform according to claim 1, wherein, The industrial large model platform further includes: an evaluation module supported by the application layer; The evaluation module is used to evaluate the functions of the industrial large model through at least one of a variety of evaluation methods; wherein, the variety of evaluation methods include: automatic evaluation and manual evaluation.

9. An industrial large model platform system, characterized in that, The industrial large model platform system includes: an industrial large model, and a cyber-physical system that interacts with the industrial large model; The industrial large model is a model obtained by training through the industrial large model platform according to any one of claims 1 to 8.

Images