Industrial sewing instruction multi-level analysis method and system fusing real-time state of equipment

Abstract

The invention discloses an industrial sewing instruction multi-level analysis method fusing the real-time state of equipment, and the method comprises the steps: firstly decomposing a complex instruction into an atomic operation intention sequence through a special instruction decomposition model; secondly, performing parameter standardization by combining the domain entity analysis model with a knowledge base; thirdly, inferring the real-time physical state of the model fusion equipment according to the context numerical value; and finally, generating a part of completely structured machine executable data through an intention synthesis step. By means of the cascade architecture exceeding the RAG, natural language instructions of workers can be converted into accurate and unambiguous machine intentions in a high-fidelity mode, and core technical support is provided for intelligent and reliable industrial sewing machine control.

Description

Technical Field

[0001] This invention belongs to the field of Natural Language Processing (NLP) technology in artificial intelligence, and specifically relates to a multi-level parsing method and system for industrial sewing instructions that integrates real-time equipment status. More specifically, this invention discloses a parsing method and corresponding system specifically for industrial sewing machine control scenarios, capable of processing natural language containing domain terminology, fuzzy descriptions, implicit intentions, and complex instructions, and parsing it into precise, unambiguous, structured operational data. Background Technology

[0002] Introducing natural language interaction into industrial sewing machine control is key to improving human-machine collaboration efficiency. However, natural language commands in industrial scenarios have unique complexities that existing general-purpose natural language understanding (NLU) technologies struggle to handle effectively.

[0003] Traditional NLU technologies, such as those based on "intent recognition + slot filling," have overly rigid architectures and cannot handle implicit intents, compound instructions, fuzzy and relative instructions, as well as a large number of domain-specific terms commonly encountered in industrial environments.

[0004] The emergence of Large Language Models (LLMs) has brought new possibilities. To overcome the problems of general-purpose LLMs lacking domain knowledge and exhibiting "illusions," a mainstream technical approach is Retrieval-Augmented Generation (RAG). This method assists LLMs in generating answers by retrieving relevant information from external knowledge bases, achieving significant success in general question-answering and dialogue scenarios. However, when attempting to directly apply the classic RAG process to industrial control domains that require the generation of precise, structured machine instructions, its inherent architectural flaws become apparent, revealing the following core pain points that are difficult to overcome:

[0005] 1. Limited Contextual Understanding: Classic RAG or dialogue models rely solely on dialogue history when processing context, while industrial control scenarios demand real-time physical state of equipment. The lack of mechanisms for perceiving physical state in existing technologies leads to fundamental limitations in processing industrial commands, primarily manifested in:

[0006] It cannot quantify relative instructions (such as "faster");

[0007] Unable to handle state-dependent operations (such as "wait for the needle to stop before lifting the presser foot");

[0008] The safety and suitability of the instructions cannot be assessed.

[0009] Therefore, how to effectively integrate physical states into the intent resolution process is a key technical problem that needs to be solved in this field.

[0010] 2. Inability to handle complex and sequential tasks. The classic RAG workflow is typically a one-step "retrieve-merge-generate" process, adept at answering single "what" or "how" questions. However, it lacks an inherent mechanism to decompose a complex instruction containing multiple sequential steps (e.g., "use pattern #5, slow down at the corner, and cut the line at the end"). RAG cannot reliably break this instruction down into three independent, ordered subtasks and parse them separately.

[0011] 3. Mismatched Architectural Goals: Built for "Dialogue" Rather Than "Control". The core goal of classic RAGs is to generate natural, fluent, and information-rich text responses, with its entire optimization focused on improving the conversational experience. However, the ultimate goal of industrial control is to generate fully deterministic, unambiguous, and strictly formatted structured data (such as JSON) to drive machine execution. The open-ended text generation characteristic of RAGs is precisely the source of "uncertainty" that industrial control must strive to avoid. Directly applying RAGs is futile.

[0012] Therefore, the market urgently needs a new intent parsing architecture that transcends the classic RAG paradigm and is specifically designed for industrial control scenarios. This architecture must combine the language capabilities of LLM with the deterministic requirements of industrial control to achieve high-fidelity conversion between fuzzy natural language and precise, structured machine instructions. Summary of the Invention

[0013] The purpose of this invention is to address the architectural deficiencies of existing technologies, particularly classical Retrieval Enhanced Generation (RAG) technology, in industrial control applications. It provides a multi-level parsing method and system for industrial sewing instructions that integrates real-time equipment status, resolving the fundamental contradiction between RAG and other technologies' inability to effectively handle composite instructions, their inability to perceive the physical state of equipment to quantify relative instructions, and the gap between their open generation goals and the highly deterministic output required for industrial control.

[0014] To achieve the above objectives, this invention abandons the traditional "one-step" end-to-end generation paradigm of RAG and innovatively proposes a cascaded multi-level parsing architecture. This architecture decomposes the complex intent parsing task into a series of goal-oriented, functionally cohesive steps, and equips key steps with specialized large language models, forming a collaborative model system. This process includes the following innovative steps, aiming to address the pain points of classic RAG one by one:

[0015] Instruction decomposition steps (addressing the pain point of RAG's inability to handle complex tasks):

[0016] To handle complex instructions, this invention incorporates a dedicated instruction decomposition step. Utilizing a "dedicated instruction decomposition model" fine-tuned for the instruction decomposition task, a complex instruction that may contain multiple operations is first decomposed into one or more independent, ordered sequences of atomic operation intentions. For example, "use pattern number 5, automatically cut the line at the end" is decomposed into [set pattern, set end action]. This step is a prerequisite for subsequent accurate parsing and a crucial element missing from the classic RAG architecture.

[0017] Entity parsing steps (adapted and optimized based on RAG):

[0018] For each atomic operation intent decomposed in the previous step, a "domain entity resolution model" based on the RAG architecture is used for processing. However, unlike classic RAG, the knowledge base in this invention (such as the process terminology knowledge base and the process template knowledge base) is not used for open-ended answers, but rather serves as a "translation dictionary" and "configuration template," precisely mapping domain terms and process patterns to standardized parameter names and values. This step adapts RAG technology from a "question-answering tool" to a "precise translator."

[0019] Value inference steps (addressing the pain point of RAG's inability to perceive physical state):

[0020] To handle relative instructions, this invention incorporates a dedicated value inference step. The core of this step lies in integrating the device's real-time state information into the reasoning process of the large language model. The system first acquires the target device's real-time physical state data (such as current speed) through an interface and combines it with the atomic operation intent (such as "speed a little faster") to form an input that includes both linguistic and physical state information. This combined input is then provided to a "contextual numerical inference model" specializing in numerical reasoning. Through this design, this invention provides a reliable technical path for quantifying relative instructions.

[0021] Intent synthesis and verification steps (ensuring deterministic output and overcoming the drawbacks of RAG open generation):

[0022] All atomic operation intentions and their determined parameters processed in the preceding steps are combined into a pre-formatted, fully structured data (such as JSON), rather than natural language text. This step ensures the absolute determinism and machine readability of the final output. Furthermore, logical consistency self-checking further guarantees the reliability of the instructions, fundamentally avoiding the "illusion" and uncertainty risks associated with classic RAG open generation.

[0023] This invention also provides a multi-level intent parsing system for industrial sewing processes. The system includes: an instruction decomposition module (with a built-in dedicated instruction decomposition model), an entity parsing module (with a built-in domain entity parsing model and coupled with a domain knowledge base retrieval interface), a value inference module (with a built-in context value inference model and coupled with a device state center communication interface), and an intent synthesis module. Each module executes the corresponding steps in the above method in sequence.

[0024] Compared with the prior art, the present invention, through the above-mentioned original technical solution, brings the following beneficial effects:

[0025] Architectural innovation that transcends the classic RAG paradigm: The "decomposition-analysis-inference-synthesis" four-step cascaded architecture proposed in this invention fundamentally solves the three core pain points of classic RAG in processing industrial control instructions (composite tasks, physical state awareness, and output uncertainty), providing a new and more suitable implementation path for highly reliable instruction parsing.

[0026] It achieves a deep understanding of complex industrial-grade instructions: The invention’s unique “decomposition-analysis-inference-synthesis” multi-level parsing process breaks down complex semantic understanding tasks into a series of simpler and more controllable sub-tasks, thereby enabling precise processing of compound instructions, implicit intentions and ambiguous instructions that traditional NLU cannot handle, with understanding capabilities and accuracy far exceeding existing technologies.

[0027] By introducing state awareness, the present invention enables precise quantification of relative instructions: By introducing the current state of the device as context in the parsing process, the present invention enables the model to understand relative instructions such as "faster" and "loosen up" and convert them into specific, executable absolute values, making human-computer interaction more natural and smooth.

[0028] Outputting highly standardized and reliable structured data: Unlike RAG's open text generation, the final output of this invention is a completely unambiguous structured intent data that has undergone multi-level processing and logical verification. It provides extremely clean and reliable input for the subsequent "deterministic verification and instruction generation" stages, and is the key cornerstone for achieving stable operation of the entire intelligent control system. Attached Figure Description

[0029] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below.

[0030] Figure 1 This is a schematic diagram of the functional module structure of an industrial sewing instruction multi-level parsing system that integrates the real-time status of the equipment according to an embodiment of the present invention.

[0031] Figure 2This is a flowchart of a multi-level parsing method for industrial sewing instructions based on the real-time status of a fusion device, according to an embodiment of the present invention.

[0032] Figure 3 This is a schematic diagram illustrating the core data flow of the present invention, wherein:

[0033] Figure 3 A shows a text example of a complex natural language instruction entered by a user.

[0034] Figure 3 C illustrates the relevant knowledge fragments retrieved by the system from the domain knowledge base during step S202 (entity parsing based on retrieval enhancement).

[0035] Figure 3 D shows the current device state data acquired by the system during step S203 (context association and state-aware value inference).

[0036] Figure 3 E demonstrates a JSON-formatted example of the final synthesized, complete structured intent data after all processing steps.

[0037] Explanation of markings in the attached diagram:

[0038] 100-Multi-level Intent Resolution System

[0039] 110-Instruction Decomposition Module

[0040] 120-Entity Parsing Module

[0041] 121-Technology Terminology Knowledge Base

[0042] 122-Process Template Knowledge Base

[0043] 130-Value Inference Module

[0044] 131-Equipment Status Center

[0045] 140 - Intent Synthesis Module

[0046] Flowchart of the 200-level multi-level intent parsing method

[0047] S201 - Instruction Decomposition Steps

[0048] S202 - Entity Resolution Steps

[0049] S203-Value Inference Steps

[0050] S204 - Intent Synthesis and Verification Steps Detailed Implementation

[0051] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. To make the objectives, technical solutions, and advantages of the present invention clearer, a typical industrial sewing scenario will be used as an example to illustrate in detail how the multi-level intent parsing method of the present invention overcomes the limitations of classic Retrieval Augmentation (RAG) technology.

[0052] Example 1

[0053] A multi-level parsing method for industrial sewing instructions that integrates real-time equipment status, such as... Figure 2 and Figure 3 As shown, it includes the following steps:

[0054] S201. Receive natural language instruction text and decompose the instruction text into one or more atomic operation intentions using a pre-trained first large language model.

[0055] Specifically, the training of this first major language model uses a sequence of compound instructions and corresponding atomic operation intentions.

[0056] In this embodiment, the first pre-trained language model training method is as follows:

[0057] 1. Training data construction: Collect or manually label a large number of instruction data pairs (X,Y) in the field of industrial sewing, where X is the original natural language instruction text that may contain multiple operations (e.g., "Use beaded stitch, speed up a bit"), and Y is the corresponding JSON format string of atomic operation intention sequence that has been manually segmented and standardized (e.g., ["Set stitch type", "Adjust sewing speed"]).

[0058] 2. Fine-tuning process: Select a suitable pre-trained large language model as the base (e.g., Tongyi Qianwen, Llama, ChatGLM, etc.). Use the dataset constructed above to perform supervised fine-tuning (SFT) on this base model. The goal of fine-tuning is to enable the model to learn the mapping relationship from input X to output Y.

[0059] 3. Loss Function and Optimization: During training, the cross-entropy loss function is typically used to calculate the difference between the model's predicted token sequence and the true label Y. Optimizers such as AdamW are used to minimize the loss function, and the model parameters are updated through backpropagation.

[0060] 4. Training Completion Criteria: Training is considered complete when the model's performance metrics (such as exact match rate, BLEU score, etc.) on an independent validation set reach a preset threshold or no longer show significant improvement, or when training reaches a preset number of iterations (Epochs). The model at this point is considered a "dedicated instruction decomposition model".

[0061] In this embodiment, when sewing the cuffs of a batch of shirts, a new process instruction needs to be issued to the currently used industrial sewing machine. The industrial sewing machine used in this embodiment has the device ID: SN-A087. The instruction is given via the human-machine interface terminal:

[0062] "Set this machine to 'shirt cuff' mode, use beaded stitch, speed it up a bit, and remember to run it back and forth at the end."

[0063] This sentence (e.g.) Figure 3 As shown in A, the natural language instruction text is input into the multi-level intent parsing system 100 of the present invention.

[0064] The instruction text is fed into the instruction decomposition module 110. This module incorporates a pre-trained primary language model, a "dedicated instruction decomposition model" fine-tuned for the instruction decomposition task. The sole responsibility of this model is to identify task boundaries within the instruction, rather than understanding the specific content. Due to its specialized training, it can efficiently and accurately decompose the aforementioned complex instruction into an ordered sequence containing four atomic operation intentions:

[0065] ["Set process mode", "Set stitch type", "Adjust sewing speed", "Set end action"]

[0066] The value of this step lies in the fact that by preprocessing with a dedicated model, a "chaotic" and holistic task is broken down into a series of clear and independently processable subtasks, providing high-quality input for subsequent more refined analysis.

[0067] S202. For each atomic operation intention, the entity parameters in the atomic operation intention are identified and standardized using a pre-trained second language model and in combination with at least one preset domain knowledge base.

[0068] Specifically, the domain knowledge base includes: a process terminology knowledge base that stores the mapping relationship between domain terms and standard parameter names, and a process template knowledge base that stores the mapping relationship between a specific process and a set of standard parameter combinations.

[0069] The training data for the second language model includes atomic operation intentions, knowledge retrieved from the domain knowledge base, and standardized entity parameters with labels. It is trained to fuse the atomic operation intentions and the knowledge retrieved from the domain knowledge base to generate standardized entity parameters.

[0070] In this embodiment, the second major language model training method for pre-training is:

[0071] 1. Training Data Construction: Construct triplet data in the format (P, C, Z). Where P is an atomic operation intent (e.g., "Set stitch type: use beaded stitch"); C is a relevant knowledge fragment retrieved from the knowledge base by simulating the RAG process (e.g., {"Term":"beaded stitch","Standard name":"picot_stitch"}); Z is a manually annotated, correct standardized entity parameter JSON string (e.g., {"stitch_protocol_id":407}).

[0072] 2. Input Formatting: During training, P and C are concatenated into the model's input Prompt using a fixed template, for example: "Task: Based on the provided knowledge, parse the following atomic intents.\nKnowledge: {C}\nAtomic Intents: {P}\nParsing Result: ".

[0073] 3. Fine-tuning and Optimization: Similarly, a suitable base model is selected for supervised fine-tuning. The goal of the model is to learn to generate an output that is completely consistent with Z, given P and C. The loss function and optimizer selection are similar to those of the first model.

[0074] 4. Training Completion Marker: Training is complete when the exact match rate (ExactMatch) between the JSON generated by the model on the validation set and the tags reaches a preset high standard (e.g., above 99%). At this point, the model is a "domain entity parsing model," possessing the ability to accurately "translate" using external knowledge.

[0075] In this embodiment, the entity parsing module 120 processes the atomic intent sequence generated in the previous step one by one. The core of this module is the "domain entity parsing model," which is a pre-trained second-largest language model. It works in a retrieval-enhanced generation (RAG) manner and is adept at combining external knowledge with the current task for accurate translation.

[0076] For the atomic intent "Set process mode", the module recognized the keyword "shirt cuff". This triggered a search of the internal process template knowledge base 122 (e.g., Figure 3(As shown in C). The knowledge base returns the relevant template: {"Process Name":"Shirt Cuff","Default Parameters":{"Needle Count":2.5,"Thread Tension":4.2,"Presser Foot Pressure":"Medium"}}. The domain entity parsing model integrates atomic intents with this knowledge, resolving multiple standard parameters.

[0077] For the atomic intent "Setting Stitch Type", the module recognizes the keyword "picot stitch". It triggers a search of the process terminology knowledge base 121 and finds the mapping: {"Term":"picot stitch","Standard Name":"picot_stitch","Protocol ID":407}. The model uses this mapping to accurately standardize "picot stitch".

[0078] For the atomic intent "set end action", the module recognizes the keyword "bartack". Similarly, it queries the process terminology knowledge base 121, finds the mapping: {"terminology":"bartack","standard name":"bartack","protocol value":1}, and completes the standardization.

[0079] Through this "retrieval + model fusion" mechanism, multiple vague and colloquial expressions are transformed into precise and standardized parameter names and values, greatly improving the reliability of domain terminology processing.

[0080] Step 3: Identify the relative instructions that depend on the current state of the device, obtain the current state data of the device, and use the pre-trained third language model to infer the relative instructions into specific parameter target values.

[0081] Furthermore, the current state data of the device is provided as contextual information to the third language model to assist it in making inferences.

[0082] The training data for the third language model includes relative instructions, current device state data, and labeled target parameter values. It is trained to infer the specific target parameter values ​​based on the relative instructions and the current device state data.

[0083] In this embodiment, the third major language model training method for pre-training is:

[0084] 1. Training Data Construction: Construct triplet data in the format (R, S, T). Where R is the atomic operation intention containing a relative description (e.g., "Adjust sewing speed: speed up a bit"); S is the current real-time status data of the device (e.g., {"current_speed_rpm":2000}); T is the specific target value of the parameter set manually or calculated according to rules (e.g., {"target_speed_rpm":2300}).

[0085] 2. Numerical Reasoning and Context Fusion: The training focus is on enabling the model to learn the non-linear mapping relationship between qualitative terms in R (such as "faster" and "loosen up") and the current numerical values ​​in S. The input Prompt can be designed as: "Task: Based on the current state of the device, infer the target value relative to the instruction.\nCurrent state: {S}\nRelative instruction: {R}\nInference result: ".

[0086] 3. Fine-tuning and Loss Functions: For tasks requiring precise numerical output, in addition to the standard cross-entropy loss, regression loss functions such as Mean Squared Error (MSE) can be introduced as auxiliary methods, or numerical values ​​can be extracted and compared after the model outputs structured results. This can better guide the model in numerical inference.

[0087] 4. Training Completion Criterion: Training is complete when the mean absolute error (MAE) or mean absolute percentage error (MAPE) between the values ​​inferred by the model on the validation set and the true label values ​​is lower than a preset minimum threshold. The model at this point is called the "contextual numerical inference model".

[0088] In this embodiment, the atomic intent "adjust sewing speed" is processed, where "faster" is a relative instruction. At this time, the value inference module 130 is activated. This module has a built-in "contextual numerical inference model" specializing in numerical reasoning, namely the third major language model. It first sends a request to the device status center 131 through the interface to obtain the current real-time status of the SN-A087 sewing machine (e.g., ...). Figure 3 As shown in D): {"current_speed":2000} (unit: rpm). This data serves as key context information for parsing this relative instruction.

[0089] Then, the value inference module 130 provides the physical state context (current speed) along with the atomic intention ("adjust sewing speed") to the context value inference model. Because the model has undergone extensive quantitative inference training involving device states, it has learned to understand the relative meaning of "a point." In this example, it calculates based on its internal quantification logic (e.g., learned or pre-defined rules: for the medium speed range, "a point" represents an increase of 15%): 2000 * (1 + 0.15) = 2300. Thus, the vague intention is inferred as a specific numerical value.

[0090] Step 4: Combine all atomic operation intentions and their parameter target values ​​into a pre-formatted structured intention data.

[0091] Finally, the intent synthesis module 140 collects and integrates the results of all processed subtasks to generate a unified, structured final intent data. During this process, it also performs logical self-checks, such as checking for conflicts between the "beaded edge stitch" mode and the default stitch settings in the "shirt cuff" process template, and merging them according to preset rules (e.g., explicit instructions have higher priority than template defaults).

[0092] Finally, the system outputs a document as follows: Figure 3 The data shown in E is fully structured JSON data. This data is clear, unambiguous, and has complete parameters, and can be directly submitted to the subsequent instruction generation and validation module.

[0093]

[0094]

[0095]

[0096] As demonstrated in this embodiment, the method of the present invention, through a cascading process driven by a specialized model, can systematically convert an extremely complex industrial instruction, rich in colloquial language and implicit information, into standardized intent data that a machine can easily understand and process with high fidelity. This clearly demonstrates how the present invention handles complex tasks that RAG struggles with through its unique "instruction decomposition" step, introduces the physical state context lacking in RAG through its "value inference" step, and ensures the absolute determinism required for industrial control by ultimately generating structured JSON rather than natural language text. This lays a solid foundation for achieving truly intelligent and reliable voice control of industrial equipment.

[0097] Example 2

[0098] This embodiment provides a multi-level parsing system for industrial sewing instructions that integrates real-time equipment status, such as... Figure 1 As shown, it includes:

[0099] Instruction decomposition module 110 receives natural language instruction text and decomposes the instruction text into one or more atomic operation intentions;

[0100] The entity parsing module 120, for each atomic operation intention, identifies and standardizes the entity parameters in the atomic operation intention by combining at least one preset domain knowledge base, wherein the domain knowledge base includes a process term knowledge base 121 and a process template knowledge base 122.

[0101] The value inference module 130 identifies relative instructions that depend on the current state of the device, obtains the current state data of the device, and infers the relative instructions into specific parameter target values.

[0102] The equipment's status data for the current year is obtained through the equipment status center.

[0103] The intent synthesis module 140 synthesizes all atomic operation intents and their parameter target values ​​into a pre-formatted structured intent data.

[0104] The instruction decomposition module has a built-in dedicated instruction decomposition model; the entity parsing module has a built-in domain entity parsing model; and the value inference module has a built-in context value inference model.

[0105] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A multi-stage parsing method of industrial sewing instructions fusing real-time states of devices, characterized in that, The method comprises the following steps: Step 1, receiving a natural language instruction text, and decomposing the instruction text into one or more atomic operation intentions by using a pre-trained first large language model; Step 2, identifying and standardizing entity parameters in each atomic operation intention by using a pre-trained second large language model in combination with a pre-set at least one domain knowledge base; Step 3, identifying a relative instruction that needs to rely on the current state of the equipment, obtaining the current state data of the equipment, and inferring the relative instruction into a specific parameter target value by using a pre-trained third large language model; Step 4, combining all atomic operation intentions and parameter target values into a structured intention data in a pre-set format.

2. The method of claim 1, wherein, The domain knowledge base comprises a process art language knowledge base storing the mapping relationship between process art terms and standard parameter names, and a process template knowledge base storing the mapping relationship between a specific process and a set of standard parameter combinations.

3. The method of claim 1, wherein, In step 3, the current state data of the equipment is provided as context information to the third large language model to assist it in inference.

4. The method of claim 1, wherein, In step 4, it also includes checking whether there is a logical conflict between the parameters parsed by different atomic operation intentions.

5. The method of claim 1, wherein, The first large language model is a command decomposition special-purpose model fine-tuned for command decomposition tasks, and the training data used by the first large language model includes composite instructions and atomic operation intention sequences corresponding to the instructions.

6. The method of claim 1, wherein, The second large language model is a domain entity parsing model based on a retrieval-enhanced generation architecture, and the training data used by the second large language model includes atomic operation intentions, knowledge retrieved from the domain knowledge base, and labeled standardized entity parameters.

7. The method of claim 1, wherein, The third large language model is a context numerical inference model fine-tuned for numerical reasoning tasks, and the training data of the third large language model includes relative instructions, current state data of the equipment, and labeled parameter target values.

8. The method of claim 1, wherein, The first large language model, the second large language model, and the third large language model are the same large language model instance, and the functional specialization is realized by calling different adapter modules or prompt engineering templates specialized for the corresponding steps in different steps.

9. A multi-stage intent resolution system for an industrial sewing process, the system comprising: It comprises: An instruction decomposition module that receives a natural language instruction text and decomposes the instruction text into one or more atomic operation intentions; An entity parsing module that identifies and standardizes entity parameters in each atomic operation intention in combination with a pre-set at least one domain knowledge base; A value inference module that identifies a relative instruction that needs to rely on the current state of the equipment, obtains the current state data of the equipment, and infers the relative instruction into a specific parameter target value; An intention synthesis module that combines all atomic operation intentions and parameter target values into a structured intention data in a pre-set format.

10. The system of claim 9, wherein, The instruction decomposition module is built-in with a command decomposition special-purpose model; the entity parsing module is built-in with a domain entity parsing model; and the value inference module is built-in with a context numerical inference model.

Images