Agricultural large model system based on multi-modal fusion technology

Abstract

The invention provides an agricultural large model system based on a multi-modal fusion technology. The agricultural large model system comprises a dialogue management module, a task scheduling engine, an agricultural knowledge base, a model and application library and a large voice model base, the dialogue management module comprises an intelligent interaction terminal, a voice converter, an image processor, a dialogue wrapper and a prompt template; the task scheduling engine comprises a task template, a task scheduler, a task executor and a task monitor; the data source of the agricultural knowledge base comprises an agricultural knowledge graph, an object database, a relational database and a text and graph database; the model and application library comprises a crop identification model, a pest identification model, a spectrum identification model and a plurality of business systems.

Description

Technical Field

[0001] The present application relates to the field of surveying and mapping technology, and in particular to a large agricultural model system based on multimodal fusion technology. Background Art

[0002] Traditional large language models are usually trained on large-scale general corpora. Through training on large-scale text data, they can understand and generate natural, complex and multi-domain texts and have extensive knowledge coverage capabilities.

[0003] Although traditional large language models have extensive knowledge coverage capabilities, they are often insufficient in depth and professionalism for specific industries such as agriculture. They may not be able to accurately understand and process industry-specific terminology, concepts, rules, and complex business logic, resulting in poor performance when handling specific industry tasks. Specific data training is still required to improve their professionalism and domain generalization capabilities.

[0004] Although traditional large language models have strong expressive power in the text field and can handle various language forms and styles, thus providing high-quality text generation and analysis, they are relatively weak in processing multimedia data such as speech and vision, and need to be combined with other related technologies to improve the performance and effect of the model.

[0005] Moreover, in real-world scenarios, some complex tasks usually consist of multiple subtasks, thus requiring the scheduling and collaboration of multiple models, which is also beyond the capabilities of language models. Summary of the invention

[0006] In order to solve at least one of the technical problems in the background technology, the embodiment of the present application provides an agricultural large model system based on multimodal fusion technology, including a dialogue management module, a task scheduling engine, an agricultural knowledge base, a model and application library, and a large voice model base;

[0007] The dialogue management module includes an intelligent interactive terminal, a voice converter, an image processor, a dialogue wrapper and a prompt template;

[0008] The task scheduling engine includes a task template, a task scheduler, a task executor, and a task monitor; the data sources of the agricultural knowledge base include an agricultural knowledge graph, an object database, a relational database, a text and a graph database;

[0009] The model and application library includes a crop identification model, a pest identification model, a spectrum identification model and several business systems.

[0010] In one embodiment of the present application, in the agricultural large model system based on multimodal fusion technology, the intelligent interactive terminal is used to receive audio and video signals and transmit them to the large language model base as information input, and can also serve as an output interface of the large language model base for voice output and video playback;

[0011] The intelligent interactive terminal includes an ESP-32 development board, a MAX9814 amplifier, a MAX98357 power amplifier, a SU-03T voice recognition chip and an ESP32-CAM camera.

[0012] In one embodiment of the present application, in the agricultural large model system based on multimodal fusion technology, the speech converter includes a speech-to-text part and a text-to-speech part, the speech-to-text part can convert the audio signal in the speech into text, and the text-to-speech part can convert the text into a speech signal.

[0013] In one of the embodiments of the present application, in the agricultural large model system based on multimodal fusion technology, the dialogue wrapper includes an input-output wrapper and a historical context wrapper;

[0014] The input-output wrapper is used to provide a unified conversation interface, coordinate language and visual semantics, merge audio and visual inputs into text sequences, merge different input requests, and correspond to multiple different output response formats after calling the large language model;

[0015] The history context wrapper is used to maintain a history context queue to support follow-up questions during the conversation.

[0016] In one of the embodiments of the present application, the agricultural large model system based on multimodal fusion technology, the task scheduler includes a task parsing module, a task matching module, a link control module and a distribution executor;

[0017] The task parsing module introduces contextual learning and injects demonstrations into prompts, so that the large language model base can better understand the intention and standards of task decomposition;

[0018] The task matching module directly matches the task template from the task number according to the parsed task, and obtains the dependency relationship and task parameters of the task;

[0019] The link control module is used to place dependent tasks that need to be completed by other tasks in a task list with a higher priority, and form a multi-branch linked list according to the topological structure;

[0020] The allocation executor is used to select the corresponding task executor according to different task types.

[0021] Beneficial effects of this application: This application uses a large language model as the foundation, combines artificial intelligence, big data, the Internet of Things and other technologies, vertically integrates and comprehensively merges multimodal data, and constructs a large language model for the agricultural industry. It effectively improves the intelligence level of the entire chain of macro-agriculture "production, supply, marketing, service, and management" and micro-agriculture "planting, breeding, farming, management, and harvesting", and promotes the deep integration of general artificial intelligence and the agricultural field and the application of scientific and technological innovation.

[0022] Through multimodal fusion technology, the company's data, application systems, hardware, and AI models are integrated to form an agricultural large language model expert system, which can be widely used in the agricultural field and achieve the following significant results.

[0023] 1. As an agricultural expert, we provide professional information consultation for farmers. The consultation covers a wide range of content, is professional, provides a wide range of services, and is easy to communicate. No professional knowledge or system is required, and natural language can be used for communication;

[0024] 2. Connecting the enterprise's big data, AI models, business systems, and intelligent terminal devices, it has made another leap forward on the basis of enterprise informatization, achieving information intelligence and making the enterprise's services truly alive. It not only improves the user experience and increases stickiness, but also improves the quality and efficiency of services for enterprises.

[0025] 3. With the evolution of the agricultural big language model, it becomes possible to build an agricultural brain. The agricultural brain can provide precise management and decision-making support for agricultural production through accurate data analysis and prediction, improve agricultural production efficiency, and reduce resource waste. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0027] Figure 1 It is a schematic diagram of a large agricultural model system based on multimodal fusion technology provided by this application;

[0028] Figure 2 It is the circuit schematic diagram of the intelligent interactive terminal;

[0029] Figure 3 It is a flowchart of data integration;

[0030] Figure 4 This is the agricultural large model deployment architecture diagram. DETAILED DESCRIPTION

[0031] Multimodal fusion technology is a comprehensive application based on technologies such as artificial intelligence, big data analysis and computer vision. It can fuse multi-source data from different sensors and models to improve the reliability, accuracy and completeness of the data.

[0032] In the agricultural field, multimodal fusion technology can be applied to the analysis and processing of agricultural big data, so as to better realize the development of new agricultural models such as precision agriculture and smart agriculture. With the continuous advancement of agricultural informatization and intelligence, agricultural big data and agricultural intelligent technology are becoming important directions for agricultural development, and the application of multimodal fusion technology will further promote the process of agricultural modernization.

[0033] In summary, when building a large agricultural model, different data, systems, technologies, resources, etc. can be integrated to achieve more efficient and optimized operations. Specifically, it includes the following aspects:

[0034] ① Use data lake technology to integrate information from multiple data sources and establish an agricultural industry knowledge base. The integrated data sources include relational data, object data, graph data, real-time data, documents, audio, pictures, videos, application systems, knowledge graphs, rule bases and other types of data;

[0035] ② Use deep learning algorithms to train multi-agent analysis models and expand the intelligent analysis boundaries of the agricultural big language model. Use machine learning, deep learning and other algorithms to build agricultural-related identification models such as pest and disease identification, crop identification, remote sensing identification of land objects, growth cycle identification, and yield prediction. Access them through the unified entrance of the agricultural big language model and output the model prediction results;

[0036] ③ Use sensor technology to access real-time data and enhance the expressiveness of the agricultural language model. Receive voice input or broadcast audio in real time through audio devices; access or display image and video data in real time through cameras; thus achieving efficient communication between humans and machines and barrier-free dialogue.

[0037] ④ Based on the decomposition and scheduling engine of task templates, multi-branch parallel reasoning is realized. By tracking the contextual state of user requests, a complex user request is planned as a multi-step, multi-round concurrent execution task, thereby solving various forms of tasks such as language, image, audio and video, including detection, generation, classification and question answering.

[0038] See also Figure 1 As shown, the embodiment of the present application provides an agricultural large model system based on multimodal fusion technology, including a dialogue management module, a task scheduling engine, an agricultural knowledge base, a model and application library, and a large voice model base, wherein:

[0039] The dialogue management module includes an intelligent interactive terminal, a voice converter, an image processor, a dialogue wrapper and a prompt template.

[0040] The intelligent interactive terminal is used to receive audio and video signals and pass them to the large language model base as information input. It can also serve as the output interface of the large language model base for voice output and video playback. The intelligent interactive terminal includes the ESP-32 development board, MAX9814 amplifier, MAX98357 power amplifier, SU-03T voice recognition chip and ESP32-CAM camera.

[0041] The speech converter includes a speech-to-text part and a text-to-speech part. The speech-to-text part can convert the audio signal in the speech into text, and the text-to-speech part can convert the text into a speech signal.

[0042] The conversation wrapper includes an input-output wrapper and a historical context wrapper. The input-output wrapper is used to provide a unified conversation interface, coordinate language and visual semantics, merge audio and visual input into a text sequence, merge different input requests, and call a large language model to correspond to a variety of different output response formats. The historical context wrapper is used to maintain a historical context queue to support follow-up questions during the conversation process.

[0043] The task scheduling engine includes task templates, task schedulers, task executors, and task monitors; the data sources of the agricultural knowledge base include agricultural knowledge graphs, object databases, relational databases, text and graph databases.

[0044] The task scheduler includes a task parsing module, a task matching module, a link control module and an allocation executor.

[0045] The task parsing module introduces contextual learning and injects demonstrations into prompts, so that the large language model can better understand the intention and standards of task decomposition. The task matching module directly matches the task number to the task template based on the parsed task, and obtains the dependency and task parameters of the task. The link control module is used to place dependent tasks that need to be completed by other tasks in a higher priority task list, and form a multi-branch linked list according to the topological structure. The allocation executor is used to select the corresponding task executor according to different task types.

[0046] The model and application library includes crop identification models, pest and disease identification models, spectral identification models and several business systems.

[0047] 1. Technical architecture composition:

[0048] Part 1: Dialogue management module. Processes the input and output of voice, text, pictures and videos. Including intelligent interactive terminal, voice converter, image processor, dialogue wrapper and prompt template.

[0049] Part 2: Task scheduling engine. Through task decomposition and scheduling, the large language model, intelligent analysis model, and external application system are called in parallel linked lists, and finally a consistent reasoning conclusion is given. It includes task templates, task schedulers, task executors, and task monitors.

[0050] Part 3: Agricultural knowledge base. Through the lake-warehouse integration technology, structured and unstructured data are integrated to fine-tune the large language model base. The integrated data sources include agricultural knowledge graphs, object databases, relational databases, text, graph databases, etc. The fine-tuning process consists of four steps: multi-data source integration, sample data preparation, vectorized storage, and external knowledge base.

[0051] Part 4: Model and Application Library. Through machine learning and deep learning algorithms, data preparation, feature extraction, and model training are performed on the relevant thematic data in the agricultural knowledge base to build a number of analytical models in the agricultural field. Including crop identification models, pest and disease identification models, spectral identification models, and several business systems (such as smart agriculture systems, rural property rights trading systems, etc.).

[0052] Part 5: Large language model base. Based on the principle of not reinventing the wheel, we selected an open source large language model as the base, then fine-tuned it and integrated other analysis models to build the upper-level application.

[0053] 2. Key technical components:

[0054] By utilizing the technical framework and components of the open source ecosystem, combined with localized data and mature business solutions in the agricultural field, we can quickly build a large agricultural model and output industry solutions to better realize intelligent agricultural production, improve agricultural production efficiency and quality, and contribute to agricultural development.

[0055] 1. Intelligent interactive terminal: The terminal device is mainly used to receive audio and video signals and transmit them to the large voice model base as information input. It can also serve as the output interface of the large voice model base for voice output and video playback. The terminal is mainly composed of ESP-32 development board, MAX9814 amplifier, MAX98357 power amplifier, SU-03T voice recognition chip, ESP32-CAM camera and other modules. By using the ESP-32 development board, users can interact with the large voice model base through voice and camera. The circuit principle is as follows Figure 2 shown.

[0056] The core processing flow of the system is: the device monitors the wake-up word through SU-03T. After the device wakes up, it receives the voice signal sampling value from MAX9814, processes the data, obtains the Base64 encoded string of PCM audio, and then sends the PCM audio to the large voice model base server through the built-in WiFi module of ESP-32 through http request, obtains the audio after the AI ​​reply from the response, and then plays the audio of the AI ​​reply through MAX98357. Video collection is also through the voice command of the audio after monitoring, collects the face of the current interactive person to the background, calls the large voice model base recognition, and conducts human-computer interaction.

[0057] 2. Voice converter: It includes two parts. One is voice-to-text, which converts the audio signal in the voice into text through voice recognition technology. The other is text-to-speech, which converts text into voice signals. Both can be implemented using open source voice components. You can also use deep learning algorithms to build a local synthetic voice phoneme library through steps such as voice signal preprocessing, feature extraction, model training and voice recognition. Voice and text can be based on business needs and support multiple international languages ​​such as Chinese, English, and Japanese.

[0058] 3. Dialogue wrapper: It consists of two parts, input and output wrapper and history context wrapper.

[0059] The input and output packaging provides a unified conversation interface, coordinates language and visual semantics, merges audio and visual input into text sequences, merges different input requests, and calls the large language model to correspond to a variety of different output response formats. Input requests support text, voice, pictures, videos, text files, wps files, pdf, word files and other formats. Output responses support text, voice, pictures, text files, pdf, word files, programming languages ​​and other formats.

[0060] The historical context wrapper is mainly used to support follow-up questions in the conversation and improve the user experience. The wrapper maintains a historical context queue, following the first-in-first-out principle. Before each round of conversation, historical information will be injected into the current conversation. After the conversation ends, the context of the current conversation will be automatically inserted into the historical context queue.

[0061] 4. Prompt template: Prompt templates are used to generate text, voice questions and answers, perform actions, etc. The system will maintain a prompt template library. Before each conversation, the program will automatically match a template, then inject the user's input into the template, format the prompt text, and finally call the language model. If no template is matched, the default general template will be used.

[0062] The information of the prompt template consists of the following parts:

[0063] Number: Starts with "CLM_" and is in English and numbers, uniquely identifying the template;

[0064] Parameters: used to control the content generated by the template, such as specifying:

[0065] 1) Role: data analyst, security officer, hardware engineer;

[0066] 2) Emotional color: happy, sad, positive;

[0067] 3) Reply step by step: imitate the case field to generate the problem-solving process;

[0068] 4) Generate poems: five-character quatrain, seven-character quatrain, Song Dynasty poems, and modern poems;

[0069] 5) Code generation: such as Java, Python, JavaScript, and C++, which separate programming languages ​​from natural languages ​​and generate code logic to solve reasoning problems;

[0070] Case: The case of the reply content corresponding to the template, for reference by the large language model. The case corresponds to the parameter one by one.

[0071] Description: Used to explain the function of the template.

[0072] 5. Task template: define some common templates to facilitate the understanding and execution of large models. Template information includes:

[0073] Task number: All numbers are in English and numbers to uniquely identify the task;

[0074] Task type: covers different types such as text, visual, audio, etc., and can be implemented but not limited to the following tasks:

[0075] 1) Text Question Answering (QA): The basic function of the large language model base;

[0076] 2) Speech-to-text (STT): Cooperate with intelligent interactive terminals to realize speech-to-text conversion;

[0077] 3) Text-to-speech (TTS): text-to-speech, played through intelligent interactive terminals;

[0078] 4) Play audio and video (play): play the audio and video data in the background audio and video library;

[0079] 5) Crop identification (crops): Call AI models to identify various crops;

[0080] 6) Pest identification: Use AI models to identify crop diseases and pests;

[0081] 7) Environmental Identification (env): Call the AI ​​model to read image information, interpret the problems existing in a specific agricultural environment, and the measures that can be taken to solve them;

[0082] 8) Face recognition: interact with agricultural experts and introduce their resumes, skills, etc.

[0083] 9) System call (callapp): connect to the company's existing external business system to provide more professional data;

[0084] Dependency: Specify the predecessor dependent tasks of each task, which must form a directed acyclic graph. A task can have multiple predecessor dependencies;

[0085] Task parameters: Different task types have different parameters. For example, for voice question answering, you need to provide the address of the object warehouse, the type of voice, and the name of the third-party AI model; for face recognition, you need to provide the address of the object warehouse, the recognition identifier, the name of the third-party AI model, etc.

[0086] 6. Task scheduler: includes four steps: task analysis, task matching, link control, and executor allocation.

[0087] 1) Task parsing: Contextual learning is introduced, which allows the large model to better understand the intention and criteria of task decomposition by injecting several demonstrations into the prompt. Each demonstration is a set of inputs and outputs about task decomposition, the user's request and the expected task to be parsed out. The demonstration template is as follows, where the prefix is ​​the task template number:

[0088]

[0089] 2) Task matching: Based on the tasks parsed in the previous step, the task number can be directly matched to the task template, and the dependencies and task parameters of the task can be obtained.

[0090] 3) Link control: When a task depends on the completion of other tasks, the dependent task will be placed in the task list with a higher priority. According to this topological structure, a multi-branch linked list will eventually be formed.

[0091] 4) Assign executors: Select the corresponding executors according to different task types.

[0092] 1) Task executor: responsible for actually executing the tasks in the linked list, including:

[0093] 2) After all the predecessor tasks of the same task in the control link are completed at the same time, the next task will be executed;

[0094] 3) Control the start, execution, stop, concurrency and retry times of each task;

[0095] 4) Integrate the execution results of all predecessor tasks of the same task as the input of the next task;

[0096] A typical "a->c, c->c" task list execution code is as follows:

[0097]

[0098] 2) Task matching: Based on the tasks parsed in the previous step, the task number can be directly matched to the task template, and the dependencies and task parameters of the task can be obtained.

[0099] 3) Link control: When a task depends on the completion of other tasks, the dependent task will be placed in the task list with a higher priority. According to this topological structure, a multi-branch linked list will eventually be formed.

[0100] 4) Assign executors: Select the corresponding executors according to different task types.

[0101] 7. Task executor: responsible for actually executing the tasks in the linked list, including:

[0102] 1) After all the predecessor tasks of the same task in the control link are completed at the same time, the next task will be executed;

[0103] 2) Control the start, execution, stop, concurrency and retry times of each task;

[0104] 3) Integrate the execution results of all predecessor tasks of the same task as the input of the next task;

[0105] A typical "a->c, c->c" task list execution code is as follows:

[0106]

[0107] 8. Agricultural knowledge base: To build a large agricultural model, industry data is crucial. The company has many years of business experience in agriculture, and has a lot of vector data, unstructured data, and structured data. To fully integrate this data and fine-tune the large language model base, please refer to the process. Figure 3 .

[0108] The above functions include the following parts:

[0109] 1) Integration and cleaning of multiple data sources: Use data lake warehouse integration technology to centrally store these different sources and various types of data. Perform data cleaning, data integration, data conversion, and data integration to gain in-depth understanding and analysis of the data and support business decision-making;

[0110] 2) Generate sample data: Sample data should be extracted according to the characteristics of the industry model and the following categories. The sample files are divided into training sets and test sets in proportion. The files can be JSON, CSV, PDF, TXT, XML, DOCX, PNG, JPG, etc. The size of a single file is limited to about 100MB, and the images need to be compressed and not exceed 10M.

[0111] Property transaction

[0112] Business Entity

[0113] Smart Management

[0115] Financial Services

[0116] Agricultural machinery management

[0117] Agricultural technical services

[0118] Tourist accommodation

[0119] The code is as follows:

[0120]

[0121] 3) Vectorized storage: Convert text data into vectors. You can choose a Chinese-friendly vector model library such as gte-large-zh, bge-large-zh-v1.5, m3e-base, tao8k, etc. Then store it in a vector database such as FAISS for later fast retrieval or text similarity comparison. The following code fine-tunes the embedding model "bge-large-zh-v1.5":

[0122]

[0123] 4) External knowledge base: Through the previous steps, a local agricultural knowledge base is created. When a user asks a question, it is sent to the large language model, and the vector representation of the question is searched for similarity in the agricultural knowledge base. After obtaining relevant information, it is fed back to the large language model. The large language model provides answers or takes actions based on the question and related information from the agricultural knowledge base.

[0124] 9. Model and application library: With the large language model as the base and the dialogue management module as the entrance, multiple applications and intelligent models are integrated through the task scheduling engine to turn the enterprise's technological accumulation into part of the brain of smart agriculture. Intelligent models can be audio and video recognition models trained by machine learning, deep learning and other frameworks, and applications can be various business systems operated by the enterprise.

[0125] 10. Large language model base: You can choose a foreign or domestic open source large language model as the base, such as LLaMA 2, BERT, Vicuna 13-B, chatGLM3-6B, etc. The large language model base provides powerful basic functions and provides reliable basic support for all business services.

[0126] The large agricultural model is deployed using a distributed architecture, as shown in Figure 4, and can be divided into the following parts:

[0127] 1. Data domain: It is an integrated platform of the enterprise’s data lake and data warehouse, which aggregates basic agricultural data. It is the source of AI models and large agricultural model data sets, and is also the database accessed by business systems.

[0128] 2. Service governance: including two nodes: the registration center and the gateway. The registration center is used to register the agricultural model, AI model, and the interface provided by the business system. The gateway node is used to provide a unified service address, monitoring, and security to the outside world.

[0129] 3. Business domain: including business-side applications such as agricultural big models, AI models, and business systems. These business domain applications are registered and configured through the registration center, and the externally exposed interfaces are uniformly forwarded and securely controlled by the gateway. The agricultural big model also accesses the AI ​​model and business system through the gateway.

[0130] 4. Front-end: All applications in the business domain can independently provide services to the outside world through the gateway. The clients supported by the agricultural big model include PC, mobile (app, mini-program), and intelligent interactive terminals based on multi-sensor packaging.

[0131] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. An agricultural large model system based on multimodal fusion technology, characterized in that: It includes a dialogue management module, a task scheduling engine, an agricultural knowledge base, a model and application library, and a large speech model base; The dialogue management module includes an intelligent interactive terminal, a voice converter, an image processor, a dialogue wrapper and a prompt template; The task scheduling engine includes a task template, a task scheduler, a task executor, and a task monitor; the data sources of the agricultural knowledge base include an agricultural knowledge graph, an object database, a relational database, a text and a graph database; The model and application library includes a crop identification model, a pest identification model, a spectrum identification model and several business systems.

2. The agricultural large model system based on multimodal fusion technology according to claim 1 is characterized in that: The intelligent interactive terminal is used to receive audio and video signals and transmit them to the large language model base as information input, and can also serve as an output interface of the large language model base to perform voice output and video playback; The intelligent interactive terminal includes an ESP-32 development board, a MAX9814 amplifier, a MAX98357 power amplifier, a SU-03T voice recognition chip and an ESP32-CAM camera.

3. The agricultural large model system based on multimodal fusion technology according to claim 1 is characterized in that: The speech converter includes a speech-to-text part and a text-to-speech part. The speech-to-text part can convert an audio signal in the speech into text, and the text-to-speech part can convert the text into a speech signal.

4. The agricultural large model system based on multimodal fusion technology according to claim 1 is characterized in that: The dialog wrapper includes an input-output wrapper and a history context wrapper; The input-output wrapper is used to provide a unified conversation interface, coordinate language and visual semantics, merge audio and visual input into text sequences, merge different input requests, and correspond to multiple different output response formats after calling the large language model; The history context wrapper is used to maintain a history context queue to support follow-up questions during the conversation.

5. The agricultural large model system based on multimodal fusion technology according to claim 1 is characterized in that: The task scheduler includes a task parsing module, a task matching module, a link control module and an allocation executor; The task parsing module introduces contextual learning and injects demonstrations into prompts, so that the large language model base can better understand the intention and standards of task decomposition; The task matching module directly matches the task template from the task number according to the parsed task, and obtains the dependency relationship and task parameters of the task; The link control module is used to place dependent tasks that need to be completed by other tasks in a task list with a higher priority, and form a multi-branch linked list according to the topological structure; The allocation executor is used to select the corresponding task executor according to different task types.

Images