CN-121809290-B - Heterogeneous fuzzy and game-based double-layer self-adaptive evolution CAD modeling command generation method

CN121809290BCN 121809290 BCN121809290 BCN 121809290BCN-121809290-B

Abstract

A double-layer self-adaptive evolution CAD modeling command generation method based on heterogeneous fuzzy and game belongs to the technical field of artificial intelligence and computer aided design intelligent modeling. The method solves the technical problems of difficult quantitative mapping of fuzzy semantics, poor adaptability of heterogeneous commands, low executable performance of the generated commands and black box in the reasoning process in the existing natural language driven CAD command generation technology. The method constructs a semantic analysis-game generation-evolution refinement-verification backtracking-knowledge precipitation full-flow closed-loop architecture, can be applied to the fields of industrial software, three-dimensional modeling, intelligent design systems and the like, improves engineering executability, generalization capability and reasoning interpretability of command generation, balances generation diversity and feasibility, reduces model iteration cost, and provides technical support for the industrial landing of text-driven CAD modeling technology.

Inventors

WANG XIANGXIANG
LIU JUN
YU YONGBIN
Fan Manping
LI CHENBO
QIN SHENGYU
Xue Kaiyi
WANG JINGYA
Han Xindie

Assignees

电子科技大学

Dates

Publication Date: 20260508
Application Date: 20260309

Claims (8)

1. A method for generating a double-layer self-adaptive evolution CAD modeling command based on heterogeneous ambiguity and game is characterized by comprising the following steps: Step 1, analyzing and encoding an input natural language instruction to obtain a precise semantic vector fused with fuzzy semantic quantization features; Step 2, constructing a heterogeneous operator library which is suitable for the generation requirement of the CAD heterogeneous command, dynamically matching heterogeneous operator combinations based on accurate semantic vectors to generate candidate command sequences, constructing a non-zero and game framework of a generator and a verifier of a CAD command generation scene, and realizing the balance of semantic fitness and geometric constraint through double-main-body collaborative game verification; Step 3, executing structure self-adaptive evolution and fuzzy parameter refinement on the candidate command sequence after game verification to obtain an optimized command sequence; step 4, performing multidimensional executable verification on the optimized command sequence, and outputting a qualified CAD command sequence after backtracking and correcting the substandard sequence; Step 5, the system carries out self-adaptive updating, namely, incremental learning and knowledge precipitation are carried out based on the high confidence output result, and continuous optimization of model performance and multiplexing of historical knowledge are realized; the step 5 specifically comprises the following steps: Step 5.1, constructing an incremental distillation unit, adopting KL divergence quantization and minimizing commands of new and old models to generate probability distribution differences, and carrying out incremental distillation optimization by combining cross entropy loss of new tasks; step 5.2, caching semantic-command-geometric triples as high confidence samples for incremental learning, wherein the semantic-command-geometric triples comprise accurate semantic vectors, optimal command sequences after game optimization and parameter refinement, and geometric constraint graphs corresponding to the command sequences; For the entities in the semantic-command-geometry triples, calculating to obtain comprehensive matching degree based on semantic feature vectors and entity association strength between new tasks and historical task entities, and rapidly positioning historical experience matched with the new tasks; step 5.3, constructing a task self-adaptive structure updating mechanism, and introducing a parameter updating formula of the structure self-adaptive evolution: ; Wherein, the For the configuration parameters of the self-adaptive evolution layer of the t-moment structure, In order to achieve a structural learning rate, Representing parameters Is used for the gradient operator of (1), To evolve the penalty, an adaptation error of the current operator combination to the task requirements is calculated, As the weight coefficient of the entropy, Adjusting the number of operators and updating parameters according to the task complexity when performing dynamic structure reconstruction for the operator combination diversity entropy; The step 1 comprises a fuzzy attention and CAD knowledge coding framework, wherein the fuzzy attention and CAD knowledge coding framework has 6 layers, and a single-layer expression is as follows: ; Wherein, the In order to be a preliminary semantic vector, In order to pre-treat the triplet, Representing a transpose; Performing residual connection operation; normalizing the layers; Respectively the first Query, key, value linear transformation matrix of the individual attention heads, Fusing a linear transformation matrix for multi-head output; the output of the 8 attention heads is shown spliced, In order for the scaling factor to be a factor, Is that A function.
2. The method for generating the double-layer self-adaptive evolution CAD modeling command based on the heterogeneous ambiguity and the game according to claim 1, wherein the step1 is specifically as follows: Step 1.1, performing word segmentation, part-of-speech tagging and syntactic dependency analysis on natural language instructions, extracting keywords based on a semantic dictionary in the CAD field, and mapping texts to action-object-parameter triples; step 1.2, preprocessing the triples through format regulation and semantic complementation to provide standardized input; Step 1.3, fuzzy semantic disambiguation and coding are carried out, namely firstly, semantic types are matched, fuzzy quantization vectors are calculated through Gaussian membership functions, meanwhile, fuzzy attention of a set number of layers is associated with the upper part and the lower part Wen Yuyi of the triples after the CAD knowledge coding framework is captured and preprocessed, preliminary semantic vectors are obtained, and then the preliminary semantic vectors and the fuzzy quantization vectors are fused, so that accurate semantic vectors are obtained.
3. The method for generating the double-layer adaptive evolution CAD modeling command based on heterogeneous ambiguity and game according to claim 2, wherein the step 2 is specifically as follows: and 2.1, calculating operator selection probability through an operator fitness function regulated by temperature coefficients, dynamically matching operator combinations in the heterogeneous operator library based on accurate semantic vectors, and executing sequence decoding to generate candidate command sequences for CAD heterogeneous command structure adaptation, wherein the formulas are as follows: ; Wherein, the As a result of the candidate command sequence, Is the first in the heterogeneous operator library The number of operators to be used in the process, In order for the decoder to be a decoder, The probability is selected for the operator and, Is a precise semantic vector; Step 2.2, constructing a non-zero and game framework of a generator and a verifier, wherein step 2.1 is used as the generator, the verifier comprises a semantic verification head for semantic consistency verification and a geometric verification head for geometric topological constraint verification, then game benefits are evaluated and iterative optimization is carried out through a combined benefit function of the verifier, and strategy iteration is carried out until Nash equilibrium judging conditions are met, so that strategy equilibrium of the generator and the verifier is achieved; And 2.3, correcting geometric parameters of the candidate command sequences based on the deviation of geometric topology constraint verification, realizing constraint adaptation through heterogeneous operator parameter reconfiguration, and finally outputting the optimized candidate command sequences.
4. The method for generating the double-layer adaptive evolution CAD modeling command based on heterogeneous ambiguity and game according to claim 3, wherein the step 3 is specifically as follows: step 3.1, an initial operator combination population, a task complexity evaluation function is calculated based on the number of modeling objects and the number of constraint relations of a candidate command sequence, an evolution direction is locked based on task complexity, and screening, crossing and mutation optimization are performed on operator combinations through a genetic optimization unit, so that a heterogeneous reasoning structure is obtained; Step 3.2, calling a Gaussian membership function to execute continuous domain correction on the game generated parameters based on a fuzzy rule base driven parameter finishing mechanism, and obtaining accurate parameters through a multi-rule output aggregation function; And 3.3, fusing the outputs of the step 3.1 and the step 3.2 based on a gating mechanism to obtain a final command sequence.
5. The method for generating the double-layer adaptive evolution CAD modeling command based on heterogeneous ambiguity and game according to claim 4, wherein the step 4 is specifically as follows: Comprehensively scoring the command sequence by using geometric constraint consistency based on geometric topology constraint verification, logic rationality based on command dependency graph and parameter feasibility based on parameter engineering feasibility verification; and backtracking and correcting the command sequences with the comprehensive scores not reaching the standard, and outputting qualified CAD command sequences.
6. The method for generating the double-layer self-adaptive evolution CAD modeling command based on heterogeneous ambiguity and game according to claim 5, wherein the confidence level of each type of ambiguity is calculated based on the accurate semantic vector E, and the calculation formula is as follows: ; Wherein the method comprises the steps of As a weight vector for the semantic tags, For the semantic-geometric association coefficient, the confidence coefficient is calculated The dynamic weight coefficient is directly acted on the synthesis process of the fuzzy quantization vector, and the confidence coefficient is transmitted to the subsequent fuzzy logic parameter finishing step through the data stream and is used as the adjusting factor of the parameter correction force.
7. The method for generating the dual-layer self-adaptive evolution CAD modeling command based on heterogeneous ambiguity and game according to claim 6, wherein the formula of the combined benefit function of the checking party is as follows: ; Wherein, the Representing the combined benefit of the two subjects, Representing candidate command sequences generated by the generator, semantic weights Geometric weight Weight of efficiency As dynamic weight, satisfy ; The semantic fitness is obtained by calculating cosine similarity between the candidate command sequence and the feature vector; for executing efficiency gain, calculating the ratio of the actual execution time of the candidate command sequence C to the actual execution time of the similar task; for geometric constraint consistency, the calculation formula is as follows: ; Wherein, the Is a geometric constraint map of the candidate command sequence C, A geometric constraint map is designed for the target, For the feature extraction function of the graph, Representing vectors Norms.
8. The method for generating the double-layer self-adaptive evolution CAD modeling command based on heterogeneous ambiguity and game according to claim 7, wherein in step 2.2, iterative optimization aims at joint benefit maximization, the generator selects probability through a gradient descent adjustment operator, and the verifier updates dynamic weight through weight normalization.

Description

Heterogeneous fuzzy and game-based double-layer self-adaptive evolution CAD modeling command generation method Technical Field The invention belongs to the technical field of artificial intelligence and computer aided design intelligent modeling, and particularly relates to a double-layer self-adaptive evolution CAD modeling command generation method based on heterogeneous fuzzy and game. The method integrates a heterogeneous operator network, a fuzzy reasoning mechanism and a game optimization framework, realizes intelligent generation from natural language instructions to executable CAD modeling command sequences, and is suitable for the fields of industrial software, three-dimensional modeling, intelligent design systems and the like. Background With the development of artificial intelligence and natural language processing technology, text-driven CAD command generation becomes a key direction of industrial software intelligence. Through natural language description design intent and automatic conversion to CAD operation command, can reduce the use threshold obviously, promote modeling efficiency, provide the support for the quick design in multiple fields. However, the process breaks through three core bottlenecks of natural language ambiguity, CAD command isomerism and generation and constraint gaming, and the prior art lacks a systematic solution to the three core bottlenecks, so that a plurality of core defects are caused. The rule matching and template mapping method relies on manual definition semantic modes, has poor generalization capability, is difficult to adapt to complex sentences and combination operation, has high maintenance cost and insufficient expansibility, and is more critical that the method cannot analyze engineering semantics of fuzzy expressions such as rough expression, approaching expression and the like, is more difficult to adapt to heterogeneous structures of discrete operation symbols and continuous geometric parameters in CAD commands, and the manual template cannot cover all heterogeneous command combination scenes. Compared with a rule matching method, the sequence generation model based on the Seq2Seq or the Transformer can capture semantic context to a certain extent, but has obvious limitation, lacks geometric constraint consciousness, generates commands often to have problems of reasonable semantics but inexecutable, has logic defects of inconsistent parameters, incorrect operation sequence and the like, has insufficient modeling capability on heterogeneous mixed structures of discrete symbols and continuous parameters in CAD commands, lacks a mapping mechanism from fuzzy semantics to accurate geometric parameters, and cannot convert ambiguous space relations such as coincidence, alignment and the like into accurate constraint, so that logic hidden hazards are buried for subsequent optimization. While the conventional generation countermeasure network is tried to enhance the rationality of the generation result, the structure is mainly designed for image generation, still faces outstanding challenges in the field of discrete command sequence generation, and the essence of easy sinking mode collapse is that a generation-verification double-main-body cooperative game framework is not constructed, only the surface layer distribution fitting is realized through countermeasure, a discriminator can only measure the surface layer distribution similarity of the command sequence, the geometric topological constraint and the semantic logic consistency cannot be comprehensively evaluated from the game income angle, and fuzzy semantic resolution precision and heterogeneous structure suitability are not included in the game income function, so that unbalance of the countermeasure process is caused. More deeply, engineering logic and parameter constraint behind a command layer are generally ignored in the prior art, the essence is the lack of collaborative modeling for fuzzy semantics, heterogeneous structures and game optimization, the fact is that a quantitative analysis mechanism aiming at fuzzy relations is lacking, conversion from semantic ambiguity to parameter accuracy cannot be achieved, a heterogeneous operator dynamic combination architecture is lacking, command structure requirements in different design stages are difficult to adapt, a multi-target collaborative game mechanism is not established, balance between diversity generation and executable performance cannot be achieved, and inference process blackout results in unexplainable and cannot meet industrial audit requirements. To improve the above problems, some studies have attempted to introduce generation of an countermeasure network to enhance the rationality of the generated result, and to evaluate the authenticity or operability of the generated command by a arbiter. However, conventional GAN structures are designed primarily for image generation tasks, and still face many challenges in the field of discrete sym