Spinning equipment design assisting method and system based on model collaboration

Abstract

The invention belongs to the technical field of computer aided design, and provides a spinning equipment design assisting method and system based on model collaboration. The method comprises the steps that a collaborative design system is constructed, a large language model receives design requirements of a user, performs task decomposition and calls corresponding professional agent models to execute design subtasks, and the professional agent models feed simulation calculation results back to a closed-loop feedback engine; the closed-loop feedback engine analyzes feedback results of the professional agent models and generates attention prompt signals pointing to potential design conflicts or optimization bottlenecks on the basis of interdisciplinary parameter coupling relations; and the large language model performs multidisciplinary coupling analysis by using the attention prompt signal, dynamically adjusts task planning and design parameters by focusing on key parameters and key disciplines pointed by the attention prompt signal, and drives related professional agent models to perform directional iterative optimization. According to the method, the blind iteration problem can be effectively avoided, and the convergence speed of the design and the scheme quality are improved in a high-quality mode.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer-aided design, in particular to a spinning equipment design assistance method and system based on model collaboration. BACKGROUND

[0002] Spinning equipment design is a typical multi-disciplinary coupled complex system problem, and its design process involves deep collaboration of multiple professional fields such as process, control, transmission, rack and modeling. With the breakthrough of artificial intelligence technology, especially the strong ability of large language models in natural language understanding and task planning, a new paradigm for design automation is provided. In the prior art, researchers have tried to introduce large models into the field of mechanical design, and have formed the following several typical schemes: the first scheme adopts a multi-agent large model to perform collaborative design through language communication, although it stimulates design diversity, but it is difficult to handle the accuracy requirements of numerical calculation and physical simulation in mechanical design.

[0003] The second scheme integrates an external function library or professional software interface for the large model to perform specific calculation or simulation tasks. However, when multiple disciplinary models (such as process, transmission, control) are executed in parallel, this method will generate a large amount of heterogeneous simulation and parameter data for complex systems such as spinning equipment. At this time, the large language model as the overall designer faces the dilemma of information overload when performing multi-disciplinary coupled analysis. It lacks an effective mechanism to quickly and accurately identify the current most critical design conflicts, the most urgent optimization bottlenecks, or the most efficient iteration paths from the complex cross-influences. This leads to an inefficient decision-making process for the large model, and it may even ignore the deep nonlinear interactions between different disciplinary parameters, thereby making the optimization process of the entire collaborative system fall into blind iteration or slow convergence, and unable to truly realize efficient and accurate automated design closed loop.

[0004] Therefore, how to construct an intelligent feedback mechanism in the design framework based on the collaboration of large models and multi-disciplinary models to significantly improve the efficiency of multi-disciplinary coupled analysis and the convergence speed of design iteration is a technical problem to be solved at present. SUMMARY

[0005] To solve the above technical problems, the present application provides a spinning equipment design assistance method and system based on model collaboration.

[0006] The application provides a spinning equipment design auxiliary method based on model cooperation, comprising the following steps: constructing a cooperative design system, wherein the cooperative design system comprises: a large language model fine-tuned by domain knowledge, serving as a total planning and decision unit; a plurality of field-specific generative professional agent models, used for performing professional calculation and simulation of each discipline; and a closed-loop feedback engine with signal generation capability; the large language model receives user design requirements, performs task decomposition, and calls corresponding professional agent models to perform design subtasks, each professional agent model feeds back simulation calculation results to the closed-loop feedback engine; the closed-loop feedback engine analyzes the feedback results of each professional agent model, and generates attention prompt signals pointing to potential design conflicts or optimization bottlenecks based on cross-disciplinary parameter coupling relationships; and the large language model performs multi-disciplinary coupling analysis by using the attention prompt signals, focuses on key parameters and key disciplines pointed by the attention prompt signals, dynamically adjusts task planning and design parameters, and drives related professional agent models to perform directional iterative optimization.

[0007] The application also provides a spinning equipment design auxiliary system based on model cooperation, comprising a processing unit and a storage unit, wherein the processing unit calls and executes a pre-stored computer program in the storage unit to perform the following steps: constructing a cooperative design system, wherein the cooperative design system comprises: a large language model fine-tuned by domain knowledge, serving as a total planning and decision unit; a plurality of field-specific generative professional agent models, used for performing professional calculation and simulation of each discipline; and a closed-loop feedback engine with signal generation capability; the large language model receives user design requirements, performs task decomposition, and calls corresponding professional agent models to perform design subtasks, each professional agent model feeds back simulation calculation results to the closed-loop feedback engine; the closed-loop feedback engine analyzes the feedback results of each professional agent model, and generates attention prompt signals pointing to potential design conflicts or optimization bottlenecks based on cross-disciplinary parameter coupling relationships; and the large language model performs multi-disciplinary coupling analysis by using the attention prompt signals, focuses on key parameters and key disciplines pointed by the attention prompt signals, dynamically adjusts task planning and design parameters, and drives related professional agent models to perform directional iterative optimization.

[0008] The application has at least the following beneficial technical effects: the application introduces an attention prompt signal generation mechanism in the closed-loop feedback engine, so that the large language model can quickly identify key design conflicts from a large amount of multi-disciplinary data, the mechanism can significantly improve the efficiency of multi-disciplinary parameter coupling analysis through active diagnosis and intelligent guidance, concentrates optimization resources on the most critical design bottlenecks, effectively avoids the problem of blind iteration in traditional methods, and finally greatly improves the convergence speed and scheme quality of the spinning equipment cooperative design. BRIEF DESCRIPTION OF DRAWINGS

[0009] Figure 1 This is a flowchart of a model-based collaborative spinning equipment design assistance method disclosed in an embodiment of the present invention.

[0010] Figure 2 This is a schematic diagram of the architecture of the collaborative design system disclosed in an embodiment of the present invention.

[0011] Figure 3 This is a structural diagram of a model-based collaborative spinning equipment design auxiliary system disclosed in an embodiment of the present invention. Detailed Implementation

[0012] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0013] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0014] To facilitate understanding of the embodiments of the present invention, further explanations and descriptions will be provided below with reference to the accompanying drawings and specific embodiments. These embodiments do not constitute a limitation on the embodiments of the present invention.

[0015] like Figure 1 As shown, this embodiment of the invention discloses a model-based collaborative spinning equipment design assistance method 100, which includes the following steps: Step S1: Construct a collaborative design system, which includes: a large language model fine-tuned by domain knowledge as the overall planning and decision-making unit; multiple domain-specific generative professional agent models for performing professional calculations and simulations of various disciplines; and a closed-loop feedback engine with signal generation capabilities.

[0016] In this step, the solution of the present invention operates on an auxiliary platform specifically designed for the intelligent design of spinning equipment. The core of this auxiliary platform is the construction of a collaborative design system, referring to... Figure 2As shown, the collaborative design system specifically includes the following components: a large language model fine-tuned with domain knowledge. This large language model serves as the overall planning and decision-making unit of the design support system, its core role being similar to the chief designer in a traditional design process. By fine-tuning this large language model with professional knowledge in the spinning equipment field and integrating Retrieval Enhanced Generation (RAG) technology, it deeply masters knowledge from multiple disciplines such as textile technology, mechanical design principles, and materials mechanics, forming a professional knowledge system covering the entire design process of spinning equipment. This large language model is responsible for understanding the user's macro or specific design needs and making global task planning and decisions accordingly.

[0017] Multiple domain-specific generative professional agent models: These professional agent models serve as specialized execution units for the design support system, focusing on five core professional domains of spinning equipment: process, control, transmission, frame, and design. These professional agent models are constructed using advanced modeling methods such as Physical Information Neural Networks (PINs), Transformer-Encoder, and Graph Neural Networks, possessing high-fidelity numerical calculation and physical simulation capabilities. For example, the process model can dynamically calculate core process parameters such as draft ratio and distribution ratio; the transmission model can use PINN to replace traditional finite element analysis to predict the dynamic performance of the transmission system.

[0018] A closed-loop feedback engine with signal generation capabilities: This closed-loop feedback engine is the core hub connecting the large language model and various specialized agent models. It is responsible for scheduling tasks and transferring data between them, and has advanced analysis and signal generation capabilities.

[0019] Step S2: The large language model receives the user's design requirements, decomposes the tasks, and calls the corresponding professional agent models to execute the design sub-tasks. Each professional agent model feeds back the simulation calculation results to the closed-loop feedback engine.

[0020] In this step, the domain-enhanced large language model receives user-input design requirements, such as designing a high-efficiency, low-energy spinning machine. Based on its semantic understanding and logical reasoning capabilities, the large language model parses and decomposes complex design requirements, generating a global design task plan.

[0021] The large language model acts as a scheduling center, invoking corresponding professional agent models to execute specific design sub-tasks according to the task plan. For example, the large language model instructs the process small model to calculate the optimal combination of process parameters, while instructing the transmission small model to perform strength simulations on key transmission components. Each invoked professional agent model runs its high-fidelity professional calculations or physical simulations and feeds back the simulation results to the closed-loop feedback engine in real time. Simulation results include, for example, the tensile force data output by the process small model and the gear contact stress data output by the transmission small model.

[0022] Step S3: The closed-loop feedback engine analyzes the feedback results of each professional agent model and generates attention prompt signals pointing to potential design conflicts or optimization bottlenecks based on the cross-disciplinary parameter coupling relationship.

[0023] In this step, the closed-loop feedback engine receives and integrates heterogeneous simulation results data from various professional proxy models. Simultaneously, the closed-loop feedback engine has a pre-built cross-disciplinary parameter coupling relationship rule base constructed based on a multi-disciplinary design knowledge graph of spinning equipment. This rule base defines the constraints, conflicts, and causal relationships between parameters from different disciplines. For example, increasing the draft ratio in the process parameters may lead to an increase in the output torque of the transmission system, thereby causing excessive gear contact stress.

[0024] The closed-loop feedback engine matches and analyzes the integrated heterogeneous simulation results data with the rule base to proactively diagnose potential design conflicts or optimization bottlenecks in the current design state. Based on the diagnostic results, the closed-loop feedback engine generates structured attention prompts pointing to these key contradictions.

[0025] It is understood that the above attention prompts include at least: focus parameters (e.g., gear contact stress), related disciplines (e.g., process and transmission), conflict level (e.g., high risk), and suggested optimization directions (e.g., please prioritize coordinating the draw ratio and gear module).

[0026] Step S4: The large language model uses the attention cue signal to perform multidisciplinary coupling analysis, focusing on the key parameters and key disciplines indicated by the attention cue signal, dynamically adjusting the task planning and design parameters, and driving the relevant professional agent model to perform targeted iterative optimization.

[0027] In this step, the large language model receives and utilizes attentional cues from the closed-loop feedback engine during subsequent multidisciplinary coupling analysis. These cues guide the large language model to quickly focus on the most pressing key parameters and disciplines from massive amounts of multidisciplinary data.

[0028] Based on the aforementioned focus information, the large language model dynamically adjusts its task planning and design parameters, and precisely drives relevant specialized agent models to perform targeted iterative optimization. For example, based on the attention signals mentioned above, the large language model prioritizes adjusting the draft ratio parameter and then calls upon the process and transmission models again for collaborative verification, rather than blindly adjusting all parameters.

[0029] Through this intelligent closed loop of perception-diagnosis-focusing-optimization, the design assistance system can effectively avoid ineffective trial and error on non-critical parameters, significantly improve the solution efficiency and convergence speed of complex multidisciplinary coupled problems, and finally output an optimized design scheme that satisfies multi-objective constraints.

[0030] This invention introduces an attention cue signal generation mechanism into a closed-loop feedback engine, enabling large language models to quickly identify key design conflicts from massive multidisciplinary data. This mechanism, through proactive diagnosis and intelligent guidance, can significantly improve the efficiency of multidisciplinary parameter coupling analysis, concentrate optimization resources on the most critical design bottlenecks, effectively avoid the blind iteration problem in traditional methods, and ultimately achieve a high-quality improvement in the convergence speed and solution quality of collaborative design of spinning equipment.

[0031] As an example, fine-tuning the large language model using domain knowledge includes: Step S11: Obtaining a professional text dataset in the field of spinning equipment design, including design manuals, academic literature, and historical design schemes; constructing training samples based on the professional text dataset, and using an instruction fine-tuning method to perform domain adaptation training on the pre-trained large language model.

[0032] In this step, a professional text dataset in the field of spinning equipment design is obtained from the company's private database and public resources. This professional text dataset mainly includes: design manuals: containing structured knowledge such as component standards, design specifications, tolerances, and material selection for spinning equipment (such as spinning frames and winding machines); academic literature: including cutting-edge research papers and technical reports on new spinning processes, mechanism dynamics analysis, and control system optimization; and historical design schemes: the company's past practical experience, including successful design cases, simulation reports, and fault improvement records.

[0033] Training samples were constructed based on the aforementioned specialized text dataset. These training samples were organized into an instruction-input-output format; for example, the instruction was "Calculate the output torque at a given spindle speed and draft ratio," the input was specific parameters, and the expected output was the correct calculation steps and results. An instruction fine-tuning method was employed, using these constructed samples to perform domain-adaptive training on the pre-trained large language model. Understandably, this process does not change the model's basic architecture, but rather adjusts its internal parameters, enabling it to learn and internalize the specialized terminology, logical relationships, and problem-solving patterns of spinning equipment design, thereby accurately understanding and responding to domain-specific design instructions.

[0034] Step S12: Integrate the trained large language model with the retrieval enhancement generation framework, enabling it to retrieve and fuse relevant knowledge from the professional knowledge base based on the design context, thereby enhancing the professionalism and accuracy of its design decisions; wherein, the professional knowledge base is constructed by vectorizing the professional text dataset.

[0035] In this step, the large language model, fine-tuned in step S11, is integrated with the retrieval-enhanced generation framework. When performing a specific design task (e.g., a user requests to "optimize roller spacing to reduce breakage rate"), the retrieval-enhanced generation framework first automatically retrieves the most relevant knowledge fragments (such as recommended value ranges in design manuals and optimization models in literature) from the constructed professional knowledge base (built from the professional text dataset in step S11 after vectorization) based on the current design context.

[0036] Then, the large language model does not generate answers solely based on its own parameters. Instead, it integrates the retrieved relevant knowledge with the user's original question, using both as context for generating the response. This mechanism enhances the professionalism and accuracy of its design decisions, ensuring that the final output of solutions, parameter calculations, or task planning is deeply rooted in validated domain knowledge rather than model conjecture. This, in turn, significantly improves the reliability of design assistance systems in engineering design scenarios.

[0037] As an example, a closed-loop feedback engine with signal generation capability is constructed in the following way: Step S13: Construct a cross-disciplinary parameter coupling relationship rule base. This rule base is constructed based on the multi-disciplinary design knowledge graph of spinning equipment and defines the constraint relationships, conflict modes and causal relationships between parameters in the fields of process, control, transmission, frame and shape.

[0038] In this step, for example, a multidisciplinary design knowledge graph for spinning equipment is constructed based on GraphRAG technology. This knowledge graph organizes the core parameters, attributes, and their relationships in a structured entity-relationship format across five major areas: process, control, transmission, frame, and design. Based on this, a cross-disciplinary parameter coupling relationship rule base is built.

[0039] This rule base is a logical and computable expression of complex relationships in a knowledge graph. Specifically, it defines the following three types of core relationships: Constraint relationships: These define the hard conditions that parameters must meet. For example, the torque of a drive shaft must be less than the allowable torque of its material.

[0040] Conflict mode: Defines the inherent contradiction between different disciplinary parameter objectives. For example, increasing the spindle speed (process objective) to increase output will usually increase vibration (rack dynamics objective), which constitutes a conflict.

[0041] Causal relationship: Defines the chain effect caused by parameter adjustment. For example, increasing the draw ratio (process parameter) will lead to an increase in draw force (process result), which in turn will lead to an increase in the torque demand of the roller drive (transmission parameter), and may cause the motor current to exceed the limit (control parameter).

[0042] Step S14: Configure the signal generation logic module, which is set to: receive and integrate the simulation calculation results of various professional agent models, perform real-time matching analysis with the rule base, identify parameter deviations exceeding preset thresholds or conflict combinations that violate design constraints; and generate structured attention prompt signals based on the identification results, which at least include focus parameters, related disciplines, conflict levels and suggested optimization directions.

[0043] In this step, a signal generation logic module is configured, which is given the following two core functions: real-time diagnosis: the generation logic module continuously receives and integrates simulation calculation results from various professional agent models (e.g., the current tensile force calculated by the process model and the gear contact stress calculated by the transmission model). These real-time data are matched and analyzed in real time with the rule base constructed in step S13. Through this analysis, the generation logic module can identify two types of key problems: (1) parameter deviations exceeding the preset threshold: for example, the calculated gear contact stress value exceeds the material fatigue strength safety threshold; (2) conflict combinations that violate design constraints: for example, the currently set combination of "spindle speed" and "roller spacing" triggers the "high speed-small spacing" vibration conflict mode defined in the rule base.

[0044] Generate Guidance: Based on the above identification results, the generation logic module generates a structured attention cue signal. It should be noted that this attention cue signal is not a simple alarm, but a data structure containing specific action guidelines. At a minimum, it includes: focus parameters: clearly indicating the core of the problem, such as "gear contact stress"; related disciplines: indicating areas requiring collaboration, such as "process" and "transmission"; conflict level: assessing the severity of the problem, such as "high risk"; suggested optimization directions: providing preliminary solutions, such as "suggesting reducing the draw ratio or checking gear strength".

[0045] This implementation method, by constructing a cross-disciplinary parameter coupling relationship rule base and configuring a signal generation logic module, enables the closed-loop feedback engine to perform real-time diagnosis and intelligent guidance based on domain knowledge. It can transform massive simulation data into attention prompts that accurately locate design conflicts and optimization paths, thereby solving the core problems of information overload and blind decision-making in multidisciplinary collaborative design, and thus significantly improving the accuracy and efficiency of design iteration.

[0046] As an example, the large language model uses the attention cue signal to perform multidisciplinary coupling analysis, focusing on the key parameters and key disciplines indicated by the attention cue signal, and dynamically adjusting the task planning and design parameters, including: Step S41: The large language model parses the attention cue signal to obtain the conflict level, focus parameters, related disciplines, and suggested optimization directions.

[0047] In this step, the large language model performs in-depth analysis of the received attention cue signals, extracting all the key information elements contained therein, including: conflict level, used to quantify the severity of the current design conflict, such as "high risk", "medium risk" or "warning"; focus parameters, used to indicate the core parameters that cause the conflict, such as "gear contact stress"; related disciplines, used to indicate the professional fields related to the conflict, such as "process" and "transmission"; and suggested optimization directions, used to provide preliminary solutions, such as "suggest reducing the draw ratio or checking the gear strength".

[0048] Step S42: Determine the parameter optimization priority based on the conflict level; determine the core parameter set to be optimized and related professional proxy models based on the focus parameters and related disciplines; generate a specific parameter adjustment instruction set based on the suggested optimization direction; the parameter adjustment instruction set specifies the parameter identifier, adjustment direction and adjustment range to be adjusted.

[0049] In this step, the large language model makes the following comprehensive decision based on the information parsed in step S41: (1) Determine the priority of parameter optimization: Based on the conflict level, the large language model determines the execution order of optimization tasks. For example, the optimization of "high-risk" conflicts is placed with the highest priority to ensure that system resources are prioritized to solve the most urgent problems.

[0050] (2) Target optimization: Combine the focus parameters and related disciplines to accurately determine the set of core parameters to be optimized and the relevant professional proxy models. For example, if the focus parameter is "gear contact stress" and the related disciplines are "process" and "transmission", then the set of core parameters includes "drawing ratio" in the process field and "gear module" in the transmission field, and the relevant professional proxy models are "process mini-model" and "transmission mini-model".

[0051] (3) Generate specific instructions: Refer to the suggested optimization direction and generate a specific set of parameter adjustment instructions. It should be noted that this set of parameter adjustment instructions is a list of operations that can be directly executed by the professional proxy model, including: parameter identifier, such as process, draw ratio; adjustment direction, such as reduction; adjustment range, such as 5%-10%.

[0052] Step S43: Based on the parameter optimization priority, sequentially call the professional agent models corresponding to the core parameter set to execute the parameter adjustment instruction set.

[0053] In this step, the large language model acts as the scheduling center, and according to the parameter optimization priority determined in the previous steps, it sequentially calls the professional agent models corresponding to the core parameter set determined in step S42, and instructs them to execute the parameter adjustment instruction set.

[0054] For example, for the highest priority conflict, the large language model first sends an instruction to the small process model, requesting it to reduce the "drawing ratio" by 5%. After the small process model completes the calculation and returns the new "drawing force" data, the large language model then calls the small transmission model and uses the new process data to perform transmission performance simulation to verify whether the "gear contact stress" has been reduced below the safety threshold.

[0055] This implementation method parses attention cue signals into specific parameter optimization priorities, core parameter sets, and executable parameter adjustment instruction sets, enabling large language models to transform macroscopic conflict diagnosis into precise, quantifiable, and orderly optimization operations. This achieves the transition from problem identification to precise execution, thereby significantly improving the targeting, automation, and iterative convergence efficiency of multidisciplinary coupled optimization.

[0056] As an example, if the suggested optimization direction output by the closed-loop feedback engine is associated with a confidence level, then the step of generating a specific parameter adjustment instruction set based on the suggested optimization direction includes: Step S421: Combining and mapping the conflict level with the confidence level to determine the global optimization strategy for this iteration, including an aggressive optimization strategy, a robust optimization strategy, and an exploration and verification strategy.

[0057] In this step, since different design conflicts have significant differences in urgency and solution reliability, adopting a uniform optimization strategy will lead to inefficiency or even new design risks. Therefore, this implementation combines conflict level with confidence level to establish a global optimization strategy based on risk assessment: (1) When the conflict level is high and the confidence level is high, since there is a serious design problem and the solution is reliable, it is necessary to quickly eliminate the risk. At this time, an aggressive optimization strategy is adopted.

[0058] (2) When the conflict level is high but the confidence level is low, the problem is serious but the solution is highly uncertain. It is necessary to control potential risks while solving the problem. In this case, a robust optimization strategy is adopted.

[0059] (3) When the conflict level is medium / low risk and the confidence level is low, it is in a relatively safe state and suitable for accumulating design knowledge through multi-directional exploration. At this time, the exploration and verification strategy is adopted.

[0060] Step S422: Select the generation template for parameter adjustment instructions according to the global optimization strategy. Specifically: if it is an aggressive optimization strategy, generate a single instruction with a large adjustment range; if it is a robust optimization strategy, generate an instruction sequence containing a main adjustment instruction and a backup adjustment instruction; if it is an exploration and verification strategy, generate multiple parallel test instructions with small amplitudes and different directions.

[0061] This step aims to transform the aforementioned defined, abstract global optimization strategy into concrete, executable code generation logic. Specifically, the large language model selects a generation template for parameter adjustment instructions from a predefined instruction template library based on the defined global optimization strategy.

[0062] If the strategy is an aggressive optimization strategy, then the first template is selected. The logic of this template is to generate a single instruction with a large adjustment range. For example, the instruction format is {parameter identifier: stretch ratio; adjustment direction: decrease; adjustment range: 15%}, which can then quickly approach the target through a single large adjustment.

[0063] If the strategy is a robust optimization strategy, then the second template is selected. The logic of this template is to generate an instruction sequence containing a main adjustment instruction and a backup adjustment instruction. For example, the sequence is [{Main instruction: Reduce the draft ratio by 10%}, {Backup instruction: If the stress still exceeds the limit, increase the gear module by one level}].

[0064] If the strategy is an exploratory verification strategy, then the third template is selected. The logic of this template is to generate multiple small-amplitude, parallel test instructions in different directions. For example, the instruction set could be: {Test 1: Decrease draft ratio by 5%; Test 2: Increase draft ratio by 3%; Test 3: Increase roller spacing by 2%}. This logic does not immediately resolve conflicts, but rather quickly obtains data on the impact of different parameter changes on the system through parallel experiments.

[0065] Step S423: Fill in the specific details of the suggested optimization direction into the selected generation template to form the final parameter adjustment instruction set.

[0066] This step is the template instantiation process, where specific engineering parameters are filled into the template to form the final instructions that can drive the professional agent model to execute. Specifically, the large language model fills the generated template selected in step S422 with the specific content of the proposed optimization direction (e.g., suggesting "reducing the draw ratio" or "checking gear strength") as key parameters.

[0067] For aggressive optimization strategies, the specific suggestion of "reducing the stretch ratio" is entered into a single instruction template and automatically matched with a large magnitude (such as 15%) to form the final instruction {parameter identifier: stretch ratio; adjustment direction: reduce; adjustment magnitude: 15%}.

[0068] For the robust optimization strategy, "reducing the draw ratio" is taken as the main instruction, and "checking the gear strength" is converted into a backup instruction, forming an ordered instruction sequence.

[0069] For the exploration and verification strategy, a set of parallel test instructions for exploration is generated using core parameters such as "stretch ratio" and "roller spacing".

[0070] This implementation method achieves a leap from single instruction generation to multi-strategy adaptive decision-making by establishing a combination mapping mechanism between conflict level and confidence level. This enables the design-aided system to dynamically select aggressive optimization, robust optimization, or exploratory verification strategies based on the urgency of design risks and the reliability of solutions, and generate a matching set of parameter adjustment instructions. This significantly improves the intelligence level and execution efficiency of the optimization process while ensuring design safety.

[0071] As an example, the conflict level and the confidence level are combined and mapped to determine the global optimization strategy for this iteration, including: step S4211: constructing an adversarial evaluation network with noise injection, taking the conflict level and confidence level as input feature vectors, and adding random Gaussian noise to generate noisy feature vectors.

[0072] Due to the complex nonlinear coupling relationships between spinning process parameters (such as draft ratio and yarn tension, spindle speed and energy consumption), and the dynamic changes in the sensitivity of each parameter with equipment operating conditions, simple fixed preset rules are insufficient for accurate strategy decision-making. Furthermore, interference factors such as equipment vibration and sensor errors in real-world engineering environments lead to uncertainties in conflict level and confidence assessment. Therefore, this implementation introduces a deep learning-based intelligent decision-making mechanism. By constructing an adversarial evaluation network including noise injection, feature weighting based on an attention mechanism, and a reinforcement learning strategy network, it achieves intelligent strategy decision-making that counteracts interference and adaptively weights factors, thereby improving the system's decision-making quality and robustness in real industrial environments.

[0073] In this step, the conflict level is quantified into a continuous value within the interval [0,1], where 0.8-1.0 represents high risk (e.g., excessive decapitation rate), 0.5-0.8 represents medium risk (e.g., high energy consumption), and 0-0.5 represents low risk. Simultaneously, the confidence level is also quantified into a value within the interval [0,1], reflecting the reliability of the proposed optimization direction. These two features are combined into a two-dimensional feature vector [Xc,Xs], where Xc represents the conflict level and Xs represents the confidence level.

[0074] Next, Gaussian noise ε ~ N(0, 0.1) with a mean of 0 and a standard deviation of 0.1 is added to the feature vector [Xc, Xs] to generate the noisy feature vector [Xc + εc, Xs + εs], which is the interference signal. This noise level is derived from statistical analysis of the measurement error of the spinning equipment parameters and can effectively simulate the measurement fluctuations in actual production.

[0075] Through this adversarial training method, the network can learn to maintain stable feature representations under noise interference, which is particularly suitable for dealing with parameter noise problems unique to spinning equipment, such as draft force fluctuations and spindle speed measurement errors.

[0076] Step S4212: The noisy feature vector is nonlinearly transformed by a multilayer perceptron to extract an anti-interference deep feature representation. Based on the anti-interference deep feature representation, the dynamic weights of the conflict level feature and the confidence feature are calculated using an attention mechanism. The anti-interference deep feature representation is then weighted and fused to generate a strategy preference score.

[0077] In this step, the noisy feature vector is input into a three-layer MLP network, which includes: an input layer with two neurons that receive the noisy feature vector [Xc+εc, Xs+εs].

[0078] Hidden layer: Nonlinear transformation is performed through 16 neurons, using the ReLU activation function.

[0079] Output layer: 8 neurons, outputting an 8-dimensional robust deep feature representation H={h1,h2,...,h8}. This deep feature representation H has been trained for noise robustness and can effectively resist interference signals in the input data.

[0080] Subsequently, the deep feature representation is dynamically weighted based on an attention mechanism. Since the components in the deep feature H contribute differently to the final policy decision, a learnable attention parameter vector is used. Calculate the weight coefficients for each feature: ;in, It is a learnable attention parameter vector. Let j be the depth feature representation.

[0081] This weighting mechanism can automatically adjust the importance of each feature based on the current design state. For example, in high-speed spinning, vibration-related features are given higher weights; in precision spinning scenarios, tension control features are given higher weights.

[0082] Weighted fusion of deep feature representations: Finally, the weighted comprehensive feature representation that reflects the key information of the current design status is... Feature representation Input three parallel scoring functions: ;in, These are scoring functions for aggressive, robust, and exploratory strategies, respectively. Through training with a large amount of spinning equipment design data, they can accurately assess the applicability of each strategy in the current design state.

[0083] Step S4213: Input the policy preference score into the reinforcement learning policy network, combine it with the feedback of historical optimization results, and output the global optimization policy for this iteration.

[0084] In this step, a reinforcement learning policy network based on DQN is constructed, whose state space includes: Current strategy preference score .

[0085] Historical optimization performance metrics: for example, the improvement rate of the objective function over the past 5 iterations.

[0086] Equipment operating status characteristics: such as cumulative operating time and wear coefficient of key components.

[0087] The action space allows for the selection of three strategies. The reward function is designed as follows: ;in, It is the improvement amount of the objective function. It is the measure of improvement in the degree of violation of constraints. It is a negative reward for computational resource consumption; weighting coefficient Adjustments should be made according to the specific spinning process requirements.

[0088] Understandable This is used to quantify the degree of improvement that the optimization process brings to the main design objectives. In spinning equipment design, the objective function is usually a comprehensive representation of multiple objectives, including: production efficiency improvement rate, calculated by comparing the theoretical output before and after iteration, ∆output = (output of the new scheme - output of the original scheme) / output of the original scheme; quality index improvement rate, including improvement in yarn CV value, reduction in breakage rate, etc., calculated according to the weighted improvement of each quality index; and energy consumption reduction rate, calculated by the reduction in equipment operating energy consumption through a power model.

[0089] The specific calculation formula is as follows: .in The weighting coefficients for each objective are determined based on specific design requirements.

[0090] Used to evaluate the improvement of a design scheme in meeting various constraints. Constraints in spinning equipment design include: process constraints, such as the range of draft ratio and twist; structural constraints, such as the maximum allowable stress and deformation; and performance constraints, such as vibration amplitude limits and noise levels.

[0091] The degree of constraint violation is calculated using the constraint violation quantity function: ,in, Let inequality constraint function be used. This is the equality constraint function.

[0092] Positive values ​​indicate a reduced degree of constraint violation and an improved feasibility of the design scheme.

[0093] This is used to control the computational cost of the optimization process and avoid excessive resource consumption. It includes: simulation computation time cost (total time for each specialized proxy model to execute simulations); data transmission cost (resource consumption for data transmission and processing between system modules); and storage resource cost (storage requirements for intermediate results and iteration history). The specific calculation method is as follows: ;in, This is the normalized value of the simulation time. This is the normalized value of the data transmission volume. This is a normalized value for storage usage. This represents the corresponding cost coefficient.

[0094] The DQN-based reinforcement learning policy network is trained using the Q-learning algorithm: Among them, state Includes the current policy bias and device state, representing the environmental state perceived by the reinforcement learning policy network in the current iteration cycle; actions The corresponding strategy selection refers to the global optimization strategy chosen by the reinforcement learning policy network from the action space for this iteration, i.e., choosing one from {aggressive optimization strategy, robust optimization strategy, and exploration / validation strategy}; reward Based on the optimization effect calculation, it represents the feedback benefit immediately obtained after executing action a (i.e., adopting a certain optimization strategy) and completing this iteration; This represents the new environmental state perceived by the policy network after executing action a and entering the next iteration; Indicates the next state It considers all possible actions. In this way, the large language model can learn the optimization patterns of spinning equipment over long-term operation. In the new equipment stage, it tends to choose aggressive strategies to quickly find the best option, and in the equipment wear-out period, it switches to robust strategies to ensure stability, thus achieving adaptive optimization throughout the entire life cycle.

[0095] This implementation simulates parameter fluctuations in actual spinning equipment operation by constructing an adversarial evaluation network with noise injection. It utilizes a multilayer perceptron to extract anti-interference depth features and dynamically assigns weights to conflict level and confidence features based on an attention mechanism. Finally, it employs a reinforcement learning policy network combined with historical optimization results to achieve adaptive decision-making for the global optimization strategy. This effectively solves the strategy decision-making challenges of traditional methods under complex operating conditions such as multi-parameter coupling, measurement noise interference, and long-term performance degradation in spinning equipment.

[0096] like Figure 3As shown in the figure, this invention discloses a model-based collaborative spinning equipment design assistance system 200. The system includes a processing unit 2001 and a storage unit 2002. The processing unit 2001 calls and executes a computer program pre-stored in the storage unit 2002 to perform the following steps: constructing a collaborative design system, which includes: a large language model fine-tuned with domain knowledge as the overall planning and decision-making unit; multiple domain-specific generative professional agent models for performing professional calculations and simulations in various disciplines; and a closed-loop feedback engine with signal generation capabilities.

[0097] The large language model receives user design requirements, decomposes tasks, and calls the corresponding professional agent models to execute design sub-tasks. Each professional agent model feeds back the simulation calculation results to the closed-loop feedback engine.

[0098] The closed-loop feedback engine analyzes the feedback results of various professional agent models and generates attention prompts pointing to potential design conflicts or optimization bottlenecks based on cross-disciplinary parameter coupling relationships.

[0099] The large language model utilizes the attention cue signals to perform multidisciplinary coupling analysis, focusing on the key parameters and key disciplines indicated by the attention cue signals, dynamically adjusting task planning and design parameters, and driving relevant professional agent models to perform targeted iterative optimization.

[0100] It should be noted that the model-based collaborative spinning equipment design auxiliary system 200 of the present invention should also include at least an I / O interface 2003, a network port 2004, a bus 2005, etc., which will not be described in detail here.

[0101] As an example, fine-tuning the large language model using domain knowledge includes: acquiring a professional text dataset in the field of spinning equipment design, including design manuals, academic literature, and historical design schemes; constructing training samples based on the professional text dataset; performing domain adaptation training on the pre-trained large language model using an instruction fine-tuning method; and integrating the trained large language model with a retrieval enhancement generation framework, enabling it to retrieve and integrate relevant knowledge from a professional knowledge base based on the design context, thereby enhancing the professionalism and accuracy of its design decisions.

[0102] As an example, a closed-loop feedback engine with signal generation capabilities is constructed in the following way: a cross-disciplinary parameter coupling relationship rule base is built, which is based on the multi-disciplinary design knowledge graph of spinning equipment and defines the constraint relationships, conflict modes and causal relationships between parameters in the fields of process, control, transmission, frame and shape.

[0103] The configuration signal generation logic module is set to: receive and integrate the simulation calculation results of various professional agent models, perform real-time matching analysis with the rule base, identify parameter deviations exceeding preset thresholds or conflict combinations that violate design constraints; and generate structured attention prompt signals based on the identification results, which at least include focus parameters, related disciplines, conflict levels, and suggested optimization directions.

[0104] As an example, the large language model uses the attention cue signal to perform multidisciplinary coupling analysis, focusing on the key parameters and key disciplines indicated by the attention cue signal, and dynamically adjusting the task planning and design parameters, including: the large language model parses the attention cue signal to obtain the conflict level, focus parameters, related disciplines, and suggested optimization directions.

[0105] Based on the conflict level, the parameter optimization priority is determined. Based on the focus parameters and the related disciplines, the set of core parameters to be optimized and the relevant professional proxy models are determined. Based on the suggested optimization direction, a specific parameter adjustment instruction set is generated, which specifies the parameter identifier, adjustment direction, and adjustment range to be adjusted.

[0106] Based on the parameter optimization priority, the professional agent models corresponding to the core parameter set are called sequentially to execute the parameter adjustment instruction set.

[0107] As an example, if the suggested optimization direction output by the closed-loop feedback engine is associated with a confidence level, then generating a specific parameter adjustment instruction set based on the suggested optimization direction includes: combining and mapping the conflict level with the confidence level to determine the global optimization strategy for this iteration, including an aggressive optimization strategy, a robust optimization strategy, and an exploration and verification strategy.

[0108] The template for generating parameter adjustment instructions is selected based on the global optimization strategy. Specifically: if it is an aggressive optimization strategy, a single instruction with a large adjustment range is generated; if it is a robust optimization strategy, an instruction sequence containing a main adjustment instruction and a backup adjustment instruction is generated; if it is an exploratory verification strategy, multiple parallel test instructions with small amplitudes and different directions are generated.

[0109] Fill in the specific details of the suggested optimization direction into the selected generation template to form the final parameter adjustment instruction set.

[0110] As an example, combining the conflict level with the confidence level to determine the global optimization strategy for this iteration includes: constructing an adversarial evaluation network with noise injection, using the conflict level and confidence level as input feature vectors, and adding random Gaussian noise to generate noisy feature vectors.

[0111] The noisy feature vector is nonlinearly transformed by a multilayer perceptron to extract anti-interference deep feature representations; and a dynamic weight allocation between conflict level features and confidence features is calculated based on an attention mechanism to generate a strategy preference score.

[0112] The policy preference score is input into the reinforcement learning policy network, and combined with the feedback from historical optimization results, the global optimization policy for this iteration is output.

[0113] Although the invention has been specifically shown and described with reference to preferred embodiments, those skilled in the art will understand that various modifications in form and detail may be made without departing from the spirit and scope of the invention. Accordingly, the disclosed invention should be considered merely illustrative and limited only by the scope specified in the appended claims.

Claims

1. A model-based collaborative design assistance method for spinning equipment, characterized in that, The process includes the following steps: Constructing a collaborative design system, comprising: a large language model fine-tuned with domain knowledge as the overall planning and decision-making unit; multiple domain-specific generative professional agent models for performing professional calculations and simulations in each discipline; and a closed-loop feedback engine with signal generation capabilities. The large language model receives user design requirements, decomposes tasks, and calls the corresponding professional agent models to execute design sub-tasks. Each professional agent model feeds back simulation calculation results to the closed-loop feedback engine. The closed-loop feedback engine analyzes the feedback results of each professional agent model and, based on cross-disciplinary parameter coupling relationships, generates attention prompt signals pointing to potential design conflicts or optimization bottlenecks. The large language model uses these attention prompt signals to perform multi-disciplinary coupling analysis, focusing on the key parameters and key disciplines indicated by the attention prompt signals, dynamically adjusting task planning and design parameters, and driving relevant professional agent models to perform targeted iterative optimization.

2. The model-based collaborative design assistance method for spinning equipment according to claim 1, characterized in that: Fine-tuning the large language model using domain knowledge includes: acquiring a professional text dataset in the field of spinning equipment design, including design manuals, academic literature, and historical design schemes; constructing training samples based on the professional text dataset; and performing domain adaptation training on the pre-trained large language model using an instruction fine-tuning method; integrating the trained large language model with a retrieval enhancement generation framework, enabling it to retrieve and integrate relevant knowledge from a professional knowledge base based on the design context, thereby enhancing the professionalism and accuracy of its design decisions; wherein, the professional knowledge base is constructed from the professional text dataset after vectorization processing.

3. The model-based collaborative design assistance method for spinning equipment according to claim 1, characterized in that: A closed-loop feedback engine with signal generation capabilities is constructed as follows: A cross-disciplinary parameter coupling relationship rule base is built, which is based on a multi-disciplinary design knowledge graph of spinning equipment, defining constraint relationships, conflict modes, and causal relationships among parameters in the fields of process, control, transmission, frame, and styling; a signal generation logic module is configured to: receive and integrate simulation calculation results from various professional proxy models, perform real-time matching analysis with the rule base, identify parameter deviations exceeding preset thresholds or conflict combinations violating design constraints; and generate structured attention prompt signals based on the identification results, including at least the focus parameter, related disciplines, conflict level, and suggested optimization direction.

4. The model-based collaborative design assistance method for spinning equipment according to claim 3, characterized in that: The large language model utilizes the attention cue signals for multidisciplinary coupling analysis, focusing on the key parameters and key disciplines indicated by the attention cue signals, and dynamically adjusting task planning and design parameters. This includes: the large language model parsing the attention cue signals to obtain conflict level, focus parameters, related disciplines, and suggested optimization directions; determining parameter optimization priorities based on the conflict level; determining the core parameter set to be optimized and related professional agent models based on the focus parameters and related disciplines; generating a specific parameter adjustment instruction set based on the suggested optimization directions, specifying the parameter identifier, adjustment direction, and adjustment range to be adjusted; and sequentially calling the professional agent models corresponding to the core parameter set to execute the parameter adjustment instruction set according to the parameter optimization priorities.

5. The model-based collaborative design assistance method for spinning equipment according to claim 4, characterized in that: If the suggested optimization direction output by the closed-loop feedback engine is associated with a confidence level, then generating a specific set of parameter adjustment instructions based on the suggested optimization direction includes: combining and mapping the conflict level with the confidence level to determine the global optimization strategy for this iteration, including an aggressive optimization strategy, a robust optimization strategy, and an exploratory verification strategy; selecting a generation template for parameter adjustment instructions according to the global optimization strategy, specifically: if it is an aggressive optimization strategy, generating a single instruction with a large adjustment range; if it is a robust optimization strategy, generating an instruction sequence containing a main adjustment instruction and a backup adjustment instruction; if it is an exploratory verification strategy, generating multiple parallel test instructions with small amplitudes in different directions; and filling the specific content of the suggested optimization direction into the selected generation template to form the final set of parameter adjustment instructions.

6. The model-based collaborative design assistance method for spinning equipment according to claim 5, characterized in that: The process of combining and mapping the conflict level with the confidence level to determine the global optimization strategy for this iteration includes: constructing an adversarial evaluation network with noise injection, using the conflict level and confidence level as input feature vectors, and adding random Gaussian noise to generate noisy feature vectors; performing a nonlinear transformation on the noisy feature vectors using a multilayer perceptron to extract anti-interference deep feature representations; calculating the dynamic weights of the conflict level features and confidence level features based on the anti-interference deep feature representations using an attention mechanism, and performing weighted fusion on the anti-interference deep feature representations to generate a strategy preference score; inputting the strategy preference score into a reinforcement learning policy network, and combining feedback from historical optimization effects to output the global optimization strategy for this iteration.

7. A model-based collaborative design assistance system for spinning equipment, characterized in that, The system includes a processing unit and a storage unit. The processing unit calls and executes pre-stored computer programs in the storage unit to perform the following steps: constructing a collaborative design system, which includes: a large language model fine-tuned with domain knowledge as the overall planning and decision-making unit; multiple domain-specific generative professional agent models for performing professional calculations and simulations in various disciplines; and a closed-loop feedback engine with signal generation capabilities. The large language model receives user design requirements, decomposes tasks, and calls the corresponding professional agent models to execute design sub-tasks. Each professional agent model feeds back simulation calculation results to the closed-loop feedback engine. The closed-loop feedback engine analyzes the feedback results of each professional agent model and generates attention prompt signals pointing to potential design conflicts or optimization bottlenecks based on cross-disciplinary parameter coupling relationships. The large language model uses the attention prompt signals to perform multi-disciplinary coupling analysis, focusing on the key parameters and key disciplines indicated by the attention prompt signals, dynamically adjusting task planning and design parameters, and driving relevant professional agent models to perform targeted iterative optimization.

8. The model-based collaborative spinning equipment design assistance system according to claim 7, characterized in that: A closed-loop feedback engine with signal generation capabilities is constructed as follows: A cross-disciplinary parameter coupling relationship rule base is built, which is based on a multi-disciplinary design knowledge graph of spinning equipment, defining constraint relationships, conflict modes, and causal relationships among parameters in the fields of process, control, transmission, frame, and styling; a signal generation logic module is configured to: receive and integrate simulation calculation results from various professional proxy models, perform real-time matching analysis with the rule base, identify parameter deviations exceeding preset thresholds or conflict combinations violating design constraints; and generate structured attention prompt signals based on the identification results, including at least the focus parameter, related disciplines, conflict level, and suggested optimization direction.

9. The model-based collaborative spinning equipment design assistance system according to claim 8, characterized in that: The large language model utilizes the attention cue signals for multidisciplinary coupling analysis, focusing on the key parameters and key disciplines indicated by the attention cue signals, and dynamically adjusting task planning and design parameters. This includes: the large language model parsing the attention cue signals to obtain conflict level, focus parameters, related disciplines, and suggested optimization directions; determining parameter optimization priorities based on the conflict level; determining the core parameter set to be optimized and related professional agent models based on the focus parameters and related disciplines; generating a specific parameter adjustment instruction set based on the suggested optimization directions, specifying the parameter identifier, adjustment direction, and adjustment range to be adjusted; and sequentially calling the professional agent models corresponding to the core parameter set to execute the parameter adjustment instruction set according to the parameter optimization priorities.

10. A model-based collaborative spinning equipment design assistance system according to claim 9, characterized in that: If the suggested optimization direction output by the closed-loop feedback engine is associated with a confidence level, then generating a specific set of parameter adjustment instructions based on the suggested optimization direction includes: combining and mapping the conflict level with the confidence level to determine the global optimization strategy for this iteration, including an aggressive optimization strategy, a robust optimization strategy, and an exploratory verification strategy; selecting a generation template for parameter adjustment instructions according to the global optimization strategy, specifically: if it is an aggressive optimization strategy, generating a single instruction with a large adjustment range; if it is a robust optimization strategy, generating an instruction sequence containing a main adjustment instruction and a backup adjustment instruction; if it is an exploratory verification strategy, generating multiple parallel test instructions with small amplitudes in different directions; and filling the specific content of the suggested optimization direction into the selected generation template to form the final set of parameter adjustment instructions.

Images