A deterministic constraint solving method and system under candidate pool local dynamic activation

Abstract

The application discloses a kind of candidate pool local dynamic activation determination constraint solving method and system, belong to mechanical intelligent design technical field, including: obtaining user unstructured mechanical design demand and converting into structured demand state;Construct the candidate pool of to-be-clarified field, rely on type relationship and two-step forward-looking value to select to-be-clarified field and output clarification question;To user response carries out source record, type check and hard constraint pre-check, obtains the structured demand state of completion verification;According to the initial candidate component pool of scale controlled of part function role generation, complete candidate pool local dynamic activation;Activated candidate component is input into deterministic constraint solver uniform check, and circulates until feasible assembly configuration or trigger termination condition is obtained Iteration;The application realizes candidate pool local update, reduces invalid calculation, improves mechanical assembly scheme generation efficiency and constraint solving reliability.

Description

Technical Field

[0001] This invention belongs to the field of mechanical intelligent design technology, specifically a deterministic constraint solution method and system under local dynamic activation of candidate pool. Background Technology

[0002] In intelligent design scenarios for mechanical products, users often submit unstructured design requirements in natural language, with ambiguous information and hidden constraints. Traditional solutions, which directly match components, are prone to misunderstandings of the requirements. Most existing technologies use a global component candidate pool for constraint solving, resulting in a massive candidate set. Each time a constraint conflict occurs, all component candidates need to be re-searched and filtered, generating a large amount of unnecessary computation and high system overhead. Furthermore, the lack of a precise location mechanism after a conflict prevents the focus on the functional role of the conflicting component to supplement candidate components; instead, the entire candidate pool needs to be refreshed, leading to slow iterative convergence. In addition, conventional design systems lack standardized requirement clarification mechanisms, cannot quantitatively assess the priority of fields to be clarified, and struggle to select key items for clarification based on field influence and information gain. This easily leads to overlooking key design parameters or repeated interactions with low-value fields, resulting in low human-computer interaction efficiency. Some solutions rely solely on large models to directly generate assembly schemes, lacking deterministic constraint solvers for rule-based verification. This leads to issues such as interface incompatibility, parameter out-of-bounds errors, and inconsistent assembly coupling, making it difficult to guarantee the feasibility of the solutions. Conversely, some constraint solving systems lack local dynamic candidate activation capabilities and employ coarse-grained methods to handle conflicts, failing to balance solution efficiency with output quality. Therefore, there is an urgent need for a mechanical design constraint solving method capable of structural transformation of requirements, precise requirement clarification, local dynamic activation of the candidate pool, and targeted conflict resolution. This method would reduce computational redundancy, accelerate iterative convergence, and improve the reliability of generated intelligent mechanical design schemes. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention proposes a deterministic constraint solving method and system under local dynamic activation of the candidate pool. The method involves: acquiring unstructured mechanical design requirements from users and converting them into structured requirement states; constructing a candidate pool of fields to be clarified, selecting fields based on type relationships and two-step look-ahead value, and outputting clarification questions; recording the source of user responses, verifying their types, and pre-verifying hard constraints to obtain a verified structured requirement state; generating an initial candidate component pool with controlled scale according to the functional roles of components, thus completing the local dynamic activation of the candidate pool; inputting the activated candidate components into a deterministic constraint solver for unified verification, iterating cyclically until a feasible assembly configuration is obtained or a termination condition is triggered; this invention achieves local updates of the candidate pool, reduces invalid calculations, and improves the efficiency of mechanical assembly scheme generation and the reliability of constraint solving.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] A method for solving deterministic constraints under local dynamic activation of a candidate pool, comprising:

[0006] S1: Obtain unstructured mechanical design requirements from user input and convert them into structured requirement states that include field types, importance, acquisition costs, status flags, and constraint relationships;

[0007] S2: Construct a candidate pool of fields to be clarified based on the structured requirement status, select fields to be clarified from the candidate pool using type relationships and two-step look-ahead value, and output the clarification question;

[0008] S3: Receive the response content returned by the user for the clarification question, perform source record, type validation and hard constraint pre-validation on the response content, and generate a structured requirement state that has completed the validation;

[0009] S4: Based on the structured requirement status that has been verified, generate an initial candidate component pool with controlled size according to the functional roles of the components, and call the language model to output the component candidate number. Then, perform whitelist verification, functional role matching, usage scenario adaptation and duplicate item removal on the component candidate number to complete the local dynamic activation of the candidate pool.

[0010] S5: Input the candidate parts obtained after local dynamic activation into the deterministic constraint solver for unified verification; if the verification passes, output the feasible assembly configuration; if the verification fails, parse the conflict information output by the deterministic constraint solver, locate the functional role of the parts corresponding to the conflict, and call the language model to supplement the candidate parts only for the functional role of the parts corresponding to the conflict.

[0011] S6: Perform loop iterations until a feasible assembly configuration is output or the preset candidate supplement termination condition is met.

[0012] Specifically, the steps of S1 include:

[0013] Unstructured mechanical design requirements text input by users is obtained based on a natural language processing interface;

[0014] Based on the unstructured mechanical design requirements, a natural language processing word segmentation algorithm is used to obtain the required word sequence;

[0015] Based on the demand word sequence, entity words are extracted using a named entity recognition algorithm to generate an initial entity set;

[0016] Based on the initial entity set, a rule-based matching method is used to map each entity word to a predefined field type, generating field type annotation results;

[0017] Based on the field type labeling results, the importance of each field is calculated using the analytic hierarchy process (AHP).

[0018] Based on importance and combined with pre-set field collection time thresholds, a cost evaluation function is used to calculate the acquisition cost;

[0019] Based on the acquisition cost, a status flag is set according to the range of the cost value; the status flags include unprocessed, pending clarification, verified, and locked.

[0020] Based on state labeling, the dependency relationships between fields are analyzed using association rule mining algorithms to establish constraint relationships and generate structured requirement states.

[0021] Specifically, the steps of S2 include:

[0022] Based on the structured requirement status, fields marked as unprocessed or pending clarification are selected, and an initial candidate pool of fields pending clarification is constructed.

[0023] For each candidate field in the initial candidate pool of fields to be clarified, the number of fields that directly and indirectly constrain the candidate field in the mechanical design constraint network is calculated and the first influence score is generated.

[0024] Based on the field type and importance, complete cases similar to the candidate fields are retrieved from the historical design case library. The value distribution of the corresponding candidate fields in the similar complete cases is statistically analyzed and the expected reduction in information entropy is calculated to generate a second information gain score.

[0025] The first influence score and the second information gain score are weighted and summed to calculate the two-step forward value of each candidate field. The candidate field with the largest two-step forward value is selected as the field to be clarified.

[0026] Based on the field type of the field to be clarified and its constraint relationship, the corresponding question sentence is matched from the preset clarification question template library, the clarification question is generated and output.

[0027] Specifically, the steps of S3 include:

[0028] Receive the response content returned by the user through the user interaction interface, and record the data source type of the response content;

[0029] Based on the field type of the field to be clarified, the corresponding type validator is invoked to perform type validation on the response content. If the type validation fails, an error message is generated and the user is requested to re-enter the information. If the type validation passes, based on the constraint relationship, the associated fields with a direct hard dependency relationship with the field to be clarified and their hard constraint conditions are extracted, and hard constraint pre-validation is performed on the response content. If the hard constraint pre-validation passes, the response content is filled into the structured requirement status as a validated value, and the status mark of the corresponding field is updated to validated. If the hard constraint pre-validation fails, a conflict report is generated and returned to the user.

[0030] After all fields to be clarified have passed type validation and hard constraint pre-validation, the structured requirement state that has been validated is generated.

[0031] Specifically, the process of generating an initial candidate component pool of controlled size based on the structured requirement status after verification and according to the functional roles of the components includes:

[0032] Parse the structured requirement status that has been verified, and extract all verified fields and their verified values;

[0033] Based on the preset mechanical function role classification tree, each verified field is matched with its corresponding component function role to generate a function role requirement mapping table.

[0034] For each functional role in the functional role requirement mapping table, retrieve components that meet the technical parameter range of the corresponding functional role from the material master database of the enterprise resource planning system to form a candidate component set.

[0035] A preset upper limit threshold for the size of the candidate pool is obtained. When the number of parts in any candidate part set exceeds the upper limit threshold for the size of the candidate pool, a maximum-minimum distance sampling algorithm based on space filling design is used to uniformly extract a subset of the candidate part set so that the size of the extracted candidate part set does not exceed the upper limit threshold for the size of the candidate pool, thereby generating an initial candidate part pool with controlled size.

[0036] Specifically, the step of calling the language model to output component candidate numbers, and then performing whitelist verification, functional role matching, usage scenario adaptation, and duplicate item removal on the component candidate numbers to complete the local dynamic activation of the candidate pool, includes:

[0037] The technical parameters of each component in the initial candidate component pool are encapsulated into a sequence of prompt words. The pre-trained language model interface is called to obtain the list of component candidate numbers corresponding to each functional role output by the language model.

[0038] Based on a preset whitelist, remove numbers from the candidate part number list that are not in the whitelist to generate a first intermediate list;

[0039] Based on the functional role requirement mapping table, retain the numbers of functional roles in the first intermediate list that match the verified fields, and generate a second intermediate list;

[0040] Based on the working environment parameters in the structured requirement status that has been verified, the numbers that do not meet the working environment adaptability standards in the second intermediate list are removed, and a third intermediate list is generated.

[0041] The third intermediate list is deduplicated, and the final retained numbers are used as the set of local candidate component numbers after activation.

[0042] Specifically, the step of inputting the candidate components obtained through local dynamic activation into the deterministic constraint solver for unified verification includes:

[0043] Obtain the activated set of local candidate component numbers, and based on the material master database, extract the detailed technical parameter table corresponding to each number in the activated set of local candidate component numbers;

[0044] The detailed technical parameter table, all verified fields in the structured requirement state that has been verified, and the constraint relationships are jointly encoded into the input fact set of the deterministic constraint solver.

[0045] The deterministic constraint solver is started; the deterministic constraint solver has a set of preset verification rules, including numerical boundary verification rules based on interval arithmetic, interface type compatibility verification rules based on the interface standard library, and cross-component coupling consistency verification rules based on the assembly connection diagram; the deterministic constraint solver performs numerical boundary verification, interface type compatibility verification, and cross-component coupling consistency verification in sequence based on the input fact set, and outputs a unified verification result.

[0046] Specifically, if the verification passes, a feasible assembly configuration is output; if the verification fails, the conflict information output by the deterministic constraint solver is parsed, and the functional roles of the components corresponding to the conflict are located, including:

[0047] When the unified verification result indicates that all the verification rules are met, the verification is deemed to have passed; based on the activated local candidate component number set, the 3D model file and bill of materials file corresponding to each number are extracted from the product data management system; the 3D model file, bill of materials file and the activated component number set are associated and packaged, and all verified fields in the structured requirement state that has been verified and the constraint association relationship are attached as metadata to the packaged file to generate a feasible assembly configuration, and the feasible assembly configuration is output to the computer-aided design system;

[0048] When the unified verification result indicates that there are unmet verification rules, the verification is determined to fail and the conflict information output by the deterministic constraint solver is obtained. The conflict information includes a conflict type identifier, the part number involved in the conflict, and the expression of the violated constraint condition. Based on the part number involved in the conflict, the functional role requirement mapping table is queried in reverse to determine the functional role of the part corresponding to the part number involved in the conflict and mark it as a conflict functional role.

[0049] Specifically, the step of calling the language model to supplement component candidates only for the functional roles of the components corresponding to the conflict includes:

[0050] Based on the conflict type identifier and the violated constraint expression, supplementary search keywords are generated for the conflicting functional roles;

[0051] Based on the supplementary search keywords, a secondary search is performed from the material master database to obtain a set of supplementary candidate parts for the conflicting functional roles;

[0052] The technical parameters of each component in the supplementary candidate component set are encapsulated into a new prompt word sequence, and the pre-trained language model interface is called again to obtain the supplementary component candidate number list output by the language model;

[0053] The components in the supplementary component candidate number list are used as newly activated local candidate components, replacing or adding them to the activated local candidate component number set to form an updated local candidate component number set.

[0054] Specifically, the steps of S6 include:

[0055] Using the updated set of local candidate component numbers as new input, the step of performing unified verification on the candidate component input deterministic constraint solver obtained through local dynamic activation is repeated.

[0056] At the start of each iteration, the loop counter is updated. When the deterministic constraint solver outputs a result indicating that the verification failed, the current value of the loop counter is compared with a preset maximum iteration threshold. If the current value is less than or equal to the maximum iteration threshold, the process continues to parse the conflict information output by the deterministic constraint solver and locate the functional role of the component corresponding to the conflict. If the current value is greater than the maximum iteration threshold, the iteration is stopped and the candidate supplement termination condition is determined to have been met, generating a diagnostic report containing all candidate components that have been traversed and their corresponding conflict causes. If the deterministic constraint solver outputs a verification that passes in any iteration, the iteration is immediately stopped and the feasible assembly configuration is output.

[0057] A deterministic constraint-solving system under local dynamic activation of a candidate pool includes:

[0058] The requirement structuring conversion module is used to acquire unstructured mechanical design requirements input by the user and output the structured requirement status.

[0059] The question generation module constructs a candidate pool of fields to be clarified based on the structured requirement status, calculates the two-step look-ahead value, selects the candidate field with the largest two-step look-ahead value as the field to be clarified, and generates and outputs the clarification question by matching the clarification question template library.

[0060] The user response verification processing module is used to receive the response content returned by the user in response to the clarification question and perform verification. After all fields to be clarified are verified, the structured requirement status of verification completion is output.

[0061] The candidate pool local dynamic activation module parses the structured requirement status after verification, constructs a functional role requirement mapping table, generates an initial candidate component pool with controlled scale, and calls the language model to obtain the component candidate number to complete the local dynamic activation of the candidate pool.

[0062] The deterministic constraint solution verification module is used to read the technical parameters of the activated local candidate parts, perform unified verification, and output the verification results, feasible assembly configurations, or conflict information.

[0063] The conflict localization module is used to parse the conflict information output by the deterministic constraint solver, reverse locate the functional role of the component corresponding to the conflict, generate supplementary search keywords, call the language model to supplement the component candidates for the conflict functional role, and update the local candidate component number set.

[0064] Compared with the prior art, the beneficial effects of the present invention are:

[0065] 1. This invention proposes a deterministic constraint solution method under local dynamic activation of the candidate pool. It extracts entities and maps fields from unstructured natural language design requirements of users, constructing a structured requirement state with importance, acquisition cost, state markers, and constraint relationships. A two-step forward-looking value comprehensive consideration of the network influence and information gain of field constraints is adopted, prioritizing high-value fields to generate clarifying questions and avoiding meaningless human-computer question-and-answer sessions. Simultaneously, the method performs type verification and hard constraint pre-verification on the response content, intercepting parameter errors and constraint conflicts in advance. This reduces solution failures caused by defective requirement input during the component matching stage, effectively improving the accuracy of requirement understanding and reducing the number of invalid interactions.

[0066] 2. This invention proposes a deterministic constraint solving method under local dynamic activation of the candidate pool. It employs a technical path combining local dynamic activation of the candidate pool with deterministic constraint solving to reduce system computational overhead and improve the efficiency and reliability of iterative solutions for assembly schemes. This invention generates an initial candidate pool of components with controlled size. After outputting candidate numbers via a language model, multi-dimensional filtering and verification are performed to complete local candidate activation. When a conflict occurs in constraint solving, instead of globally refreshing all component candidates, the functional role of the conflicting component is located, and component candidates are only added to the conflicting role, reducing redundant computation caused by repeated searches and verifications. A deterministic constraint solver is used to perform rule-based verification of numerical boundaries, interface compatibility, and cross-component assembly coupling consistency, compensating for the lack of logical consistency in schemes generated by pure language models. Combined with a cyclic iteration mechanism, assembly conflicts are continuously repaired before the termination condition is reached, ensuring the engineering feasibility of the assembly configuration scheme and accelerating the algorithm's convergence speed. It can stably output feasible assembly configurations with accompanying 3D models, bills of materials, and metadata, adapting to the actual use scenarios of computer-aided design systems. Attached Figure Description

[0067] Figure 1 This is a schematic diagram of a deterministic constraint solution method under local dynamic activation of the candidate pool according to the present invention;

[0068] Figure 2 This is a flowchart illustrating the principle of a deterministic constraint solution method under local dynamic activation of the candidate pool according to the present invention.

[0069] Figure 3 This is a system architecture diagram for solving deterministic constraints under local dynamic activation of the candidate pool according to the present invention. Detailed Implementation

[0070] Example 1:

[0071] Please see Figure 1 and Figure 2 The present invention provides an embodiment of a deterministic constraint solution method under local dynamic activation of a candidate pool, comprising the following steps:

[0072] S1: Obtain unstructured mechanical design requirements from user input and convert them into structured requirement states that include field types, importance, acquisition costs, status flags, and constraint relationships;

[0073] The specific steps of S1 include:

[0074] S1.1: Obtain unstructured mechanical design requirement text input by the user based on a natural language processing interface;

[0075] Furthermore, in actual operation, the natural language processing interface adopts a locally deployed industrial-specific natural language inference service interface. The interface call uses the HTTP POST request method, and the maximum character length of the text input received by the interface in a single instance is set to 1200 characters. This setting is based on the statistical results of user natural language input in mechanical design scenarios. The number of natural language descriptions for most mechanical design requirements will not exceed 1200 characters, which can fully accommodate the entire design description content entered by the user, while avoiding the problem of excessive interface memory consumption caused by excessively long text input. Users enter the original requirements description related to mechanical design in the text input box on the front-end interactive page. The input content can include various colloquial and fragmented text content such as equipment performance indicators, external dimension limitations, working environment conditions, cost budgets, assembly constraints, etc. The interface will completely read all the text content in the input box, filter invalid spaces, line breaks, and invisible special control characters at the beginning and end of the text, retain valid characters, numbers, and unit symbols, and send the cleaned original text to the word segmentation processing flow. The example of the unstructured mechanical design requirements text input by the actual user selected in this embodiment is as follows: A small material handling equipment needs to be designed. The overall weight of the equipment cannot exceed 180kg, the working environment temperature is -10℃ to 45℃, the equipment needs to have a two-stage reduction transmission structure, the maximum length of the overall assembled shape cannot exceed 1200mm, the maximum width cannot exceed 800mm, the total material procurement cost of a single piece of equipment cannot exceed 12,000 yuan, and the equipment needs to support continuous uninterrupted operation for more than 8 hours.

[0076] S1.2: Based on the unstructured mechanical design requirements, a natural language processing word segmentation algorithm is called to obtain the required word sequence. The word segmentation algorithm is existing technology in this field and is not an inventive solution of this application, so it will not be described in detail here.

[0077] Furthermore, this embodiment selects a word segmentation algorithm optimized for the mechanical industry. The algorithm pre-constructs 23,000 entries from a specialized mechanical dictionary, covering component names, performance parameters, units of measurement, operating condition descriptions, and constraint terms. The matching priority of these specialized dictionary entries is set higher than that of general word segmentation dictionaries. This is based on the fact that mechanical design requirements texts contain a large number of industrial technical terms, and prioritizing matching with the specialized dictionary can prevent these technical terms from being incorrectly segmented into multiple fragmented words. The maximum forward matching window length of the word segmentation algorithm is set to 15 Chinese characters. This is based on the fact that the character length of most complex technical terms in the mechanical industry does not exceed 15 Chinese characters. Setting the window length too high will increase computation time, while setting it too low will easily lead to errors in segmenting technical terms. The cleaned original text is then fed into the word segmentation algorithm, which segments the text segment by segment, removing modal particles and function words that have no actual business meaning, and outputting an ordered sequence of requirement words. The unstructured mechanical design requirement text selected in this embodiment, after word segmentation, yields the following requirement word sequence: design, small material handling equipment, equipment, overall weight, not exceeding, 180kg, working environment temperature, -10℃, 45℃, two-stage reduction transmission structure, assembly completed, maximum length, not greater than, 1200mm, maximum width, not greater than, 800mm, single unit, total material procurement cost, not exceeding, 12000 yuan, equipment, support, continuous uninterrupted operation, 8 hours.

[0078] S1.3: Based on the sequence of required words, entity words are extracted using a named entity recognition algorithm to generate an initial entity set. The named entity recognition algorithm is existing technology in this field and is not an inventive solution of this application, so it will not be described in detail here.

[0079] Furthermore, in this embodiment, the named entity recognition algorithm adopts an industrial text entity recognition model. During the model training phase, 17,000 annotated mechanical design requirement corpora are used for training. The entity categories supported by the model recognition are divided into five major categories: performance parameter entities, size parameter entities, cost parameter entities, operating environment entities, and structural configuration entities. During the model inference phase, the entity recognition confidence filtering threshold is set to 0.75. This threshold is based on multiple sets of comparative tests from the model testing phase. A threshold of 0.75 can filter out a large number of incorrect recognition results with low confidence, while not discarding correctly recognized valid entities. If the entity recognition confidence filtering threshold is set too high, some valid entities will not be extracted; if the threshold is set too low, a large number of incorrectly recognized invalid entities will be introduced. The requirement word sequence output in the previous step is traversed sequentially, and entity judgment is performed word by word. Recognition results with a confidence level greater than or equal to 0.75 are output as valid entity words. All valid entity words together form the initial entity set. For the example requirements of this embodiment, the final initial entity set includes: overall weight, 180kg, working environment temperature, -10℃, 45℃, two-stage reduction transmission structure, maximum length, 1200mm, maximum width, 800mm, total material procurement cost, 12000 yuan, and continuous uninterrupted working time, 8 hours.

[0080] S1.4: Based on the initial entity set, a rule matching method is used to map each entity word to a predefined field type to generate field type annotation results. The rule matching method is existing technology in this field and is not an inventive solution of this application, so it will not be described in detail here.

[0081] Furthermore, this embodiment pre-maintains a predefined field type mapping rule library. This library contains 58 commonly used field types in mechanical design. Each rule includes an entity keyword matching expression, the target field type, and a description of the field value data format. Rule matching is executed in descending order of priority, with higher-priority rules matching first. Once a match is found, subsequent lower-priority rules are not executed. The rule priority value ranges from 1 to 100, with higher values ​​indicating higher priority. The priority of rules for core fields related to performance, size, and cost is uniformly set to 80 or higher, while the priority of rules for secondary descriptive fields is set between 20 and 50. Each entity word in the initial entity set is traversed and compared sequentially with the matching expression in the rule library. When an entity word matches the matching expression of any rule, the entity is mapped to the corresponding field type, and the value carried by the entity is recorded. For the example requirements of this embodiment, the overall weight of the entity is mapped to the "Machine Weight" field type, with a value of 180kg; the operating environment temperature of the entity is mapped to the "Ambient Operating Temperature" field type, with a lower limit of -10℃ and an upper limit of 45℃; the two-stage reduction transmission structure of the entity is mapped to the "Transmission Stage" digital segment type, with two values; the maximum length of the entity's external shape is mapped to the "Maximum Length of the Machine" field type, with a value of 1200mm; the maximum width of the entity is mapped to the "Maximum Width of the Machine" field type, with a value of 800mm; the total material procurement cost of the entity is mapped to the "Total Material Budget" field type, with a value of 12000 yuan; and the continuous uninterrupted working time of the entity is mapped to the "Continuous Working Time" field type, with a value of 8 hours. After all entities are mapped, the complete field type annotation results are output.

[0082] S1.5: Based on the field type labeling results, the importance of each field is calculated by calling the analytic hierarchy process (AHP). The AHP is a prior art in this field and is not an inventive solution of this application, so it will not be described in detail here.

[0083] Furthermore, in the execution of the analytic hierarchy process in this embodiment, a pairwise comparison judgment matrix is ​​first constructed. The number of rows and columns of the judgment matrix is ​​equal to the total number of fields identified in this instance. In this example, a total of 7 valid fields were identified, so a 7×7 judgment matrix is ​​constructed. The comparison scale of the judgment matrix elements adopts a scale system from one to nine. The scale values ​​are filled in based on the business experience provided in advance by mechanical design experts. For the design scenario of small material handling equipment, the overall weight, overall dimensions, and total material budget are high-influence fields, and the pairwise comparison scale is set to 7; the working environment temperature and transmission stages are medium-influence fields, and the pairwise comparison scale is set to 4; the continuous uninterrupted working time has a relatively low weight, and the pairwise comparison scale is set to 2. After the judgment matrix is ​​constructed, a matrix consistency check is performed. The maximum consistency ratio threshold allowed for the consistency check is set to 0.1. The maximum consistency ratio threshold is a standard threshold commonly used in the industrial application of the analytic hierarchy process. When the calculated consistency ratio is less than or equal to 0.1, it means that the judgment matrix logic is reasonable, and the weights can continue to be calculated; if the consistency ratio is greater than 0.1, the operator is prompted to readjust the pairwise comparison scale values. In this example, the consistency ratio calculated by the judgment matrix is ​​0.07, which meets the verification condition. The matrix eigenvectors are then calculated and normalized. Each component obtained after normalization corresponds to the importance value of a field, and the sum of all field importance values ​​equals 1. The output of this example scenario is as follows: overall machine weight importance 0.22, overall machine maximum length importance 0.18, overall machine maximum width importance 0.18, total material budget importance 0.2, operating environment temperature importance 0.12, number of transmission stages importance 0.06, and continuous uninterrupted working time importance 0.04.

[0084] S1.6: Based on importance and combined with a pre-set threshold for field collection time, a cost evaluation function is used to calculate the acquisition cost;

[0085] Furthermore, this embodiment pre-sets a threshold for field acquisition time, measured in seconds. In this mechanical design requirements analysis scenario, this threshold is set to 45 seconds. This threshold is based on the acceptable waiting time for user interaction clarification. If the estimated time required for complete confirmation of any field exceeds 45 seconds, the acquisition cost of that field is considered too high. The cost evaluation function's input includes field importance and the average historical acquisition time for the field. The average historical acquisition time is statistically obtained from a historical design case library, using the most recent 1200 historical design records as the sample size. This statistically analyzes the average interaction time consumed to complete information confirmation for the same field in past requirements analysis stages. The cost evaluation function outputs acquisition costs categorized into four levels: 1, 2, 3, and 4. A higher level value indicates a higher acquisition cost. When the average historical data collection time for a field is less than 15 seconds and the field importance is greater than 0.1, the acquisition cost level is assigned as 1. When the average historical data collection time for a field is greater than or equal to 15 seconds and less than 45 seconds, and the field importance is greater than 0.05, the acquisition cost level is assigned as 2. When the average historical data collection time for a field is less than 45 seconds and the field importance is less than or equal to 0.05, the acquisition cost level is assigned as 3. When the average historical data collection time for a field is greater than or equal to 45 seconds, regardless of the importance value, the acquisition cost level is assigned as 4. For this example, the average historical data collection time for the overall machine weight, maximum overall machine length, maximum overall machine width, and total material budget are 8 seconds, 7 seconds, 7 seconds, and 9 seconds respectively, all with an acquisition cost level of 1. The average historical data collection time for the operating environment temperature is 12 seconds, with an acquisition cost level of 1. The average historical data collection time for the number of transmission stages is 22 seconds, with an acquisition cost level of 2. The average historical data collection time for continuous uninterrupted working duration is 28 seconds, with an acquisition cost level of 3.

[0086] S1.7: Based on the acquisition cost, set a status flag according to the range of the cost value; the status flag includes unprocessed, pending clarification, verified, and locked;

[0087] Furthermore, in this embodiment, the allocation logic for status markers is based on the acquisition cost level, combined with whether a field has already obtained a clear and valid value. Fields that have already obtained a valid and compliant value from the user's original input text have their initial status marker set to "Unprocessed." If a field has not obtained a valid value, or if the obtained value is ambiguous or incomplete, it is directly marked as "Pending Clarification." After subsequent user interaction and confirmation, and after the value has undergone type and constraint validation, the status marker is updated to "Validated." When a field value has undergone multiple rounds of validation and confirmation, and subsequent modifications are not permitted, the status marker is updated to "Locked." In this example, all seven fields have obtained their corresponding values ​​from the user's original input, therefore, all of them are initially marked as "Unprocessed."

[0088] S1.8: Based on state labeling, the dependency relationship between fields is analyzed using the association rule mining algorithm, constraint association relationship is established, and structured requirement state is generated. The association rule mining algorithm is existing technology in this field and is not an inventive solution of this application, so it will not be described in detail here.

[0089] Furthermore, in this embodiment, the input data source for the association rule mining algorithm is selected from 2500 complete structured requirement records in the historical design case library. The association rule mining algorithm sets a minimum support threshold of 0.12 and a maximum confidence threshold of 0.82. These two sets of parameter settings are based on multiple sets of comparative experiments. The minimum support of 0.12 can filter out occasional association rules with too low a frequency of occurrence, while the maximum confidence of 0.82 ensures that the mined field dependencies have high business credibility. The association rule mining algorithm traverses the historical requirement records, mining association rules where different fields appear simultaneously and their values ​​constrain each other, and outputs the direct dependencies, indirect dependencies, hard constraints, and soft constraints between fields. For example, the overall weight field will have a direct constraint relationship with the number of transmission stages and the total material budget, the overall dimensions and component selection parameters have an indirect constraint relationship, and the working environment temperature will form a hard constraint on the component material selection. All the mined dependency constraints are bound to each field, and the field type, importance, acquisition cost, status flag, and constraint relationships obtained in the previous steps are integrated together to finally generate a complete structured requirement state.

[0090] S2: Construct a candidate pool of fields to be clarified based on the structured requirement status, select fields to be clarified from the candidate pool using type relationships and two-step look-ahead value, and output the clarification question;

[0091] The specific steps of S2 include:

[0092] S2.1: Based on the structured requirement status, filter out the fields marked as unprocessed or pending clarification, and construct an initial candidate pool of fields pending clarification;

[0093] Furthermore, the system iterates through all fields within the structured requirement state, reads the status flag corresponding to each field, and extracts all fields marked as "unprocessed" or "awaiting clarification" to form an initial candidate pool of fields awaiting clarification. In the example scenario of this embodiment, all 7 fields are initially marked as "unprocessed," therefore the initial candidate pool of fields awaiting clarification contains a total of 7 fields.

[0094] S2.2: For each candidate field in the initial candidate pool of fields to be clarified, iterate through and calculate the number of fields that are directly and indirectly constrained by the candidate field in the mechanical design constraint network, and generate the first influence score;

[0095] Furthermore, this embodiment pre-builds a mechanical design constraint network, which is stored in a graph structure. Each node in the constraint network corresponds to a requirement field, and the edges of the graph represent constraint relationships between fields, distinguishing between direct and indirect constraint edges. For each candidate field in the initial candidate pool of fields to be clarified, starting from the network node corresponding to that candidate field, the number of directly connected nodes (i.e., the number of direct constraint fields) is counted, followed by the number of indirect constraint fields that require one or more intermediate nodes to connect. The first influence score is calculated as follows: the number of direct constraint fields multiplied by a weight coefficient of 2.5, plus the number of indirect constraint fields multiplied by a weight coefficient of 1. The weight coefficients are set based on the fact that direct constraints have a significantly higher impact on the design results than indirect constraints. In this example, the total machine weight field has 4 direct constraint fields and 6 indirect constraint fields, resulting in a first influence score of 4 × 2.5 + 6 × 1 = 16; the total material budget field has 3 direct constraint fields and 5 indirect constraint fields, resulting in a first influence score of 12.5. The remaining fields are calculated according to the same rules.

[0096] S2.3: Based on the field type and importance, retrieve completed cases similar to the candidate field from the historical design case library, statistically analyze the value distribution of the candidate field in similar completed cases, calculate the expected reduction in information entropy, and generate a second information gain score;

[0097] Furthermore, in this embodiment, when retrieving similar cases, the case similarity calculation dimensions include field type matching and field importance numerical deviation. The case similarity filtering threshold is set to 0.65, meaning only historical cases with a similarity greater than or equal to 0.65 are included in the statistical set. The threshold is set based on the premise that if the similarity is too low, the reference value of the case will decrease significantly. For the filtered set of similar cases, all values ​​of the candidate field that have appeared in historical cases are counted, and the frequency of different values ​​is counted. Based on the frequency distribution, the expected reduction in information entropy is calculated, and the calculated expected reduction in information entropy is directly used as the second information gain score. When the number of historical similar cases is less than thirty, the algorithm will perform a decay processing on the score, with a decay coefficient set to 0.6, to avoid score distortion caused by small sample statistics.

[0098] S2.4: The first influence score and the second information gain score are weighted and summed to calculate the two-step prospective value of each candidate field, and the candidate field with the largest two-step prospective value is selected as the field to be clarified;

[0099] Furthermore, in this embodiment, the weighting parameters for the weighted summation are set as follows: the first influence score is weighted at 0.6, and the second information gain score is weighted at 0.4. The weighting parameters are set based on the fact that the influence of the field itself on the constraint network has a slightly higher priority for solving mechanical design constraints than the information gain of historical cases. The two-step look-ahead value is equal to the first influence score multiplied by 0.6, plus the second information gain score multiplied by 0.4. After calculating the values ​​of all fields in the candidate pool, the numerical values ​​of all two-step look-ahead values ​​are compared, and the candidate field with the largest value is selected as the field to be clarified in this round.

[0100] S2.5: Based on the field type of the field to be clarified and its constraint relationship, match the corresponding question sentence from the preset clarification question template library, generate the clarification question and output it.

[0101] Furthermore, this embodiment includes a pre-defined clarification question template library, which stores 110 clarification question templates specifically for mechanical design scenarios. Each template is bound to a corresponding field type, and some templates also include placeholders related to constraint relationships. Once the field to be clarified is determined, the field type is read, and all templates bound to that field type are matched first. Simultaneously, the constraint relationships of that field are read, and the names and value ranges of the associated fields in the constraint relationships are filled into the corresponding placeholders within the template. This is then combined to form a complete natural language clarification question, which is pushed to the front-end interactive interface for display to the user. For example, if the field to be clarified is the total weight of the machine, and the associated constraint is the total material budget, the generated clarification question would be: "Please confirm that the maximum weight of the equipment is 180kg. This value directly affects the total material procurement budget. Please confirm whether this parameter is accurate."

[0102] S3: Receive the response content returned by the user for the clarification question, perform source record, type validation and hard constraint pre-validation on the response content, and generate a structured requirement state that has completed the validation;

[0103] The specific steps of S3 include:

[0104] S3.1: Receive the response content returned by the user through the user interaction interface, and record the data source type of the response content;

[0105] Furthermore, in this embodiment, the data source types of the response content are divided into three categories: the first category is manual input from the user interface, the second category is obtained by parsing imported external requirement documents, and the third category is imported and backfilled from historical cases. While reading the response content, the system records which source type the response belongs to. The recorded information is stored in the corresponding field's metadata in the structured requirement status for subsequent problem tracing. After receiving the response content, basic text cleaning is performed to remove redundant spaces and line breaks, retaining numbers, units, and descriptive text.

[0106] S3.2: Based on the field type of the field to be clarified, call the corresponding type validator to perform type validation on the response content; if the type validation fails, generate an error message and request the user to re-enter; if the type validation passes, based on the constraint relationship, extract the related fields that have a direct hard dependency relationship with the field to be clarified and their hard constraint conditions, and perform hard constraint pre-validation on the response content; if the hard constraint pre-validation passes, fill the response content as a validated value into the structured requirement status, and update the status mark of the corresponding field to validated; if the hard constraint pre-validation fails, generate a conflict report and return it to the user.

[0107] Furthermore, in this embodiment, each field type is configured with a dedicated type validator. Different field types correspond to different validation rules. The size field validator checks whether it contains a valid length unit and whether the value is within the reasonable range of mechanical design. In this example, the reasonable length range is set to 5mm to 5000mm. The weight field validator checks the validity of the weight unit, with a reasonable value range set to 0.5kg to 2000kg. The budget cost field checks the currency unit, with a reasonable value range set to 100 yuan to 500,000 yuan. When the response content cannot match the valid data format corresponding to the field, or the value exceeds the reasonable range, the type validation is deemed to have failed. The front end displays an error message, clearly informing the user that the current input does not meet the data format requirements of the field and requesting the user to re-enter the correct content.

[0108] Furthermore, after successful type validation, the system reads the constraint relationships of the field to be clarified, filtering out related fields with direct hard dependencies and hard constraints. For example, the total weight field and the total material budget have direct hard constraints; the larger the total weight value, the higher the corresponding lower limit of material procurement cost. The system substitutes the newly submitted response value into the hard constraints for pre-validation, determining whether the new value will conflict with the already validated related field values. If all hard constraint pre-validation conditions are met, the response content is saved as the validated value of that field, and the field's status is changed from unprocessed or pending clarification to validated. If the hard constraint pre-validation finds a conflict, a conflict report is generated. The conflict report specifies the conflicting field name, the current input value of the field, the validated value of the related field, and the specific constraint violated. The complete conflict report is output to the user interface.

[0109] S3.3: After all fields to be clarified have passed type validation and hard constraint pre-validation, generate the structured requirement state that has been validated.

[0110] Furthermore, after generating the structured requirement state that has been verified, an integrity check step is also included: based on the structured requirement state that has been verified, all verified fields and their corresponding verified values ​​are extracted to construct a set of fields that have been verified; a preset integrity check rule is obtained, which includes a list of required fields and logical consistency rules between fields; based on the list of required fields, it is checked whether the set of fields that has been verified covers all required fields. If there are any missing fields, the status of the missing field is marked as pending clarification and added to the candidate pool of fields pending clarification; based on the logical consistency rules, the values ​​of multiple logically related fields are checked for consistency. If the check fails, a logical conflict report is generated and a re-clarification process for the conflicting fields is triggered.

[0111] This embodiment pre-defines integrity check rules. The list of required fields for small material handling equipment includes 11 fields: maximum weight of the whole machine, maximum length of the whole machine, maximum width of the whole machine, maximum height of the whole machine, minimum operating temperature, maximum operating temperature, maximum total material procurement budget, number of transmission structure stages, continuous working time, rated load, and protection level. The rule configuration file records the field identifier for each required field. The system extracts the set of fields that have been verified and compares them one by one with the list of required fields. If any field in the list is not present in the set of fields that have been verified, the status of the missing field is changed to "awaiting clarification," and it is added back to the candidate pool of fields awaiting clarification. It will then re-enter the clarification selection process in S2. A total of 207 logical consistency rules are configured between fields. For example, the value of the minimum operating temperature must be less than the maximum operating temperature, the value of the maximum length of the whole machine must be greater than the minimum theoretical length required for component assembly, and the total material budget must be greater than the sum of the minimum procurement costs of all core components. Execute the logical consistency rule verification one by one. If any rule verification fails, generate a logical conflict report. The report specifies the field name involved in the conflict, the current value of each field, the logical rule violated, and triggers the re-clarification process for the conflicted field. The status of the conflicted field is reset to pending clarification and it is put back into the candidate pool of fields pending clarification.

[0112] Furthermore, after generating the structured requirement state that has been verified, the method also includes a data clustering anomaly review step: based on the structured requirement state that has been verified, all verified fields and their corresponding verified values ​​are extracted to construct a verified dataset; a density-based clustering algorithm is called to perform cluster analysis on the verified dataset to identify potential outlier value combinations; when the cluster analysis result indicates the presence of outlier values, the verified field corresponding to the outlier value is marked as pending review, and a review prompt message is generated; the review prompt message is output to the user interface, and confirmation or modification instructions input by the user for the outlier value are received; the structured requirement state that has been verified is updated based on the confirmation or modification instructions. The density-based clustering algorithm is existing technology in this field and is not an inventive solution of this application, and will not be described in detail here.

[0113] In this embodiment, a density-based clustering algorithm is used. The neighborhood radius parameter is set to 0.24, and the minimum number of sample points within the neighborhood is set to 8. These parameters are set based on the following: the training and parameter tuning process uses a dataset of 4100 historical normal mechanical design requirements. After comparing multiple sets of parameters, a neighborhood radius of 0.24 and a minimum number of sample points of 8 achieve the best outlier identification effect on this dataset. This ensures that normal and reasonable requirements are not identified as outliers, while effectively identifying abnormal value combinations that significantly deviate from the historical design range. Before the algorithm executes, the values ​​of multi-dimensional fields are normalized, mapping all values ​​to the range of 0 to 1, eliminating interference caused by differences in the dimensions of different fields. After execution, data sample points not within any dense cluster are identified as outlier value combinations. After identifying outliers, the corresponding fields are marked as pending review, and a review prompt message is generated. The prompt message specifies which field value combinations deviate from the distribution range of historical normal design cases, asking the user to confirm whether the set of parameters is the actual parameter required for the design. The review prompt message is displayed on the user interface, awaiting user action. Users can choose to confirm that the parameters remain unchanged, or they can choose to modify the value of the corresponding field. After receiving the user's confirmation or modification instruction, if the user confirms, the field status is changed from pending review to verified; if the user performs the modification, the field value is updated, and the status is kept as pending clarification, and the field is resubmitted to the candidate pool of fields pending clarification for a new clarification and verification process.

[0114] S4: Based on the structured requirement status that has been verified, generate an initial candidate component pool with controlled size according to the functional roles of the components, and call the language model to output the component candidate number. Then, perform whitelist verification, functional role matching, usage scenario adaptation and duplicate item removal on the component candidate number to complete the local dynamic activation of the candidate pool.

[0115] Based on the structured requirement status that has been verified, an initial candidate component pool of controlled size is generated according to the functional roles of the components, including:

[0116] S4.1: Parse the structured requirement status that has been verified and extract all verified fields and their verified values;

[0117] Furthermore, the entire content of the structured requirement status after verification is fully read, each field is traversed, fields whose status markers are not verified are filtered out, and the field identifiers and corresponding verified values ​​of the remaining fields are extracted and organized into a set of key-value pairs for subsequent functional role matching. In this embodiment, the extracted parameters are: maximum weight of the whole machine 180kg, maximum length of the whole machine 1200mm, maximum width of the whole machine 800mm, minimum operating temperature -10℃, maximum operating temperature 45℃, two transmission stages, total material procurement budget of 12,000 yuan, and continuous uninterrupted working time of 8 hours.

[0118] S4.2: Based on the preset mechanical function role classification tree, match each verified field with its corresponding component function role to generate a function role requirement mapping table;

[0119] Furthermore, this embodiment pre-defines a mechanical function role classification tree. This tree adopts a multi-level tree structure, with the root node representing the entire machine. First-level child nodes include transmission components, support frame components, power components, protective components, connecting fasteners, and load-bearing components. Each first-level child node is further subdivided into second- and third-level sub-functional roles. The entire classification tree contains a total of 67 end-component functional roles. The system traverses each verified field, reads its business meaning, matches it upwards to the corresponding end-component functional role in the classification tree, establishes a mapping relationship between field values ​​and component functional roles, and summarizes and organizes all mapping relationships to generate a functional role requirement mapping table. In this example scenario, the transmission stage digital segment matches the functional roles of two-stage reduction transmission components; the overall machine weight and overall machine dimensions match the functional roles of support frame components and load-bearing components; the operating environment temperature field matches all component functional roles; and the total material budget field creates cost constraints for all functional roles.

[0120] S4.3: For each functional role in the functional role requirement mapping table, retrieve components that meet the technical parameter range of the corresponding functional role from the material master database of the enterprise resource planning system to form a candidate component set;

[0121] Furthermore, the material master database of the enterprise resource planning system stores all available component material records within the enterprise. Each material record contains complete information such as material number, component functional role label, various technical parameters, material purchase price, supplier code, and working environment adaptation parameters. For each component functional role in the mapping table, material records whose functional role labels in the database completely match are filtered out. At the same time, components whose technical parameters do not meet the general parameter constraints given by the verified fields are filtered out. A set of candidate components is output for each functional role. In this example, the candidate component set for the two-stage reduction gear transmission component yields 142 material records; the support frame component yields 97 material records; and the load-bearing component yields 86 material records.

[0122] S4.4: Obtain a preset upper limit threshold for the size of the candidate pool. When the number of parts in any candidate part set exceeds the upper limit threshold for the size of the candidate pool, use a maximum-minimum distance sampling algorithm based on space filling design to uniformly extract a subset of the candidate part set, so that the size of the extracted candidate part set does not exceed the upper limit threshold for the size of the candidate pool, and generate an initial candidate part pool with controlled size.

[0123] Furthermore, in this embodiment, the upper limit threshold for the candidate pool size is set to 60. This threshold is set based on the processing performance of subsequent language model calls and constraint solvers. Keeping the number of components in a single set below 60 ensures both the richness of candidate solutions and the speed of the algorithm, preventing excessive processing latency due to large amounts of data processed at once. When the number of material records within a candidate component set corresponding to a certain functional role exceeds 60, the maximum-minimum distance sampling algorithm is activated for subset extraction. If the number of materials within the set is less than or equal to 60, all components are retained without sampling. In this example, the number of materials within the candidate set for the two-stage reduction transmission component is 142, exceeding the upper limit threshold of 60, requiring sampling; the number of materials within the candidate set for the support frame component is 97, exceeding the threshold of 60, requiring sampling; and the number of materials within the candidate set for the load-bearing component is 86, exceeding the threshold of 60, requiring sampling.

[0124] Furthermore, the specific implementation of the maximum-minimum distance sampling algorithm is as follows: Obtain the multi-dimensional technical parameter vector of each component in the candidate component set, where the multi-dimensional technical parameter vector includes size parameters, material parameters, and performance parameters; map the multi-dimensional technical parameter vectors of all components to a parameter space to form a parameter space point set; use the maximum-minimum distance sampling algorithm to perform subset extraction on the parameter space point set: first, randomly select a point from the parameter space point set as the first sampling point; then, select the point with the largest minimum distance to all current sampling points from the remaining points as the next sampling point, repeating this step until the number of selected sampling points reaches the upper limit threshold of the candidate pool size; use the components corresponding to the finally selected sampling point set as the initial candidate component pool with controlled size.

[0125] In this embodiment, the multidimensional technical parameter vector has 11 dimensions, including external length, external width, external height, rated load capacity, minimum withstand temperature, maximum withstand temperature, rated continuous working time, component weight, purchase price, material yield strength, and transmission output torque. Before executing the algorithm, each dimension of the 11-dimensional vector is normalized to the zero-to-one range, eliminating distance calculation biases caused by different parameter units and numerical magnitudes. Euclidean distance is used for distance calculation. When randomly selecting the first sampling point, the random seed value is fixed at 1127. A fixed random seed ensures that the sampling output results obtained from each run of the algorithm are consistent for the same set of input components, facilitating engineering debugging and reproduction. After selecting the first sampling point, all remaining unselected points are traversed, and the Euclidean distance between each point and every selected sampling point is calculated. The smallest distance value is selected, and the point with the largest minimum distance value among all remaining points is chosen as the next sampling point. This selection step is repeated until the total number of sampling points equals the upper limit threshold of the candidate pool (60), at which point the loop stops. All sampled points are mapped back to their corresponding component material numbers to obtain a subset of components for that functional role after sampling. After all functional roles have completed processing, the resulting pool of initial candidate components with controlled size is obtained.

[0126] The process involves calling the language model to output candidate component numbers, then performing whitelist verification, functional role matching, usage scenario adaptation, and duplicate item removal on these candidate numbers to achieve partial dynamic activation of the candidate pool, including:

[0127] S4.5: Encapsulate the technical parameters of each component in the initial candidate component pool into a prompt word sequence, call the pre-trained language model interface, and obtain the component candidate number list corresponding to each functional role output by the language model;

[0128] Furthermore, the invocation of the language model is performed based on a few-shot prompting learning method: Multiple successful design cases with the highest similarity to the currently verified structured requirement state are retrieved from the historical design case library. Each successful design case contains a complete structured requirement state and its corresponding final selection of component candidate numbers. Input-output pairs are extracted from the multiple successful design cases, where the input is the complete verified structured requirement state and the output is the corresponding component candidate number list. The input-output pairs are used as few-shot examples and formatted according to a preset prompt word template to generate complete prompt words containing the task description, the few-shot example, and the current structured requirement state to be processed. The complete prompt words are input into the pre-trained language model to obtain the component candidate number list output by the language model.

[0129] Furthermore, in this embodiment, similar successful design cases are retrieved. The case similarity calculation comprehensively compares the deviations of the values ​​of each verified field. The similarity screening threshold is set to 0.7, and the top six successful cases with the highest similarity ranking are selected as minority sample examples. The number of minority sample examples is set to 6, based on the limit of the prompt word context window capacity. Six sets of examples can provide sufficient reference examples for the model without causing the total length of the prompt words to exceed the limit. The preset prompt word template contains a task description paragraph, a minority sample example storage area, and a current requirement input area. The task description paragraph specifies the task objective, requiring the model to combine mechanical design requirements and select suitable part numbers from the initial candidate part number pool for output. Only the part numbers are output, without any extra text. The input-output pairs of the six sets of historical successful design cases are sequentially filled into the minority sample example storage area. Then, the current structured requirement status of the completed verification, as well as the technical parameters and part numbers of all parts in the initial candidate part pool, are filled into the current requirement input area to assemble a complete prompt word. The maximum number of output tokens for a single request of the language model interface is set to 1800, and the output temperature parameter is set to 0.2. This relatively low temperature setting ensures the output results are more deterministic, reducing invalid material numbers generated randomly. The kernel sampling probability parameter is set to 0.75. After the interface returns the results, the returned text is parsed to extract the candidate component number list for each functional role.

[0130] S4.6: Based on a preset whitelist, remove numbers from the candidate part number list that are not in the whitelist and generate a first intermediate list;

[0131] Furthermore, this embodiment pre-defines a whitelist of qualified suppliers. The whitelist records all supplier codes that the enterprise is allowed to purchase from. Each component material record is bound to a corresponding supplier code. The component candidate number list output by the language model is traversed, and the corresponding supplier code in the material master database is read according to the material number. If the supplier code does not exist in the qualified supplier whitelist, the material number is directly deleted from the list, and the remaining material numbers form the first intermediate list.

[0132] S4.7: Based on the functional role requirement mapping table, retain the numbers of functional roles in the first intermediate list that match the verified fields, and generate a second intermediate list;

[0133] Further, the functional role requirement mapping table is read to obtain the tag for each functional role to be processed. Each material number in the first intermediate list is read, and the corresponding functional role tag for that component is obtained by querying the material master database. Only material functional role tags that completely match the currently processed functional role are retained; all other mismatched material numbers are discarded. After processing, the second intermediate list is obtained.

[0134] S4.8: Based on the working environment parameters in the structured requirement state that has been verified, remove the numbers in the second intermediate list that do not meet the working environment adaptability standard, and generate a third intermediate list;

[0135] Furthermore, from the structured requirements status that has been verified, the lower limit of the operating temperature is -10℃, and the upper limit of the operating temperature is 45℃. Each material number in the second intermediate list is traversed, and the component's own temperature tolerance range is read from the material master database. The component's minimum tolerable temperature must be less than or equal to the lower limit of the required operating temperature, and the component's maximum tolerable temperature must be greater than or equal to the upper limit of the required operating temperature. If both conditions are met, the component is considered to have met environmental adaptability standards; if either condition is not met, the material number is removed. After completing the traversal and filtering, the third intermediate list is obtained.

[0136] S4.9: Perform deduplication on the third intermediate list and use the final retained numbers as the set of local candidate component numbers after activation.

[0137] Furthermore, the third intermediate list is traversed to identify duplicate material numbers. For cases where the same material number appears multiple times, only the first record is retained, and the remaining duplicate entries are deleted. After all duplicates are removed, the remaining material numbers constitute the activated local candidate component number set.

[0138] S5: Input the candidate parts obtained after local dynamic activation into the deterministic constraint solver for unified verification; if the verification passes, output the feasible assembly configuration; if the verification fails, parse the conflict information output by the deterministic constraint solver, locate the functional role of the parts corresponding to the conflict, and call the language model to supplement the candidate parts only for the functional role of the parts corresponding to the conflict.

[0139] The step of inputting candidate components obtained through local dynamic activation into the deterministic constraint solver for unified verification includes:

[0140] S5.1: Obtain the activated set of local candidate component numbers, and based on the material master database, extract the detailed technical parameter table corresponding to each number in the activated set of local candidate component numbers;

[0141] Furthermore, the material numbers within the activated local candidate component number set are read one by one, the material master database is accessed, and the complete and detailed technical parameter table corresponding to each material number is read. The parameter table contains all technical fields such as size, weight, material, interface geometry, interface physical properties, performance indicators, and procurement cost. After all the detailed technical parameter tables of the components are read, they are cached in memory for the solver to call.

[0142] S5.2: Encode the detailed technical parameter table and all verified fields in the structured requirement state that has been verified, along with the constraint relationships, into the input fact set of the deterministic constraint solver;

[0143] Furthermore, the entire detailed technical parameter table of the output is read, and each detailed technical parameter table corresponding to each component is traversed page by page. The component material number and the name and value of each technical parameter corresponding to that component are extracted. A separate fact record is generated for each technical parameter, specifying the material number, parameter name, and actual parameter value. When the text length of the fact record generated for a single technical parameter exceeds the 260-character threshold, the multi-dimensional parameters of that component are split into multiple independent fact records, ensuring that the text length of each fact record after splitting is no more than 260 characters. All verified fields in the structured requirement status that has been verified are read, and the field type, field identifier, and verified value of each field are extracted. A separate independent fact record is generated for each verified field, specifying the requirement field identifier, field type, and confirmed field value. In this example scenario, multiple fact records will be generated sequentially, corresponding to the following: maximum machine weight of 180kg, maximum machine length of 1200mm, maximum machine width of 800mm, minimum operating temperature of -10℃, maximum operating temperature of 45℃, two transmission stages, maximum material procurement budget of 12,000 yuan, and continuous uninterrupted working time of 8 hours. All constraint relationships in the verified structured requirement state are read, each constraint relationship is traversed, and the set of field identifiers, constraint type, and constraint judgment conditions involved in the constraint are extracted. Each constraint relationship is converted into an independent constraint fact record. The constraint fact record fully records the identifiers of each object involved in the constraint and the constraint judgment conditions. No fuzzy descriptions are used; all conditions are expressed deterministically. In this example, the constraint fact records generated by the constraint relationship conversion include multiple fact records such as: the sum of the total machine weight must not exceed the maximum machine weight; the sum of the cumulative component procurement costs must not exceed the maximum material procurement budget; and the operating environment temperature must fall within the component's tolerance temperature range. All generated fact records are aggregated and merged. During the merging process, deduplication verification is performed on the fact records, comparing the complete text content of the fact records and deleting fact records with completely duplicate content to avoid the solver repeatedly parsing the same conditions. In this embodiment, the maximum number of fact records that the input fact set can hold is set to 1200. This setting is based on the measured performance of the deterministic constraint solver. When the total number of fact records exceeds 1200, the time taken for a single verification process by the solver will increase significantly. When the total number of fact records after aggregation exceeds 1200, priority is given to retaining numerical boundary facts and interface constraint facts with high conflict risk, while eliminating duplicate derived secondary derivations, so that the total number of fact records obtained in the final aggregation does not exceed 1200. All fact records after aggregation, deduplication, and quantity control are uniformly formatted and packaged according to the fact expression format specified by the solver to form a complete input fact set.

[0144] S5.3: Start the deterministic constraint solver; the deterministic constraint solver has a set of preset verification rules, including numerical boundary verification rules based on interval arithmetic, interface type compatibility verification rules based on the interface standard library, and cross-component coupling consistency verification rules based on the assembly connection diagram; the deterministic constraint solver performs numerical boundary verification, interface type compatibility verification, and cross-component coupling consistency verification in sequence based on the input fact set, and outputs a unified verification result.

[0145] Furthermore, the numerical boundary verification rules based on interval arithmetic specifically include: the solver loads all fact records related to numerical parameters from the input fact set within the memory workspace. This part of the fact records includes fact records of various technical parameters of the components themselves, fact records of verified fields of the overall machine requirements, and constraint fact records obtained by converting numerical constraint association relationships; the solver pre-configures a floating-point comparison tolerance deviation parameter for interval arithmetic verification, which is set to 0.01. The setting of this parameter is based on the numerical comparison allowable error commonly used in the field of mechanical processing, and is used to avoid misjudgments caused by small calculation deviations in floating-point data storage; the solver first traverses the fact records corresponding to all component technical parameters, extracts the actual values ​​of various quantifiable parameters such as size, weight, cost, and temperature of each component, and reads the corresponding numerical upper and lower limit constraint fact records from the demand side. The solver performs an interval comparison for each constrained numerical parameter, comparing the actual value of the component parameter with the allowable range specified in the requirements. If the actual value of the component parameter falls within the range specified in the requirements, and a floating-point comparison tolerance of 0.01 is considered, the single parameter is deemed to have passed the verification. If the actual value of the component parameter exceeds the range specified in the requirements, and the interval constraint is still not met after adding the tolerance deviation, the solver generates a corresponding temporary conflict record. The temporary conflict record records the verification rule type of the current conflict, the component material number corresponding to the conflict, the name of the parameter involved in the conflict, the actual value of the parameter, and the boundary of the constraint range specified in the requirements. This completes the independent parameter verification for a single component. After the interval comparison of the numbers, the solver continues to read the summation-type constraints in the constraint fact record. In this embodiment, typical summation-type constraints include the summation of the total weight of all components and the summation of the procurement costs of all components. The solver reads the parameter values ​​corresponding to all components involved in the summation calculation from the input fact set, completes the numerical summation calculation, and then compares the calculated result with the upper limit value of the constraint given by the requirement. The comparison process also uses a floating-point number of 0.01 to compare the tolerance deviation parameter. When the summation result exceeds the requirement constraint range, the solver also generates a corresponding temporary conflict record, recording the material numbers of all components involved in the summation calculation, the actual result obtained from the summation calculation, and the constraint boundary value given by the requirement. After all independent parameter verification and summation-type parameter verification are completed, the numerical boundary verification process ends, and all generated temporary conflict records are completely saved to the temporary conflict cache area inside the solver.

[0146] Furthermore, the interface type compatibility verification rules based on the interface standard library specifically include: the solver first extracts all component material numbers participating in this verification from the input fact set, and reads the component interface type codes recorded in the fact record based on the material numbers; the solver performs validity verification on each extracted interface type code to determine whether the code actually exists in the interface standard library. If an interface type code is not in the interface standard library, the solver generates a temporary conflict record, marks the interface type unknown conflict, and records the corresponding component material number and the unknown interface code carried by the component; after completing the interface code validity screening, the solver reads the assembly constraint-related fact records stored in the input fact set, identifies component pairing relationships with assembly docking requirements, reads the interface type codes of each pairing party for each pairing party, calls the interface standard library to query the set of compatible interface types corresponding to the interface type of the first component, determines whether the interface type code of the second component exists in the set of compatible interface types, and simultaneously... The algorithm queries the set of compatible interface types corresponding to the interface type of the second component and verifies whether the interface type code of the first component is within the set of compatible interface types. Only when both bidirectional verifications pass is the compatibility verification of the paired interface types of the component group considered to have passed. If either side of the bidirectional verification fails to match the set of compatible interface types, the solver generates a corresponding temporary conflict record, which records the verification rule type to which the conflict belongs, the material numbers of the two paired components, the interface type codes of both parties, and the set of interface codes that the interface is allowed to be compatible with in the interface standard library. All generated temporary conflict records are also stored in the temporary conflict cache area inside the solver and will not overwrite the conflict records already stored in the previous numerical boundary verification stage.

[0147] Furthermore, before executing the cross-component coupling consistency verification rules based on the assembly connection diagram, the solver needs to rely on the component material number fact records, component assembly connection constraint fact records, and component assembly interface definition related fact records within the input fact set to complete the construction of the assembly connection diagram. The solver internally configures a maximum number of nodes in the assembly connection diagram, which is set to 80. This parameter setting is based on the solver graph structure memory overhead test results; when the number of component nodes participating in the verification does not exceed 80, the processing latency of the graph traversal calculation is within a controllable range. In this embodiment, after the material candidate pool is activated, the number of components participating in the verification will not exceed 60, satisfying the constraint condition of this parameter. Each node in the assembly connection diagram corresponds to a component material number participating in the verification. The solver traverses the assembly connection constraint fact records within the input fact set. When two components have an assembly connection constraint relationship, an undirected connection edge is generated between the two corresponding graph nodes. The associated information of the connection edge records the fit tolerance requirements and the minimum constraint requirements for energy transfer efficiency corresponding to the assembly relationship. After the assembly connection diagram is fully constructed, the solver traverses all the connection edges of the assembly connection diagram and extracts the component pairs associated with each connection edge, that is, the component pairs that have a direct assembly connection relationship.

[0148] Furthermore, the cross-component coupling consistency verification rules based on the assembly connection diagram specifically include: For each pair of components, the solver reads the relevant fact records of the standard assembly interface definitions of the two components stored in the input fact set. The interface definition includes interface geometric feature parameters and interface physical property parameters. The solver performs interface geometric feature verification, which includes comparison and verification of interface hole diameter, shaft diameter, positioning slot size, and mating surface shape. The solver internally sets the lower limit of the allowable tolerance for ordinary clearance fits to 0.02mm and the upper limit of the allowable tolerance for ordinary clearance fits to 0.25mm. These two sets of values ​​are set based on the general engineering tolerance range for conventional mechanical clearance assembly. The solver reads the interface geometric dimension parameters of both paired components, calculates the actual fit tolerance value obtained after assembly, and compares the actual fit tolerance value with the preset upper and lower limits of the ordinary clearance fit tolerance. If the actual fit tolerance does not fall within the range of 0.02mm to 0.25mm, the solver generates a temporary conflict record, recording the type of verification rule to which the conflict belongs, the material number of the paired parts, the key geometric dimensions of the interfaces of both parties, the calculated actual fit tolerance value, and the preset tolerance allowable range. After completing the interface geometric feature verification, the solver continues to verify the interface physical properties, focusing on verifying the theoretical energy transfer efficiency of the interface combination in the transmission scenario. The solver internally sets a minimum allowable energy transfer efficiency parameter for the transmission interface combination, which is set to 85%, based on the equipment transmission assembly business indicators corresponding to this mechanical design case. The solver reads the rated transmission torque, allowable transmission power, and other parameters stored in the physical property fact record of the two parts interface, and calculates the theoretical energy transfer efficiency value corresponding to the interface pairing based on the interface pairing relationship. If the calculated theoretical energy transfer efficiency is lower than the parameter threshold of 85%, the solver generates a temporary conflict record, recording the type of verification rule to which the conflict belongs, the material number of the paired parts, the calculated actual theoretical energy transfer efficiency, and the minimum allowable energy transfer efficiency threshold. After the geometric features and physical properties of a single set of component pairs are verified, the solver continues to process the next set of component pairs in the assembly connection diagram until all directly connected component pairs are processed. All temporary conflict records generated are still stored in the temporary conflict buffer area inside the solver, and are stored together with the conflict records generated in the previous two verification stages.

[0149] If the verification passes, the feasible assembly configuration will be output, including:

[0150] S5.4: When the unified verification result indicates that all the verification rules are satisfied, the verification is deemed to have passed;

[0151] S5.5: Based on the activated set of local candidate component numbers, extract the 3D model file and bill of materials file corresponding to each number from the product data management system;

[0152] S5.6: Associate and package the 3D model file, bill of materials file and the activated part number set, and attach all verified fields in the verified structured requirement state and the constraint relationship as metadata to the packaged file to generate a feasible assembly configuration, and output the feasible assembly configuration to the computer-aided design system.

[0153] If the verification fails, the conflict information output by the deterministic constraint solver is parsed to locate the functional role of the component corresponding to the conflict, including:

[0154] S5.7: When the unified verification result indicates that there are unmet verification rules, the verification is determined to fail and the conflict information output by the deterministic constraint solver is obtained; the conflict information includes the conflict type identifier, the part number involved in the conflict, and the specific constraint condition expression violated.

[0155] Furthermore, if any of the three verification rules contains a conflict record, the overall verification is deemed to have failed, and the complete conflict information output by the solver is read. Each conflict record within the conflict information carries a conflict type identifier, distinguishing between numerical boundary conflicts, interface type conflicts, and cross-component coupling conflicts; it also records the material numbers of one or more components involved in the conflict, as well as a complete description of the constraints violated by the conflict.

[0156] Furthermore, when the unified verification fails, if the conflict type is a numerical boundary conflict or an interface type conflict, an alternative solution retrieval step is also included: obtaining the conflict information output by the deterministic constraint solver, and parsing the conflict type identifier contained in the conflict information; when the conflict type identifier indicates a numerical boundary conflict, retrieving the parameter alternative model of the component from the material master database based on the component number involved in the conflict, wherein the parameter alternative model has different numerical parameters but maintains interface compatibility; when the conflict type identifier indicates an interface type conflict, retrieving whether there is an available interface adapter part number from the interface conversion parts library based on the component number involved in the conflict and the assembly connection diagram; if the parameter alternative model or the interface adapter part number exists, it is used as a supplementary candidate component, replacing the original conflicting component and added to the activated local candidate component number set.

[0157] Furthermore, the interface conversion parts library is a pre-built dedicated parts database. Its construction method is as follows: all adapter-type parts with interface conversion functions are screened and classified from the material master database of the enterprise resource planning system. The interface codes of both ends of the parts, the range of compatible parts models, assembly constraints, and working environment adaptation parameters are extracted, and discontinued and obsolete materials are filtered out. At the same time, the publicly available technical parameters of industry standard interface conversion parts are imported, and indexes are built according to the input interface type and output interface type to complete the storage, so that the available interface adapter part numbers can be quickly retrieved when conflicts occur.

[0158] S5.8: Based on the component numbers involved in the conflict, reverse query the functional role requirement mapping table to determine the component functional role corresponding to the component number involved in the conflict, and mark it as the conflicting functional role.

[0159] Furthermore, the system reads the component number from the conflict record, queries the material master database to obtain the component's functional role label, and then compares it with the functional role requirement mapping table to determine the component's functional role, marking that functional role as the conflicting functional role. Multiple conflict records may correspond to the same conflicting functional role; the system will deduplicate conflicting functional roles, marking the same functional role only once.

[0160] The method of invoking the language model to supplement component candidates only for the functional roles of the components corresponding to the conflict includes:

[0161] S5.9: Based on the conflict type identifier and the specific constraint expression of the violation, generate supplementary search keywords for the conflicting functional role;

[0162] Furthermore, the conflict type identifier and the description of the violated constraint are read, and the parameter restrictions in the constraints are extracted and concatenated to form supplementary search keywords for the conflicting functional roles. For example, if the conflict type is a numerical boundary conflict and the conflict is that the weight of the transmission component exceeds the overall weight constraint, the corresponding supplementary search keywords would be two-stage reduction transmission component, weight reduction, interface compatibility, and meeting the overall weight limit constraint of 180kg.

[0163] S5.10: Based on the supplementary search keywords, perform a secondary search from the material master database to obtain a set of supplementary candidate parts for the conflicting functional roles;

[0164] S5.11: Encapsulate the technical parameters of each component in the supplementary candidate component set into a new prompt word sequence, and call the pre-trained language model interface again to obtain the supplementary component candidate number list output by the language model;

[0165] Furthermore, the specific steps of S5.11 include: reading the technical parameters corresponding to all components in the supplementary candidate component set, and simultaneously reading the conflict constraints generated in S5.9, and organizing and arranging the two types of information; based on the organized component technical parameters and conflict constraints, assembling and generating a new prompt word sequence, fully incorporating the conflict-corresponding constraints into the prompt word sequence, clearly informing the language model of the constraint conflict issues that need to be avoided in this screening; reusing the pre-trained language model interface parameters that have been configured, setting the output temperature to 0.2, and setting the maximum number of output tokens per request to 1800; sending the assembled prompt word sequence into the pre-trained language model interface for execution, and receiving the text results returned by the interface; parsing and extracting the text returned by the interface, and outputting a list of supplementary component candidate numbers.

[0166] S5.12: The components in the supplementary component candidate number list are used as newly activated local candidate components, replacing or adding them to the activated local candidate component number set to form an updated local candidate component number set.

[0167] S6: Perform loop iterations until a feasible assembly configuration is output or the preset candidate supplement termination condition is met.

[0168] The specific steps of S6 include:

[0169] S6.1: Using the updated set of local candidate component numbers as new input, repeatedly execute the step of inputting the candidate components obtained through local dynamic activation into the deterministic constraint solver for unified verification;

[0170] S6.2: Update the loop counter at the start of each iteration; when the deterministic constraint solver outputs a result indicating that the verification failed, compare the current value of the loop counter with a preset maximum iteration threshold; if the current value is less than or equal to the maximum iteration threshold, continue to execute the step of parsing the conflict information output by the deterministic constraint solver and locating the functional role of the component corresponding to the conflict; if the current value is greater than the maximum iteration threshold, stop the iteration and determine that the candidate supplement termination condition has been met, and generate a diagnostic report containing all candidate components that have been traversed and the corresponding conflict reasons; if the deterministic constraint solver outputs a verification that passes in any iteration, immediately stop the iteration and output the feasible assembly configuration.

[0171] In this embodiment, the maximum iteration threshold is preset to 8 times. This setting is based on engineering experience; setting the iteration count too low provides insufficient opportunities to supplement candidate parts, while setting it too high causes the overall solution time to continuously increase, consuming a large amount of computing resources. The initial value of the loop counter is set to zero, and the value of the loop counter increases by 1 with each iteration. The solver returns a verification failure result, comparing the current value of the loop counter with the maximum iteration threshold of 8. If the counter value is less than or equal to 8, the process of conflict resolution, locating conflicting functional roles, and supplementing candidate parts continues. Once the loop counter value is greater than 8, the iteration loop is immediately stopped, a candidate supplementation termination condition is triggered, and a complete diagnostic report is generated. The diagnostic report fully records all candidate part material numbers traversed during the iteration process, as well as all conflict reasons, conflict-related fields, and conflicting part numbers that occurred in each iteration. If the solver returns a verification success in any iteration, the entire iteration process is terminated directly, and a feasible assembly configuration is output.

[0172] Furthermore, the preset candidate supplementation termination condition also includes a single iteration supplementation quantity upper limit condition: after each call to the language model to supplement component candidates, the number of newly supplemented component candidates in this iteration process is counted; a preset single iteration supplementation quantity upper limit threshold is obtained; the number of newly supplemented candidates is compared with the upper limit threshold; when the number of newly supplemented candidates exceeds the upper limit threshold, priority is sorted based on the correlation between each newly supplemented component candidate and the conflict information, and only the component candidates that are ranked first and whose quantity does not exceed the upper limit threshold are retained, while the remaining candidates that exceed the upper limit are discarded.

[0173] In this embodiment, the upper limit threshold for the number of supplementary candidates in a single iteration is set to 215. This threshold is set based on the fact that adding too many components in a single iteration will cause a sharp increase in the size of the solver's input fact set, significantly increasing the solution time. After each call to the language model to obtain the list of supplementary component candidate numbers, the total number of newly added candidate components is counted. The number of newly added candidate components is compared with the upper limit threshold of 215. If the number of newly added components is greater than 215, the correlation degree between each newly added candidate component and the current conflict information is calculated. The correlation degree is calculated by referring to the matching degree between the component's technical parameters and the conflict constraints; the higher the matching degree, the higher the correlation degree value. All newly added candidate components are prioritized according to their correlation degree values ​​from high to low. Only the top-ranked candidate components with a total number not exceeding 215 are retained, while the lower-ranked and excess candidate components are discarded and not added to the candidate set. If the number of newly added candidate components is less than or equal to 215, all are retained without discarding.

[0174] Furthermore, when the preset candidate supplementation termination condition is reached, the process also includes a constraint relaxation and feedback step: based on the generated diagnostic report containing all candidate components that have been traversed and their corresponding conflict causes, a priority-based conflict resolution strategy is used to analyze the conflict causes in the diagnostic report; based on the analysis results, the key conflict constraints causing the iteration termination are identified; relaxation calculations are performed on the key conflict constraints to determine the adjustable range of each constraint boundary value; the adjustable key conflict constraints and their corresponding relaxation ranges are fed back to the user interface, and constraint adjustment suggestions are generated; modification instructions input by the user based on the constraint adjustment suggestions are received, the constraint relationships in the structured requirement state that has been verified are updated based on the modification instructions, and the step of re-triggering the construction of the candidate pool for fields to be clarified is triggered based on the updated constraint relationships.

[0175] Furthermore, all conflict records in the diagnostic report are read. In this embodiment, the conflict priority ranking rule is that numerical boundary conflicts have higher priority than interface type conflicts, interface type conflicts have higher priority than cross-component coupling conflicts, and within the same conflict type, conflicts caused by fields with higher importance have higher priority. All conflict records are prioritized according to this conflict resolution strategy, and the ranked conflict record set is output. Based on the ranked conflict record set, the key conflict constraints causing the iteration termination are identified, and the key conflict constraint set is output. The defined historical design case library is called, and relaxation calculations are performed by referring to the parameter distribution range of similar equipment in the historical design case library. The reasonable adjustment range for the constraint boundary values ​​of the key conflict constraints is calculated, and the constraint adjustment range must not exceed the parameter range that has appeared in historical cases of similar equipment. The relaxed adjustment range corresponding to each key conflict constraint is output. The key conflict constraints, the relaxed adjustment range, and the constraint adjustment suggestions are pushed to the user interface, waiting for the user to input modification instructions. The user can choose to adjust the constraint boundary values ​​or maintain the original constraints without modification. Receive the modification instruction from the user, update the constraint relationships within the structured requirement state that has been verified based on the modification instruction, and output the updated structured requirement state that has been verified. Using the updated structured requirement state that has been verified as input, re-trigger the step of building the candidate pool of fields to be clarified, and return to the S2 process.

[0176] Example 2:

[0177] Please see Figure 3 Another embodiment of the present invention provides: a deterministic constraint solving system under local dynamic activation of a candidate pool, comprising:

[0178] The requirement structuring conversion module 10 is used to acquire unstructured mechanical design requirements input by users. Relying on natural language processing word segmentation, named entity recognition, rule matching, analytic hierarchy process, cost evaluation function and association rule mining algorithm, it outputs a structured requirement status containing field type, importance, acquisition cost, status label and constraint relationship, and completes the conversion of the original requirement from natural language text to a set of computable fields.

[0179] The question generation module 20 constructs a candidate pool of fields to be clarified based on the structured requirement status, calculates the first influence score and the second information gain score of the candidate fields to obtain two-step prospective value, selects the candidate field with the largest two-step prospective value as the field to be clarified, matches the clarification question template library to generate and output clarification questions, and realizes the priority screening of high-value fields to be clarified.

[0180] The user response verification processing module 30 is used to receive the response content returned by the user in response to the clarification question, complete the response content source record, call the corresponding type validator to perform type verification, extract the directly hard dependency related fields to perform hard constraint pre-verification, update the field status flag, and output the structured requirement status of verification completion after all fields to be clarified have been verified.

[0181] The candidate pool local dynamic activation module 40 is used to parse the structured requirement status after verification to build a functional role requirement mapping table, retrieve the material master database and generate an initial candidate component pool with controlled scale through a sampling algorithm; call the language model to obtain the component candidate number, and sequentially perform whitelist verification, functional role matching, usage scenario adaptation, and duplicate item removal to obtain the activated local candidate component number set, thus completing the local dynamic activation of the candidate pool.

[0182] The deterministic constraint solver verification module 50 is used to read the technical parameters of the activated local candidate parts, combine the verified fields with the constraint association relationship to encode the solver input fact set, call the deterministic constraint solver with built-in numerical boundary, interface compatibility, and cross-part coupling consistency verification rules to perform unified verification, and output the verification results, feasible assembly configuration or conflict information.

[0183] The conflict localization module 60 is used to parse the conflict information output by the deterministic constraint solver, reverse locate the functional role of the component corresponding to the conflict, generate supplementary search keywords, call the language model to supplement the component candidates for the conflict functional role, update the local candidate component number set, maintain the loop counter, compare the loop counter with the maximum iteration number threshold, control the iteration process, and output a diagnostic report when the termination condition is reached.

[0184] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments under the guidance of the present invention without departing from the spirit and scope of the present invention. All of these variations are within the protection scope of the present invention.

Claims

1. A method for solving deterministic constraints under local dynamic activation of a candidate pool, characterized in that, include: S1: Obtain unstructured mechanical design requirements input by the user and convert them into structured requirement states; S2: Construct a candidate pool of fields to be clarified based on the structured requirement status, select fields to be clarified from the candidate pool using type relationships and two-step look-ahead value, and output the clarification question; S3: Receive the response content returned by the user for the clarification question, perform source record, type validation and hard constraint pre-validation on the response content, and generate a structured requirement state that has completed the validation; S4: Based on the structured requirement status that has been verified, generate an initial candidate component pool with controlled size according to the functional roles of the components, and call the language model to output the component candidate number. Then, perform whitelist verification, functional role matching, usage scenario adaptation and duplicate item removal on the component candidate number to complete the local dynamic activation of the candidate pool. S5: Input the candidate components obtained after local dynamic activation into the deterministic constraint solver to perform unified verification; if the verification passes, output the feasible assembly configuration. If the verification fails, the conflict information output by the deterministic constraint solver is parsed, the functional role of the component corresponding to the conflict is located, and the language model is called only for the functional role of the component corresponding to the conflict to supplement the component candidates. S6: Perform loop iterations until a feasible assembly configuration is output or the preset candidate supplement termination condition is met.

2. The method for solving deterministic constraints under local dynamic activation of a candidate pool as described in claim 1, characterized in that, The specific steps of S1 include: Unstructured mechanical design requirements text input by users is obtained based on a natural language processing interface; Based on the unstructured mechanical design requirements, a natural language processing word segmentation algorithm is used to obtain the required word sequence; Based on the demand word sequence, entity words are extracted using a named entity recognition algorithm to generate an initial entity set; Based on the initial entity set, a rule-based matching method is used to map each entity word to a predefined field type, generating field type annotation results; Based on the field type labeling results, the importance of each field is calculated using the analytic hierarchy process (AHP). Based on importance and combined with pre-set field collection time thresholds, a cost evaluation function is used to calculate the acquisition cost; Based on the acquisition cost, a status flag is set according to the range of the cost value; the status flags include unprocessed, pending clarification, verified, and locked. Based on state labeling, the dependency relationships between fields are analyzed using association rule mining algorithms to establish constraint relationships and generate structured requirement states.

3. The method for solving deterministic constraints under local dynamic activation of a candidate pool as described in claim 2, characterized in that, The specific steps of S2 include: Based on the structured requirement status, fields marked as unprocessed or pending clarification are selected, and an initial candidate pool of fields pending clarification is constructed. For each candidate field in the initial candidate pool of fields to be clarified, the number of fields that directly and indirectly constrain the candidate field in the mechanical design constraint network is calculated and the first influence score is generated. Based on the field type and importance, complete cases similar to the candidate fields are retrieved from the historical design case library. The value distribution of the corresponding candidate fields in the similar complete cases is statistically analyzed and the expected reduction in information entropy is calculated to generate a second information gain score. The first influence score and the second information gain score are weighted and summed to calculate the two-step forward value of each candidate field. The candidate field with the largest two-step forward value is selected as the field to be clarified. Based on the field type of the field to be clarified and its constraint relationship, the corresponding question sentence is matched from the preset clarification question template library, the clarification question is generated and output.

4. The method for solving deterministic constraints under local dynamic activation of a candidate pool as described in claim 3, characterized in that, The specific steps of S3 include: Receive the response content returned by the user through the user interaction interface, and record the data source type of the response content; Based on the field type of the field to be clarified, the corresponding type validator is invoked to perform type validation on the response content. If the type validation fails, an error message is generated and the user is requested to re-enter the information. If the type validation passes, based on the constraint relationship, the associated fields with a direct hard dependency relationship with the field to be clarified and their hard constraint conditions are extracted, and hard constraint pre-validation is performed on the response content. If the hard constraint pre-validation passes, the response content is filled into the structured requirement status as a validated value, and the status mark of the corresponding field is updated to validated. If the hard constraint pre-validation fails, a conflict report is generated and returned to the user. After all fields to be clarified have passed type validation and hard constraint pre-validation, the structured requirement state that has been validated is generated.

5. The method for solving deterministic constraints under local dynamic activation of a candidate pool as described in claim 4, characterized in that, Based on the structured requirement status that has been verified, an initial candidate component pool of controlled size is generated according to the functional roles of the components, including: Parse the structured requirement status that has been verified and extract all verified fields and their verified values; Based on the preset mechanical function role classification tree, each verified field is matched with its corresponding component function role to generate a function role requirement mapping table. For each functional role in the functional role requirement mapping table, retrieve components that meet the technical parameter range of the corresponding functional role from the material master database of the enterprise resource planning system to form a candidate component set. A preset upper limit threshold for the size of the candidate pool is obtained. When the number of parts in any candidate part set exceeds the upper limit threshold for the size of the candidate pool, a maximum-minimum distance sampling algorithm based on space filling design is used to uniformly extract a subset of the candidate part set so that the size of the extracted candidate part set does not exceed the upper limit threshold for the size of the candidate pool, thereby generating an initial candidate part pool with controlled size.

6. The method for solving deterministic constraints under local dynamic activation of a candidate pool as described in claim 5, characterized in that, The process involves calling the language model to output candidate component numbers, then performing whitelist verification, functional role matching, usage scenario adaptation, and duplicate item removal on these candidate numbers to achieve partial dynamic activation of the candidate pool, including: The technical parameters of each component in the initial candidate component pool are encapsulated into a sequence of prompt words. The pre-trained language model interface is called to obtain the list of component candidate numbers corresponding to each functional role output by the language model. Based on a preset whitelist, remove numbers from the candidate part number list that are not in the whitelist to generate a first intermediate list; Based on the functional role requirement mapping table, retain the numbers of functional roles in the first intermediate list that match the verified fields, and generate a second intermediate list; Based on the working environment parameters in the structured requirement status that has been verified, the numbers that do not meet the working environment adaptability standards in the second intermediate list are removed, and a third intermediate list is generated. The third intermediate list is deduplicated, and the final retained numbers are used as the set of local candidate component numbers after activation.

7. The method for solving deterministic constraints under local dynamic activation of a candidate pool as described in claim 6, characterized in that, The step of inputting candidate components obtained through local dynamic activation into the deterministic constraint solver for unified verification includes: Obtain the activated set of local candidate component numbers, and based on the material master database, extract the detailed technical parameter table corresponding to each number in the activated set of local candidate component numbers; The detailed technical parameter table, all verified fields in the structured requirement state that has been verified, and the constraint relationships are jointly encoded into the input fact set of the deterministic constraint solver. The deterministic constraint solver is started; the deterministic constraint solver has a set of preset verification rules, including numerical boundary verification rules based on interval arithmetic, interface type compatibility verification rules based on the interface standard library, and cross-component coupling consistency verification rules based on the assembly connection diagram; the deterministic constraint solver performs numerical boundary verification, interface type compatibility verification, and cross-component coupling consistency verification in sequence based on the input fact set, and outputs a unified verification result.

8. The method for solving deterministic constraints under local dynamic activation of a candidate pool as described in claim 7, characterized in that, If the verification passes, a feasible assembly configuration will be output. If the verification fails, the conflict information output by the deterministic constraint solver is parsed to locate the functional roles of the components corresponding to the conflict, including: When the unified verification result indicates that all the verification rules are met, the verification is deemed to have passed; based on the activated local candidate component number set, the 3D model file and bill of materials file corresponding to each number are extracted from the product data management system; the 3D model file, bill of materials file and the activated component number set are associated and packaged, and all verified fields in the structured requirement state that has been verified and the constraint association relationship are attached as metadata to the packaged file to generate a feasible assembly configuration, and the feasible assembly configuration is output to the computer-aided design system; When the unified verification result indicates that there are unmet verification rules, the verification is determined to fail and the conflict information output by the deterministic constraint solver is obtained. The conflict information includes a conflict type identifier, the part number involved in the conflict, and the expression of the violated constraint condition. Based on the part number involved in the conflict, the functional role requirement mapping table is queried in reverse to determine the functional role of the part corresponding to the part number involved in the conflict and mark it as a conflict functional role.

9. The method for solving deterministic constraints under local dynamic activation of a candidate pool as described in claim 8, characterized in that, The method of invoking the language model to supplement component candidates only for the functional roles of the components corresponding to the conflict includes: Based on the conflict type identifier and the violated constraint expression, supplementary search keywords are generated for the conflicting functional roles; Based on the supplementary search keywords, a secondary search is performed from the material master database to obtain a set of supplementary candidate parts for the conflicting functional roles; The technical parameters of each component in the supplementary candidate component set are encapsulated into a new prompt word sequence, and the pre-trained language model interface is called again to obtain the supplementary component candidate number list output by the language model; The components in the supplementary component candidate number list are used as newly activated local candidate components, replacing or adding them to the activated local candidate component number set to form an updated local candidate component number set.

10. The method for solving deterministic constraints under local dynamic activation of a candidate pool as described in claim 9, characterized in that, The specific steps of S6 include: Using the updated set of local candidate component numbers as new input, the step of performing unified verification on the candidate component input deterministic constraint solver obtained through local dynamic activation is repeated. At the start of each iteration, the loop counter is updated. When the deterministic constraint solver outputs a result indicating that the verification failed, the current value of the loop counter is compared with a preset maximum iteration threshold. If the current value is less than or equal to the maximum iteration threshold, the process continues to parse the conflict information output by the deterministic constraint solver and locate the functional role of the component corresponding to the conflict. If the current value is greater than the maximum iteration threshold, the iteration is stopped and the candidate supplement termination condition is determined to have been met, generating a diagnostic report containing all candidate components that have been traversed and their corresponding conflict causes. If the deterministic constraint solver outputs a verification that passes in any iteration, the iteration is immediately stopped and the feasible assembly configuration is output.

11. A deterministic constraint solving system under local dynamic activation of a candidate pool, used to implement the deterministic constraint solving method under local dynamic activation of a candidate pool as described in any one of claims 1-10, characterized in that, include: The requirement structuring conversion module is used to acquire unstructured mechanical design requirements input by the user and output the structured requirement status. The question generation module constructs a candidate pool of fields to be clarified based on the structured requirement status, calculates the two-step look-ahead value, selects the candidate field with the largest two-step look-ahead value as the field to be clarified, and generates and outputs the clarification question by matching the clarification question template library. The user response verification processing module is used to receive the response content returned by the user in response to the clarification question and perform verification. After all fields to be clarified are verified, the structured requirement status of verification completion is output. The candidate pool local dynamic activation module parses the structured requirement status after verification, constructs a functional role requirement mapping table, generates an initial candidate component pool with controlled scale, and calls the language model to obtain the component candidate number to complete the local dynamic activation of the candidate pool. The deterministic constraint solution verification module is used to read the technical parameters of the activated local candidate parts, perform unified verification, and output the verification results, feasible assembly configurations, or conflict information. The conflict localization module is used to parse the conflict information output by the deterministic constraint solver, reverse locate the functional role of the component corresponding to the conflict, generate supplementary search keywords, call the language model to supplement the component candidates for the conflict functional role, and update the local candidate component number set.

Images