Multi-ring coupling mining mechanism intelligent optimization method inspired by large language model

Abstract

The invention provides a multi-ring coupling mining mechanism intelligent optimization method inspired by a large language model. The method comprises the following steps: constructing an optimization model comprising a design variable, an optimization target and a constraint condition; then configuring a hyper-heuristic optimization framework, and initializing a historical knowledge base, a trigger mechanism and a bottom operator library; whether the large model is called or not is determined through triggering judgment, cue words are constructed by fusing multi-source state information, the operator scheduling proportion is generated by the large model or obtained through a standby strategy, filial generations are generated according to the operator scheduling proportion, a population is updated, and meanwhile a knowledge base is updated in a rolling mode; and finally outputting an optimal design scheme. According to the method, the deep reasoning capability of a large language model is fused, dynamic scheduling and adaptive combination of various underlying operators are achieved, the defects that a traditional algorithm is insufficient in adaptive capability and prone to falling into local optimum are effectively overcome, and the global search efficiency and the solving precision in high-dimensional, nonlinear and strong-constraint complex engineering problems are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of engineering machinery technology, specifically to an intelligent optimization method for multi-ring coupled excavation mechanisms inspired by a large language model. Background Technology

[0002] Hydraulic excavators, as important high-end engineering machinery, are widely used in key sectors of the national economy such as construction, mining, transportation, and water conservancy. The multi-ring coupled excavating mechanism is the core motion execution component of a hydraulic excavator, mainly composed of the boom, stick, bucket, and working hydraulic cylinders. Its parameter design directly determines the excavator's key performance indicators such as working range, digging capacity, and power characteristics, playing a decisive role in improving the overall efficiency of the machine and is a core aspect of excavator design and development. However, due to the complex operating conditions, numerous variables, and strong coupling of constraints of excavators, the parameter design of the multi-ring coupled excavating mechanism is a typical high-dimensional, nonlinear, and strongly constrained complex engineering optimization problem.

[0003] Traditional design patterns rely on a process of "engineering experience trial calculations - physical prototype verification - iterative scheme correction," which suffers from long development cycles, high costs, and low efficiency, severely restricting product innovation and performance improvement. While modern intelligent optimization algorithms (such as genetic algorithms, differential evolution algorithms, and particle swarm optimization) possess global search capabilities, when dealing with such complex optimization problems, the fixed operators of a single algorithm cannot effectively match the dynamic optimization process. They generally suffer from insufficient adaptive capabilities, susceptibility to local optima, and limited convergence accuracy, making it difficult to meet practical engineering needs in terms of optimization results.

[0004] In recent years, general artificial intelligence technologies, represented by large models, have made breakthroughs in semantic understanding, logical reasoning, and policy generation. The deep information analysis and high-level decision-making capabilities of large models can introduce more flexible heuristic strategies and adaptive mechanisms into intelligent optimization algorithms, providing a new technical path for solving the aforementioned complex engineering optimization problems. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an intelligent optimization method for multi-ring coupled mining mechanisms inspired by large language models, constructing a hyperheuristic intelligent optimization framework with a large model as the core decision-making unit. This method leverages the deep reasoning and policy generation capabilities of the large model to dynamically schedule and combine underlying optimization operators, achieving adaptive search in the optimization process and significantly improving solution quality and optimization efficiency.

[0006] The present invention achieves the above-mentioned technical objectives through the following technical means.

[0007] A method for intelligent optimization of multi-ring coupled mining mechanisms inspired by large language models includes the following steps:

[0008] Step S01: Construct an optimization model for the multi-ring coupled excavation mechanism, including determining design variables, optimization objectives, and constraints;

[0009] Step S02: Configure and initialize the hyperheuristic optimization framework, including setting optimization parameters, initializing the historical performance knowledge base, configuring the large model decision triggering mechanism, and building the underlying optimization operator library;

[0010] Step S03: Execute a large model-driven hyperheuristic optimization loop, including the following steps:

[0011] Step S3.1: Population initialization and fitness assessment;

[0012] Step S3.2: Determine whether the large model intelligent decision-making triggering conditions are met. The decision-making triggering conditions are usually determined by the initial warm-up period and the fixed call cycle.

[0013] Step S3.3: When the triggering condition is met, collect and fuse population statistical characteristics, convergence state indicators, stagnation state information and historical performance data of the underlying operators in real time to form a comprehensive multidimensional state information set;

[0014] Step S3.4: Construct prompt words and call the large language model to construct structured prompt words, and generate the optimization guidance strategy for the next stage by calling the large model;

[0015] Step S3.5: Analyze and process the guidance strategy for generating the large model or obtain the operator scheduling ratio through a robust guarantee mechanism;

[0016] Step S3.6: When the triggering conditions for intelligent decision-making in large models are not met, the operator scheduling ratio is obtained through the policy inheritance mechanism;

[0017] Step S3.7: Generate offspring and update the population according to the operator scheduling ratio;

[0018] Step S3.8: Update the historical performance knowledge base;

[0019] Step S3.9: Determine whether the termination condition is met. If not, return to step S3.1.

[0020] Step S04: Output the optimal design scheme.

[0021] Furthermore, the optimization objectives in step S01 include the working range objective, the boom digging force objective, and the bucket digging force objective. The working range objective is defined as the envelope area of ​​the discrete point set within the reachable working area of ​​the excavating mechanism.

[0022] Furthermore, the constraints in step S01 employ a hierarchical constraint processing mechanism, including:

[0023] For inviolable hard constraints, adopt an early return strategy and impose a death penalty;

[0024] The penalty function method is used to handle soft constraints that are violated to a certain degree.

[0025] Furthermore, the underlying optimization operator library in step S02 includes a variety of optimization operators with complementary search characteristics, including at least two of the following: simulated binary crossover operator, arithmetic crossover operator, polynomial mutation operator, DE / rand / 1 differential evolution operator, and DE / current-to-rand / 1 differential evolution operator.

[0026] Furthermore, the large model intelligent decision-making triggering conditions in step S3.2 include: triggering according to a preset periodic interval after the initial warm-up period of the optimization process.

[0027] Furthermore, the prompts in step S3.4 are preset standardized text templates, including: large model role definition, optimization problem description, operator characteristic description, current optimization state quantification data, operator historical performance statistics, dynamic strategy generation guidance, and output format specifications.

[0028] Furthermore, step S3.5 includes:

[0029] When the large model is successfully invoked, the guidance strategy generated by the large model is parsed and processed to obtain the operator scheduling ratio, including parsing the return result of the large model, performing legality checks and corrections, and normalizing the obtained operator scheduling ratio.

[0030] When a large model call fails or an abnormal response occurs, the operator scheduling ratio is obtained through a robust guarantee mechanism, specifically:

[0031] The scheduling ratio Ra of the operator is generated in real time using a weighted fusion formula. The specific calculation formula is as follows:

[0032] ,

[0033] In the formula Ra i Pe represents the scheduling ratio of operator i; i Let Ex be the recent average improvement gain of operator i; i Assigning a dynamic scaling factor to operator i based on the population convergence state; St i To assign a dynamic scaling factor to operator i based on the population stagnation generation; w p w e w s These are the weight coefficients for the corresponding items.

[0034] Furthermore, the strategy inheritance mechanism in step S3.6 includes: adopting a uniform allocation strategy during the warm-up period; and inheriting the operator scheduling ratio generated by the previous large model decision after the warm-up period but not during the call interval.

[0035] Furthermore, the offspring generation in step S3.7 includes: calculating the number of offspring that each operator should generate based on the dynamic scheduling ratio, generating a random permutation sequence of operator indices of corresponding length, allocating the corresponding operators to parent individuals in the population according to this sequence to generate offspring individuals, updating the population through an environment selection mechanism, and recording the iteration performance of each operator; the environment selection adopts a greedy strategy, and an offspring individual will only be retained if its fitness is better than that of its parent.

[0036] The beneficial effects of this invention are as follows:

[0037] 1. The intelligent optimization method for multi-ring coupled mining mechanism inspired by large language models described in this invention integrates the deep reasoning and strategy generation capabilities of large language models to construct a hyperheuristic optimization framework driven by a large model. This enables dynamic scheduling and adaptive combination of various underlying intelligent optimization operators, effectively overcoming the shortcomings of traditional intelligent optimization algorithms such as insufficient adaptive capability and easy getting trapped in local optima. It significantly improves the global search efficiency and solution accuracy of the algorithm in high-dimensional, nonlinear, and strongly constrained complex engineering problems.

[0038] 2. The intelligent optimization method for multi-ring coupled mining mechanisms inspired by large language models described in this invention, in terms of model construction, quantifies the working range into the envelope area of ​​the discrete point set of reachable regions, replaces multiple independent performance objectives with a single comprehensive index, simplifies the multi-objective trade-off problem, and adopts a hierarchical constraint processing mechanism to distinguish between different processing strategies for hard constraints and soft constraints, which significantly improves the efficiency of generating feasible solutions, reduces model complexity and solution difficulty, and helps to obtain high-performance multi-ring coupled mining mechanism design schemes. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. The drawings described below are some embodiments of the present invention. For those skilled in the art, it is obvious that other drawings can be obtained from these drawings without creative effort.

[0040] Figure 1 The flowchart is a process for the intelligent optimization method of the multi-ring coupled mining mechanism inspired by the large model described in this invention.

[0041] Figure 2 This is a schematic diagram of the overall framework of the multi-ring coupled excavation mechanism optimization model described in this invention;

[0042] Figure 3 This is a schematic diagram of the multi-ring coupled excavation mechanism described in this invention;

[0043] Figure 4 A schematic diagram illustrating the calculation of the envelope area of ​​the work area;

[0044] Figure 5 A flowchart for a hyperheuristic optimization loop driven by a large model;

[0045] Figure 6 This is a comparison of the convergence curves of the method of this invention and mainstream intelligent optimization algorithms.

[0046] Figure 7 The diagram shows the change in the scheduling ratio of each underlying operator in the process of optimizing using the method of this invention.

[0047] In the picture:

[0048] 1-Boom; 2-Stick; 3-Bucket; 4-Boom hydraulic cylinder; 5-Stick hydraulic cylinder; 6-Bucket hydraulic cylinder. Detailed Implementation

[0049] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0050] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "axial," "radial," "vertical," "horizontal," "inner," and "outer," etc., indicating orientation or positional relationships based on the orientation or positional relationships shown in the accompanying drawings, are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0051] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0052] The following will, in conjunction with the accompanying drawings, take the parameter optimization design of the multi-ring coupling excavation mechanism of a certain type of 70-ton hydraulic excavator as an example to clearly and completely describe the technical solution in the embodiment of the present invention.

[0053] refer to Figure 1 As shown in the figure, this embodiment provides an intelligent optimization method for multi-ring coupled mining mechanisms inspired by a large model. The specific implementation steps are as follows:

[0054] Step S01: Construct an optimization model for a multi-ring coupled excavation mechanism, the overall framework of which is as follows: Figure 2 As shown.

[0055] Figure 3 The diagram shows the structure of a multi-ring coupled digging mechanism for a 70-ton excavator, with the origin O at the intersection of the excavator's rotation center and the stopping surface. This digging mechanism consists of a boom 1, a stick 2, a bucket 3, and hydraulic cylinders 4 (boom), 5 (stick), and 6 (bucket). The components are connected via hinge points A, B, C, D, E, F, G, H, I, and J. Figure 2 As shown, point A is the hinge point between the lower end of boom 1 and the frame; point B is the hinge point between the upper end of boom 1 and one end of stick 2; point C is the hinge point between the other end of stick 2 and bucket 3; point D is the end point of the bucket teeth; point E is the hinge point between one end of the boom hydraulic cylinder and the frame; point F is the hinge point between the other end of the boom hydraulic cylinder and the middle of stick; point G is the hinge point between one end of the stick hydraulic cylinder and the middle of boom 1; point H is the hinge point between the other end of the stick hydraulic cylinder and the middle of stick 2; point I is the hinge point between one end of the bucket hydraulic cylinder and boom 1; and point J is the hinge point between the other end of the bucket hydraulic cylinder and bucket 3.

[0056] The goal of this embodiment is to optimize the key dimensional parameters of these components and the hinge coordinates to achieve the best excavation performance.

[0057] S1.1, Design Variables: In this embodiment, 23 independent parameters are selected as design variables, forming a decision vector X = [x1, x2, …, x…]. 23These design variables cover parameters such as the position coordinates of each key hinge point, mechanism dimensions, minimum length of the hydraulic cylinder, and extension ratio. Their value range is determined based on the actual design space of the 70-ton excavator, and the specific definitions are shown in Table 1.

[0058] Table 1. Definitions and Value Ranges of Design Variables in the Optimization Model

[0059] Design variables Physical meaning Range of values unit <![CDATA[x1]]> <![CDATA[The abscissa x of hinge point A A > [50, 700] mm <![CDATA[x2]]> <![CDATA[The ordinate y of the hinge point A A > [2000, 3200] mm <![CDATA[x3]]> <![CDATA[The abscissa x of the hinge point E E > [500, 1800] mm <![CDATA[x4]]> <![CDATA[The ordinate y of the hinge point E E > [1500, 2500] mm <![CDATA[x5]]> <![CDATA[Boom length L AB > [4000, 5500] mm <![CDATA[x6]]> <![CDATA[Stick length L BC > [2500, 5000] mm <![CDATA[x7]]> <![CDATA[The distance L from hinge point C to D CD > [1800, 2600] mm <![CDATA[x8]]> <![CDATA[The distance L from hinge point C to J CJ > [500, 1000] mm <![CDATA[x9]]> <![CDATA[The distance L from hinge point D to J DJ > [2000, 2800] mm <![CDATA[x 10 ]]> <![CDATA[The distance L from hinge point A to G AG > [600, 2000] mm <![CDATA[x 11 ]]> <![CDATA[The distance L from hinge point B to G BG > [2500, 4800] mm <![CDATA[x 12 ]]> <![CDATA[The distance L from hinge point A to F AF > [2500, 4500] mm <![CDATA[x 13 ]]> <![CDATA[The distance L from hinge point B to F BF > [1000, 2500] mm <![CDATA[x 14 ]]> <![CDATA[The distance L from hinge point A to I AI > [3000, 5000] mm <![CDATA[x 15 ]]> <![CDATA[Distance L from hinge point B to I BI > [500, 1500] mm <![CDATA[x 16 ]]> <![CDATA[The distance L from hinge point B to H BH > [1000, 2500] mm <![CDATA[x 17 ]]> <![CDATA[The distance L from hinge point C to H CH > [1000, 2500] mm <![CDATA[x 18 ]]> <![CDATA[The shortest length L of the boom cylinder 1min > [2000, 3500] mm <![CDATA[x 19 ]]> <![CDATA[The shortest length L of the arm cylinder 2min > [2000, 3500] mm <![CDATA[x 20 ]]> <![CDATA[The shortest length L of the bucket cylinder 3min > [2000, 3500] mm <![CDATA[x 21 ]]> <![CDATA[Boom cylinder extension ratio λ1]]> [1.5,1.8] - <![CDATA[x 22 ]]> <![CDATA[Stick cylinder extension and retraction ratio λ2]]> [1.5,1.8] - <![CDATA[x 23 ]]> <![CDATA[Stick cylinder expansion and contraction ratio λ3]]> [1.5,1.8] -

[0060] S1.2, Optimization Objective: This embodiment aims to simultaneously improve the excavator's operating range and digging capacity. To this end, an optimization model comprising three sub-objectives is constructed, and a weighted sum method is used to aggregate them into a single-objective optimization problem. The specific objective function is as follows:

[0061] The objective function f1 for the scope of work is quantified as the tip of the bucket teeth of the excavator (i.e., Figure 3 The area of ​​the envelope of the discrete points (point D) within the workspace under the combined effective strokes of the boom cylinder, stick cylinder, and bucket cylinder is calculated using kinematic analysis, iterating through the extension and retraction of each hydraulic cylinder to obtain a series of point D coordinates. Then, the area of ​​the envelope of these discrete points (Area) is calculated using the convex hull algorithm and the shoelace theorem (shoelace formula). Figure 4 As shown. To transform it into a unified minimization problem and ensure that the objective function value is positive, the objective function is defined as:

[0062]

[0063] In the formula, c1 is the first constant, ensuring that f1 is always positive. In this embodiment, c1 = 500 m 2 ; This indicates that when the design variable takes the value X, the area of ​​the discrete point set envelope of D is... The unit is m 2 .

[0064] Objective function of stick digging force f2: Within the main digging area, considering the theoretical thrust of the active cylinder, the locking capability of the passive cylinder, and the tilting and adhesion constraints of the entire machine, the theoretical digging force F of the entire machine under stick digging conditions is calculated through mechanical analysis. 斗杆 And take its maximum value. Similarly, the objective is transformed into a minimization form:

[0065]

[0066] In the formula, c2 is the second constant to ensure that f2 is always positive. In this embodiment, c2 = 1000 kN. This indicates the theoretical digging force F of the entire machine when the design variable takes the value X. 斗杆 The maximum value is ;

[0067] Objective function for bucket digging force f3: Similar to the boom digging force, the theoretical digging force F of the entire machine under bucket digging conditions is calculated within the main digging area. 铲斗 And take its maximum value. The objective is transformed into a minimization form:

[0068]

[0069] In the formula, c3 is the third constant to ensure that f3 is always positive. In this embodiment, c3 = 1000 kN. This indicates the theoretical digging force F of the entire machine when the design variable takes the value X. 铲斗 The maximum value is ;

[0070] The final single-objective optimization function F(X) is constructed using a weighted sum method:

[0071]

[0072] In the formula, w1 is the weighting coefficient of the objective function of the working range; w2 is the weighting coefficient of the objective function of the boom digging force; w3 is the weighting coefficient of the objective function of the bucket digging force; in this embodiment, for the sake of simplicity, w1=w2=w3=1 is taken.

[0073] S1.3, Constraint Handling: To ensure the engineering feasibility of the design scheme, the optimization model incorporates multi-dimensional constraints such as geometric structure, kinematics, mechanics, excavation performance, and stability. To efficiently handle these strongly coupled constraints, this invention adopts the aforementioned hierarchical constraint handling mechanism and aggregates all constraint violation metrics into a unified constraint violation quantity denoted as CV(X).

[0074] Hard constraint layer: For inviolable physical constraints such as motion interference, singular configurations of mechanisms, and geometric failures, an "early return" strategy is adopted. Once a violation is detected, a death penalty (assigning a very large penalty value) is imposed on the individual, making it priority for elimination in subsequent selections and avoiding invalid calculations.

[0075] Soft constraint layer: For performance indicators that allow for moderate deviations, such as transmission angle range, overall machine stability, and excavation map standards, a penalty function method is used to quantify them and incorporate them into the fitness evaluation. By applying appropriate penalties, effective information in the boundary region between the feasible and infeasible regions is preserved, guiding the algorithm to converge towards the feasible region and avoiding premature abandonment of potential high-quality solutions.

[0076] This hierarchical mechanism establishes clear constraint priorities during the optimization process by differentiating penalty strengths. It can provide strong selection pressure to quickly eliminate severely infeasible solutions while protecting potential boundary solutions, thereby significantly improving search efficiency in complex feasible domains.

[0077] Step S02: Configure and initialize the large model-driven hyperheuristic optimization framework.

[0078] S2.1, Based on the optimization model constructed in step S1, set the core optimization parameters.

[0079] In this embodiment, the dimension of the design variables is 23, and the value range is shown in Table 1; the population size is set to 300, and the maximum number of iterations is 500, that is, the maximum number of function evaluations is set to 500×300=150000.

[0080] S2.2, Initialization of key components of the optimization framework:

[0081] Historical performance knowledge base initialization: A fixed-capacity circular record area is established to store the performance data of each operator in recent iterations. In this embodiment, the capacity of this record area is set to 10, meaning the system always retains historical performance information from the most recent 10 generations. The system follows the "first-in, first-out" principle to continuously update the knowledge base, ensuring that decisions are based on the latest optimization dynamics.

[0082] Large Model Decision Trigger Mechanism Configuration: This section sets the large model invocation strategy. In the initial stage of the optimization process (defined as the first 50 generations in this embodiment), a "warm-up period" is defined, during which the large model is not invoked to conserve computing resources. After the warm-up period, the large model is scheduled to be invoked at fixed intervals. In this embodiment, this interval is set to 10 generations, meaning that a large model intelligent decision is triggered every 10 generations of optimization.

[0083] Construction of the underlying optimization operator library: The initialization includes a variety of optimization operators with different search characteristics. As the underlying operator library of the hyperheuristic framework, the underlying optimization operator library in this embodiment includes five operators: simulated binary crossover operator (denoted as OP1), arithmetic crossover operator (denoted as OP2), polynomial mutation operator (denoted as OP3), DE / rand / 1 difference evolution operator (denoted as OP4), and DE / current-to-rand / 1 difference evolution operator (denoted as OP5).

[0084] Step S03: Execute the large model-driven hyperheuristic optimization loop. The overall process is as follows: Figure 5 As shown. The following sub-steps are executed repeatedly within a preset number of iterations until the termination condition is met (reaching the maximum number of iterations or the convergence criterion), as follows:

[0085] S3.1: Population Initialization and Fitness Evaluation. Within the feasible boundary of the 23-dimensional design variables, an initial population of 300 individuals is randomly generated. For each individual in the population, its corresponding fitness value is calculated. This fitness value comprehensively reflects the individual's (solution's) quality in terms of performance objectives and engineering constraints.

[0086] In this embodiment, the fitness value Fit(X) = objective function value F(X) + constraint violation amount CV(X).

[0087] S3.2: The large-scale intelligent decision-making mechanism triggers the judgment.

[0088] At the beginning of each generation, it is determined whether the conditions for calling the large model are met. The judgment logic is as follows: the current iteration generation is greater than 50 (the warm-up period has passed), and the current generation is an integer multiple of 10 (the call interval is met). If both conditions are met, the large model decision is triggered, and steps S3.3-3.5 are executed to obtain the operator scheduling ratio; otherwise, step S3.6 is executed to obtain the operator scheduling ratio.

[0089] S3.3: Multi-source state information perception and fusion.

[0090] In this embodiment, when a large model decision is triggered, the system collects and fuses state information from four dimensions in real time to form a multi-dimensional state information set:

[0091] (1) Population statistical characteristics: the best, worst, and average fitness values ​​of the current population, as well as the standard deviation of the fitness values. The standard deviation is used to quantify the diversity level of the population.

[0092] (2) Convergence state index: The convergence ratio between 0 and 1 is calculated by the formula (maximum fitness - minimum fitness) / maximum fitness, which is used to characterize the degree of convergence of the population. The smaller the value, the more concentrated the population is and the stronger the convergence.

[0093] (3) Stagnation status information: Records the number of iterations (stagnation generations) since the last improvement of the population's optimal fitness. This value directly reflects the activity of the search process.

[0094] (4) Operator historical performance: Extract the cumulative performance improvement value and the frequency of each operator in the last 10 generations from the historical performance knowledge base, and calculate its average improvement benefit.

[0095] S3.4: Prompt Term Construction and Large Model Invocation. Based on the state information set fused in step S3.3, a structured prompt term is constructed to ensure that the output of the large model meets engineering requirements. In this embodiment, the prompt term template is as follows:

[0096] Role Definition: As a high-level decision-making component of the hyperheuristic algorithm, you are responsible for intelligently selecting the proportion of search operators used for the optimization problem. Your selection should be based on the current population state data, combined with the characteristics of the search operators and their historical performance. Avoid simply distributing operators evenly; instead, select them strategically based on the optimization state.

[0097] Problem background: The problem is a single-objective optimization problem, with the optimization direction being minimization, variable dimension, population size, and current iteration number.

[0098] Operator characteristics: There are 5 optional operators, each with its own characteristics:

[0099] 1. Simulated binary crossover operator (suitable for fine-grained search, more effective when the population converges)

[0100] 2. Arithmetic crossover operator (balances exploration and exploitation, suitable for populations with moderate diversity)

[0101] 3. Polynomial mutation operator (increases diversity, more effective when the population converges prematurely)

[0102] 4. DE / rand / 1 operator (strong global search, suitable for populations with high diversity)

[0103] 5. DE / current-to-rand / 1 operator (local development, suitable for the later stages of population convergence).

[0104] Current population state quantification data: optimal fitness value, average fitness value, fitness standard deviation, convergence ratio, number of generations without improvement.

[0105] Historical performance statistics of operators: performance and number of calls for each operator in the last 10 iterations;

[0106] Guidelines for dynamic strategy generation:

[0107] If there are no more than 10 consecutive improved algebras, increase the use of exploratory operators;

[0108] If the convergence degree is <0.1, increase the use of expansive operators;

[0109] Currently, AI decisions are made every 10 iterations, with a uniform distribution used for the first 50 iterations.

[0110] Output requirements and format specifications:

[0111] Return only an array of 5 ratios, formatted as [0.19, 0.32, 0.11, 0.21, 0.17], containing 5 numbers with a sum of 1. Each ratio must be between 0.05 and 0.8. Prioritize assigning higher ratios to operators with higher average improvement values, and balance exploration and development. Do not return any explanations or additional text.

[0112] In this embodiment, the DeepSeek API is used to send prompts filled with real-time data to the DeepSeek large model, which then generates optimization strategy guidance for the next stage. The large model called in step S3 can be a large-scale natural language processing model deployed in the cloud or locally, including but not limited to well-known mainstream large models at home and abroad such as DeepSeek, Qwen, Kimi, Llama, Doubao, Hunyuan, and Zhipu Qingyan, and their interaction is implemented through API.

[0113] S3.5: Large Model Strategy Analysis and Processing.

[0114] S3.5.1, when the API call is successful, receives and parses the text response from the large model, extracting a numerical array of the scheduling ratios of the five operators. Then, it performs validity checks and standardization: first, it ensures that each ratio value falls within a preset reasonable range [0.05, 0.80] to prevent monopolization by a single operator; second, it normalizes the ratio values ​​of all operators to ensure the sum is 1, ultimately obtaining a valid dynamic scheduling ratio.

[0115] In S3.5.2, when a large model API call fails, a response times out, or the returned content cannot be parsed, a robust guarantee mechanism is adopted, and the system automatically switches to a local heuristic backup strategy. This strategy uses a dynamic weighted formula to integrate the recent average improvement gain of the operator, the population convergence state, and the number of stalled generations to generate the operator's scheduling ratio Ra in real time, ensuring uninterrupted optimization. The specific calculation formula is as follows:

[0116]

[0117] In the formula Ra i Pe represents the scheduling ratio of operator i; i Let Ex be the recent average improvement gain of operator i; i Assigning a dynamic scaling factor to operator i based on the population convergence state; St i To assign a dynamic scaling factor to operator i based on the population stagnation generation; w p w e w s These are the weight coefficients for the corresponding items. In this embodiment, we take w. p =w e =0.4, w s =0.2.

[0118] S3.6: Strategy Inheritance Mechanism. If the large model decision is not triggered, the operator scheduling ratio is determined according to the following rules: During the warm-up period (first 50 generations), the usage ratio of the five operators is evenly distributed, that is, each operator accounts for 0.2; after the warm-up period but not in the call interval generation, the scheduling ratio generated by the previous large model decision is directly used.

[0119] S3.7: Offspring Generation and Population Update. First, based on the scheduling ratios of the operators obtained in S3.5, S3.6, or S3.7, calculate the number of offspring individuals that each operator should generate. Then, create a random permutation sequence with a total length equal to the population size, and assign operators to individuals in the population according to this sequence to generate offspring. After calculating the offspring fitness, a greedy selection strategy is used to update the population: for each parent-offspring pair, the offspring replaces the parent only if the offspring's fitness is better than the parent's, ensuring that the overall quality of the population does not monotonically decrease. Simultaneously, record the average of all improvements generated by each operator in this generation as its performance gain for this iteration.

[0120] S3.8: Update the historical performance knowledge base. Store the operator performance data for the current iteration recorded in step S3.7 into the historical performance knowledge base. If the base is full (capacity 10), follow the "first-in, first-out" principle to overwrite the oldest data.

[0121] S3.9: After completing the population update for each iteration, determine whether the maximum number of iterations or the convergence criterion (the optimal solution shows no significant improvement for 20 consecutive generations) has been reached. If the termination condition is not met, return to step S3.1 for the next generation of optimization.

[0122] Step S04: If the termination condition is met, output the historical best individual recorded throughout the entire optimization process. The 23 design variable parameters corresponding to it constitute the final optimization scheme of the multi-ring coupled mining mechanism.

[0123] To verify the effectiveness and comprehensive performance of the large language model-inspired multi-ring coupled mining mechanism intelligent optimization method (LLM-HHO) proposed in this invention, a systematic comparative experiment was conducted. Four mainstream intelligent optimization algorithms widely used in engineering optimization were selected as comparison benchmarks: Genetic Algorithm (GA), Particle Swarm Optimization (PSO), Differential Evolution (DE), and Covariance Matrix Adaptive Evolutionary Strategy (CMA-ES). All comparison algorithms were run under a unified experimental environment with the same population size (300) and maximum number of iterations (500), and each was run independently 10 times to eliminate the influence of random factors, thus ensuring the fairness and reliability of the comparison results. The optimization performance statistics of each algorithm are shown in Table 2, and the convergence process is compared as follows: Figure 6 As shown.

[0124] Table 2 Comparison of optimization performance of each algorithm

[0125] algorithm Success rate optimal value Mean (Standard Deviation) Runtime (seconds) LLM-HHO 100% 1363.3 1497.4（89.46） 8796.6(764.69) GA 40% 1487.8 1532.9（51.37） 11839.5(1503.38) PSO 0% - - - DE 0% - - - CMAES 0% - - -

[0126] As shown in Table 2, the LLM-HHO method proposed in this invention significantly outperforms traditional optimization algorithms in all key performance indicators. Specifically, LLM-HHO successfully converged to a feasible solution in all 10 independent runs, achieving a success rate of 100%, demonstrating excellent algorithm stability and robustness. In contrast, GA, the best-performing algorithm in the comparison group, only achieved a success rate of 40%, while PSO, DE, and CMA-ES completely failed on this complex high-dimensional optimization problem with strong constraints, failing to obtain any feasible solutions, further highlighting the limitations of mainstream intelligent algorithms in handling such problems. Regarding the quality of solutions, the optimal objective function value (1363.3) and average objective function value (1497.4) obtained by LLM-HHO are significantly better than those of the GA algorithm, demonstrating its superior global search capability and convergence accuracy. Of particular note is that although LLM-HHO requires calling external large model APIs during the optimization process, its average running time (8796.6 seconds) is still significantly lower than that of the GA algorithm (11839.5 seconds). This result demonstrates that by introducing an intelligent decision-making mechanism based on a large model, the present invention effectively guides the search process to focus on high-potential areas, significantly reducing redundant computation and ineffective exploration, thereby achieving a substantial improvement in convergence speed and computational efficiency over "blind search" through "intelligent guidance".

[0127] Figure 6 The convergence curves shown further reveal the dynamic characteristics of each algorithm during the search process. LLM-HHO exhibits a rapid decline in the early stages of optimization, effectively avoiding local optima. In the later stages of optimization, its convergence curve continues to decline smoothly, demonstrating strong continuous optimization capabilities and stable convergence characteristics. In contrast, mainstream intelligent optimization algorithms such as GA show significant stagnation in their convergence curves early on, becoming trapped in local optima and struggling to escape, ultimately limiting their optimization effectiveness.

[0128] The superior performance of LLM-HHO stems from its core adaptive mechanism. For example... Figure 7 As shown, throughout the optimization process, the large model can dynamically adjust the invocation ratio of underlying operators based on the real-time search status, achieving an intelligent balance between exploration and exploitation. The strategy flexibility and decision-making accuracy demonstrated by this mechanism far surpass those of a single algorithm or an adaptive strategy based on fixed rules, thus ensuring the global convergence and search efficiency of the algorithm in complex high-dimensional spaces.

[0129] In summary, the comparative experiments fully validated the effectiveness and advancement of the proposed method across multiple dimensions, including success rate, solution quality, convergence speed, and operational efficiency. For the typical high-dimensional, nonlinear, and strongly constrained engineering optimization problem of multi-ring coupled excavation mechanisms, LLM-HHO not only significantly outperforms mainstream intelligent optimization algorithms in solution stability and final solution quality, but also demonstrates a clear advantage in overall optimization efficiency. This embodiment successfully provides a set of mechanism parameter design schemes with excellent comprehensive performance that fully meets all engineering constraints for this 70-ton hydraulic excavator, fully demonstrating the technological advancement and wide application potential of this invention in solving complex engineering optimization problems.

[0130] It should be understood that although this specification is described according to various embodiments, not every embodiment contains only one independent technical solution. This way of describing the specification is only for clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

[0131] The detailed descriptions listed above are merely specific illustrations of feasible embodiments of the present invention and are not intended to limit the scope of protection of the present invention. All equivalent embodiments or modifications made without departing from the spirit of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent optimization of a multi-ring coupled mining mechanism inspired by a large language model, characterized in that, Includes the following steps: Step S01: Construct an optimization model for the multi-ring coupled excavation mechanism, including determining design variables, optimization objectives, and constraints; Step S02: Configure and initialize the hyperheuristic optimization framework, including setting optimization parameters, initializing the historical performance knowledge base, configuring the large model decision triggering mechanism, and building the underlying optimization operator library; Step S03: Execute a large model-driven hyperheuristic optimization loop, including the following steps: Step S3.1: Population initialization and fitness assessment; Step S3.2: Determine whether the large model intelligent decision-making triggering conditions are met. The decision-making triggering conditions are usually determined by the initial warm-up period and the fixed call cycle. Step S3.3: When the triggering condition is met, collect and fuse population statistical characteristics, convergence state indicators, stagnation state information and historical performance data of the underlying operators in real time to form a comprehensive multidimensional state information set; Step S3.4: Construct prompt words and call the large language model to construct structured prompt words, and generate the optimization guidance strategy for the next stage by calling the large model; Step S3.5: Analyze and process the guidance strategy for generating the large model or obtain the operator scheduling ratio through a robust guarantee mechanism; Step S3.6: When the triggering conditions for intelligent decision-making in large models are not met, the operator scheduling ratio is obtained through the policy inheritance mechanism; Step S3.7: Generate offspring and update the population according to the operator scheduling ratio; Step S3.8: Update the historical performance knowledge base; Step S3.9: Determine whether the termination condition is met. If not, return to step S3.

1. Step S04: Output the optimal design scheme.

2. The intelligent optimization method for multi-ring coupled mining mechanisms inspired by large language models according to claim 1, characterized in that, The optimization objectives in step S01 include the working range objective, the boom digging force objective, and the bucket digging force objective. The working range objective is defined as the envelope area of ​​a discrete set of points within the reachable working area of ​​the excavating mechanism.

3. The intelligent optimization method for multi-ring coupled mining mechanisms inspired by large language models according to claim 1, characterized in that, The constraints in step S01 employ a hierarchical constraint processing mechanism, including: For inviolable hard constraints, adopt an early return strategy and impose a death penalty; The penalty function method is used to handle soft constraints that are violated to a certain degree.

4. The intelligent optimization method for multi-ring coupled mining mechanisms inspired by large language models according to claim 1, characterized in that, The underlying optimization operator library in step S02 includes a variety of optimization operators with complementary search characteristics, including at least two of the following: simulated binary crossover operator, arithmetic crossover operator, polynomial mutation operator, DE / rand / 1 differential evolution operator, and DE / current-to-rand / 1 differential evolution operator.

5. The intelligent optimization method for multi-ring coupled mining mechanisms inspired by large language models according to claim 1, characterized in that, The triggering conditions for large-scale intelligent decision-making in step S3.2 include: triggering according to a preset periodic interval after the initial warm-up period of the optimization process.

6. The intelligent optimization method for multi-ring coupled mining mechanisms inspired by large language models according to claim 1, characterized in that, The prompts in step S3.4 are preset standardized text templates, which include: large model role definition, optimization problem description, operator characteristic description, current optimization state quantification data, operator historical performance statistics, dynamic strategy generation guidance, and output format specifications.

7. The intelligent optimization method for multi-ring coupled mining mechanisms inspired by large language models according to claim 1, characterized in that, Step S3.5 includes: When the large model is successfully invoked, the guidance strategy generated by the large model is parsed and processed to obtain the operator scheduling ratio, including parsing the return result of the large model, performing legality checks and corrections, and normalizing the obtained operator scheduling ratio. When a large model call fails or an abnormal response occurs, the operator scheduling ratio is obtained through a robust guarantee mechanism, specifically: The scheduling ratio Ra of the operator is generated in real time using a weighted fusion formula. The specific calculation formula is as follows: ， In the formula Ra i Pe represents the scheduling ratio of operator i; i Let Ex be the recent average improvement gain of operator i; i Assigning a dynamic scaling factor to operator i based on the population convergence state; St i To assign a dynamic scaling factor to operator i based on the population stagnation generation; w p w e w s These are the weight coefficients for the corresponding items.

8. The intelligent optimization method for multi-ring coupled mining mechanisms inspired by large language models according to claim 1, characterized in that, The strategy inheritance mechanism in step S3.6 includes: adopting a uniform allocation strategy during the warm-up period; and inheriting the operator scheduling ratio generated by the previous large model decision after the warm-up period but not during the call interval.

9. The intelligent optimization method for multi-ring coupled mining mechanisms inspired by large language models according to claim 1, characterized in that, The offspring generation in step S3.7 includes: calculating the number of offspring to be generated by each operator according to the dynamic scheduling ratio, generating a randomized sequence of operator indices of the corresponding length, allocating the corresponding operators to the parent individuals in the population according to this sequence to generate offspring individuals, updating the population through an environment selection mechanism, and recording the iteration performance of each operator; the environment selection adopts a greedy strategy, and the offspring individual will only be retained if its fitness is better than that of its parent.

Images