A knowledge reasoning method and system for design consultation

Abstract

The application relates to the technical field of design consultation, and discloses a knowledge reasoning method for design consultation, which comprises the following steps: acquiring diversified knowledge data and preprocessing to generate a benchmark knowledge base; extracting knowledge correlation features to determine reasoning rules; constructing reasoning chains and reasoning nodes; verifying the effectiveness of the reasoning nodes and distinguishing normal and abnormal reasoning states; reasoning in a hierarchical order in the normal state, identifying abnormal nodes, calculating reasoning deviation and performing path correction until the reasoning is completed in the abnormal state. The system comprises data acquisition, rule output, reasoning generation, state evaluation and reasoning optimization modules. The application improves the efficiency and accuracy of design consultation knowledge reasoning by integrating diversified data, scientifically formulating reasoning rules, dynamically optimizing reasoning paths and accurately processing abnormalities, and is suitable for intelligent reasoning decision in the field of design consultation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of design consultation, in particular to a knowledge reasoning method and system for design consultation. BACKGROUND

[0002] In the field of design consultation, with the rapid development of information technology and the increasing complexity of industry demand, how to efficiently integrate and utilize diversified knowledge data to achieve accurate and intelligent reasoning decision has become a key problem to be solved. Traditional knowledge reasoning methods have many limitations when dealing with complex knowledge systems in design consultation.

[0003] From the perspective of knowledge data acquisition and processing, design consultation requirements usually involve design requirement documents, historical case data, industry specification data and other types and sources of knowledge data. However, existing methods are difficult to systematically integrate these diversified data, with prominent data redundancy and conflict problems, and lack effective semantic standardization processing means, resulting in chaotic knowledge unit structure and inability to form a unified and standardized reference knowledge base, which seriously affects the accuracy and efficiency of subsequent reasoning.

[0004] In terms of knowledge association feature extraction and reasoning rule determination, traditional methods often ignore the semantic association degree and association strength difference between knowledge nodes, and cannot scientifically classify and prioritize knowledge association features. For example, the reasoning priority of aggregation features and dispersion features cannot be clearly distinguished, making the reasoning process lack hierarchy and logic, and making it difficult to develop reasonable reasoning rules based on the internal association characteristics of knowledge, resulting in unreasonable reasoning path construction and low reasoning efficiency.

[0005] For reasoning path construction and reasoning node management, existing technologies lack dynamic adjustment mechanisms when constructing reasoning chains and reasoning nodes. When abnormal reasoning nodes occur in the reasoning process, they cannot be identified and processed in a timely and accurate manner. For example, it is difficult to effectively calculate the reasoning deviation amount and correct the path according to the deviation amount, resulting in a reasoning process that is easily trapped in a stalemate and unable to guarantee the continuity and reliability of reasoning. At the same time, the effectiveness verification means for reasoning nodes is single, lacking a scientific verification model and parameter evaluation system, making it difficult to accurately judge the state of reasoning nodes in real time and discover abnormal situations in the reasoning process in a timely manner.

[0006] In addition, when dealing with abnormal reasoning nodes, traditional methods lack path correction strategies based on historical data and dynamic adjustment factors. It is difficult to analyze and correct based on multi-dimensional data such as the number of historical backtracking times and reasoning deviation amount of abnormal reasoning nodes, resulting in insufficient pertinence and effectiveness of abnormal handling, making it difficult to quickly restore normal reasoning flow and affecting the efficiency and quality of the entire design consultation knowledge reasoning. SUMMARY

[0007] The present application aims to provide a design consultation-oriented knowledge reasoning method and system to solve the problems in the background art.

[0008] To achieve the above-mentioned purpose, the present application provides the following technical solutions: a design consultation-oriented knowledge reasoning method, the method comprising:

[0009] Obtaining diversified knowledge data of design consultation requirements and performing knowledge preprocessing to generate a benchmark knowledge base;

[0010] Extracting knowledge correlation features of each benchmark knowledge base and determining reasoning rules of different knowledge bases according to the knowledge correlation features;

[0011] According to the reasoning rules, constructing a reasoning path for each benchmark knowledge base to form a plurality of reasoning chains and reasoning nodes of each reasoning chain;

[0012] Validating each reasoning node to obtain a reasoning state of each node, wherein the reasoning state comprises a normal reasoning state and an abnormal reasoning state;

[0013] In the normal reasoning state, the benchmark knowledge base under each reasoning node is correlated and reasoned in a hierarchical order;

[0014] In the abnormal reasoning state, an abnormal reasoning node is identified, and a reasoning deviation amount of the benchmark knowledge base under the abnormal reasoning node is calculated;

[0015] Based on the reasoning deviation amount, the abnormal reasoning node is corrected in a path until all the benchmark knowledge bases complete reasoning in a hierarchical order and the process is terminated.

[0016] Preferably, the obtaining of diversified knowledge data of design consultation requirements and the knowledge preprocessing to generate a benchmark knowledge base comprises:

[0017] Collecting design requirement documents, historical case data and industry specification data, integrating each into an independent knowledge unit, and assigning a unique identifier to each knowledge unit;

[0018] Performing redundancy filtering on the data in each independent knowledge unit to remove duplicate or conflicting knowledge entries;

[0019] Performing semantic standardization processing on the filtered independent knowledge units, unifying the data structure of the knowledge units, and outputting the independent knowledge units with unified structure as the benchmark knowledge base.

[0020] Preferably, the extraction of knowledge correlation features of each benchmark knowledge base comprises:

[0021] Obtain the knowledge nodes under each of the aforementioned benchmark knowledge bases and sort them according to their association strength;

[0022] Extract the semantic correlation between adjacent knowledge nodes and label it as a classification condition parameter;

[0023] Obtain a preset association threshold, and divide each classification condition parameter into intervals based on the association threshold to generate multiple association intervals;

[0024] The number of classification condition parameters within each of the associated intervals is counted and recorded as feature distribution parameters. Based on the feature distribution parameters, the knowledge association features of each benchmark knowledge base are determined.

[0025] The knowledge association features include aggregate features and dispersed features, and the reasoning priority of dispersed features is lower than that of aggregate features.

[0026] Preferably, determining the knowledge association features of each benchmark knowledge base based on the feature distribution parameters includes:

[0027] Obtain the feature distribution parameters under each of the aforementioned benchmark knowledge bases;

[0028] The feature distribution parameters under the same benchmark knowledge base are sorted from high to low according to their values, and the proportion weight of the feature distribution parameter with the highest value is calculated.

[0029] Obtain the weight threshold and compare the weight threshold with the proportion weight of the feature distribution parameter with the highest value;

[0030] If the weight percentage is greater than the weight threshold, then the benchmark knowledge base corresponding to the feature distribution parameter is determined to have an aggregation feature.

[0031] If the percentage weight is less than or equal to the weight threshold, then the benchmark knowledge base corresponding to the feature distribution parameter is determined to have a dispersion characteristic.

[0032] Preferably, determining the reasoning rules for different knowledge bases based on the knowledge association features includes:

[0033] Obtain the knowledge association features of each of the aforementioned benchmark knowledge bases;

[0034] The maximum value of the association interval corresponding to the aggregated feature is used as the hierarchical inference interval;

[0035] Summarize the hierarchical reasoning intervals of the benchmark knowledge base under all the aforementioned aggregated features, and arrange them in ascending order of interval value to generate a reasoning priority sequence;

[0036] Hierarchical reasoning is performed on each of the benchmark knowledge bases according to the reasoning priority sequence, and after the reasoning of the benchmark knowledge base under the aggregation feature is completed, the reasoning of the benchmark knowledge base under the dispersion feature is added.

[0037] The generation of the inference priority sequence includes:

[0038] Obtain the historical adjustment records of the hierarchical reasoning interval of the benchmark knowledge base under the aforementioned aggregation feature;

[0039] Extract the actual effective duration of the hierarchical reasoning interval in the historical reasoning path and calculate the deviation ratio between it and the preset duration;

[0040] A dynamic adjustment factor is generated based on the deviation ratio, and the current level inference interval is weighted and corrected according to the adjustment factor.

[0041] The weighted hierarchical inference intervals are aggregated, the inference priority sequence is rearranged, and the adjusted sequence is synchronized to the path construction of subsequent inference nodes.

[0042] Preferably, the validity verification of each of the inference nodes includes:

[0043] Obtain the reasoning results of the baseline knowledge base under each reasoning node, and perform vector mapping transformation to generate multiple verification vectors;

[0044] Call the preset verification function, input the verification vector into the verification function, and record the output value as the verification parameter;

[0045] Obtain the verification threshold and compare the verification parameters with the verification threshold;

[0046] If the verification parameter is less than the verification threshold, the inference node is determined to be in a normal inference state.

[0047] If the verification parameter is greater than or equal to the verification threshold, the inference node is determined to be in an abnormal inference state.

[0048] Preferably, calculating the inference bias of the baseline knowledge base under the abnormal inference node includes:

[0049] Obtain the actual reasoning nodes of the baseline knowledge base that have not completed the associated reasoning under the abnormal reasoning node;

[0050] Calculate the logical interval between the actual inference node and the abnormal inference node, and calibrate it as the real-time deviation.

[0051] The abnormal inference nodes are backtracked to extract the historical deviation of their associated historical nodes;

[0052] Obtain a preset fault tolerance threshold, and terminate the backtracking operation when the historical deviation is less than or equal to the fault tolerance threshold;

[0053] Call the deviation calculation function, input the real-time deviation and historical deviation into the calculation function, and record the output result as the inference deviation.

[0054] Preferably, the step of correcting the path of the abnormal inference node based on the inference deviation includes:

[0055] The historical backtracking count of the abnormal inference nodes is counted and recorded as adjustment reference values;

[0056] Obtain the adjustment threshold and compare the adjustment reference value with the adjustment threshold;

[0057] When the adjustment reference value is greater than or equal to the adjustment threshold, the abnormal inference node is logically offset according to the inference deviation, a corrected inference node is generated, and the correlation inference is performed under the corrected inference node.

[0058] When the adjustment reference value is less than the adjustment threshold, the inference deviation is continuously collected until the adjustment reference value reaches or exceeds the adjustment threshold, at which point path correction is triggered.

[0059] Preferably, the vector mapping transformation includes:

[0060] Extract semantic entities and relational attributes from the reasoning results and construct multidimensional feature vectors;

[0061] The multidimensional feature vectors are subjected to dimensionality reduction processing to generate a standardized mapping result in a low-dimensional vector space;

[0062] Based on the preset semantic weight matrix, the standardized mapping results are weighted and calculated to generate the verification vector, which is then input into the verification function.

[0063] Preferably, the present invention also includes a knowledge reasoning system for design consulting, the system comprising:

[0064] The data acquisition module is used to acquire diverse knowledge data related to design consulting needs, perform knowledge preprocessing, and generate a benchmark knowledge base.

[0065] The rule output module is used to extract the knowledge association features of each of the benchmark knowledge bases and determine the reasoning rules of different knowledge bases based on the knowledge association features.

[0066] The reasoning generation module is used to construct reasoning paths for each benchmark knowledge base according to the reasoning rules, forming multiple sets of reasoning chains, and reasoning nodes for each reasoning chain.

[0067] A state evaluation module is used to verify the validity of each inference node and obtain the inference state of each node.

[0068] Under the normal reasoning state, the baseline knowledge base under each reasoning node is associated and reasoned in hierarchical order.

[0069] In the abnormal reasoning state, abnormal reasoning nodes are identified, and the reasoning deviation of the benchmark knowledge base under the abnormal reasoning node is calculated.

[0070] The reasoning optimization module is used to correct the path of the abnormal reasoning node based on the reasoning deviation, until all the benchmark knowledge bases have completed the reasoning in hierarchical order and then the process terminates.

[0071] Compared with the prior art, the beneficial effects of the present invention are:

[0072] In the knowledge data processing stage, design requirement documents, historical case data, and industry standard data are collected and integrated into independent knowledge units, each assigned a unique identifier, effectively achieving systematic management of diverse knowledge data. Redundancy filtering removes duplicate or conflicting knowledge entries, ensuring the accuracy and consistency of the knowledge data. Semantic standardization unifies the data structure of knowledge units, generating a standardized and orderly benchmark knowledge base, laying a solid foundation for subsequent reasoning, avoiding reasoning errors caused by data chaos, and improving the usability of knowledge data and the reliability of the starting point for reasoning.

[0073] In terms of knowledge association feature extraction and inference rule determination, by acquiring knowledge nodes and sorting them according to their association strength, extracting semantic association degrees and dividing them into intervals, the aggregated and dispersed features of the benchmark knowledge base can be accurately identified. It is clearly stated that the inference priority of dispersed features is lower than that of aggregated features, making the inference process more hierarchical and logical. The determination of knowledge association features based on feature distribution parameters and weight thresholds ensures the scientific nature of the inference rule formulation. The introduction of historical adjustment records and the deviation ratio of actual effective duration to generate dynamic adjustment factors, and the weighted correction of hierarchical inference intervals, realizes the dynamic optimization of the inference priority sequence. This allows the inference rules to be flexibly adjusted according to the actual inference situation, improving the adaptability of the inference rules and the rationality of the inference process.

[0074] During the construction of the inference path, multiple inference chains and nodes are formed according to the inference rules, creating a clear inference framework. By validating the inference nodes and utilizing vector mapping transformation, preset verification functions, and verification thresholds, the normal or abnormal state of the inference nodes can be determined in real time and accurately. This provides strong support for timely detection of problems in the inference process and ensures the effectiveness of the inference path and the controllability of the inference process.

[0075] In handling abnormal inference nodes, the real-time deviation is obtained by calculating the logical interval between the actual inference node and the abnormal inference node. Combined with the historical deviation extracted from historical backtracking, the inference deviation is accurately calculated using a deviation measurement function, providing a quantitative basis for anomaly handling. Based on the comparison between the number of historical backtrackings and the adjustment threshold, different path correction strategies are adopted to achieve precise correction of abnormal inference nodes. Logical offsets generate corrected inference nodes and execute associated inference, or inference deviation is continuously collected until correction is triggered, ensuring the targeting and effectiveness of anomaly handling, improving the fault tolerance and recovery efficiency of the inference process, and guaranteeing the coherence and reliability of inference. Attached Figure Description

[0076] Figure 1 This is a schematic diagram illustrating the working principle of the knowledge reasoning method for design consulting described in this invention.

[0077] Figure 2 Design diagrams generated for the benchmark knowledge base;

[0078] Figure 3 Design diagram for determining knowledge association features;

[0079] Figure 4 Design diagram for validating inference nodes;

[0080] Figure 5 Design diagram for path correction of abnormal reasoning nodes. Detailed Implementation

[0081] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0082] Please see Figures 1-5 This invention relates to a knowledge reasoning method for design consulting, and the technical solution is described in detail below with reference to specific implementation methods.

[0083] The system collects diverse knowledge data from design consulting scenarios via interfaces, including design requirement documents, historical case data, and industry standard data. First, each type of data is integrated into independent knowledge units, each assigned a unique identifier for precise location and management. Then, redundancy filtering is performed on the data within each knowledge unit, removing duplicate or conflicting knowledge entries through data comparison algorithms. For example, different expressions of the same parameter in design specifications are calibrated using an industry standard library. After filtering, the knowledge units undergo semantic standardization to unify the data structure; for example, design parameters in different formats are converted to the standard format defined by XMLSchema. Finally, a standardized baseline knowledge base is output.

[0084] The system parses each baseline knowledge base, extracts knowledge nodes (such as design parameters, case features, and regulatory clauses), and calculates the association strength between nodes based on the cosine similarity algorithm, sorting them from high to low strength values. For adjacent knowledge nodes, semantic association degree is calculated using a semantic analysis model (such as Word2Vec) to generate classification condition parameters. A preset association threshold (such as 0.6) is used to divide the classification condition parameters into intervals, for example, high association interval (≥0.8), medium association interval (0.6-0.8), and low association interval (<0.6). The number of parameters in each interval is counted to form feature distribution parameters (such as 40% of parameters in the high association interval), which is used to determine whether the knowledge association feature is aggregated or dispersed (see subsequent implementation examples for specific judgment rules). Inference rules are determined based on feature type: hierarchical inference rules are used for knowledge bases corresponding to aggregated features, and supplementary inference rules are used for knowledge bases corresponding to dispersed features.

[0085] For a baseline knowledge base with aggregated features, the maximum value of the correlation interval is used as the hierarchical reasoning interval (e.g., if the maximum value of the high correlation interval is 0.9, then the interval is 0.9). A reasoning priority sequence is generated based on the interval value from smallest to largest. For example, if the hierarchical reasoning intervals for three aggregated knowledge bases are 0.5, 0.7, and 0.9 respectively, the priority sequence is Knowledge Base A (0.5) → Knowledge Base B (0.7) → Knowledge Base C (0.9). Reasoning chains are constructed according to the priority order, with each reasoning step corresponding to a reasoning node. Each node contains the knowledge unit and correlation rules for the current reasoning. For knowledge bases with distributed features, reasoning is added in parallel after the reasoning of the aggregated knowledge base is completed, forming a branched reasoning chain.

[0086] The system acquires the output of the inference nodes, extracts semantic entities and relational attributes, and constructs multi-dimensional feature vectors (such as design parameter vectors containing dimensions like material, size, and performance). Principal Component Analysis (PCA) is used to reduce the dimensionality of the vectors, mapping them to a lower-dimensional vector space. Then, a weighted calculation is performed using a preset semantic weight matrix (e.g., 30% weight for design specifications and 50% weight for case experience) to generate a validation vector. A preset validation function (such as an SVM classifier) ​​is called to evaluate the validation vectors, outputting validation parameters (such as classification probability values). A preset validation threshold (e.g., 0.7) is used; if the validation parameter is <0.7, the system is considered to be in a normal inference state; if it is ≥0.7, the system is considered to be in an abnormal inference state.

[0087] During normal reasoning, the baseline knowledge base under each reasoning node performs associated reasoning sequentially according to the reasoning priority sequence and hierarchical reasoning interval. For example, after knowledge base A in the priority sequence completes the reasoning of the current node, it triggers the reasoning of knowledge base B at an interval of 0.5 time units (which can be dynamically adjusted according to computing resources), and so on, forming a linearly progressive reasoning chain.

[0088] When an abnormal inference node is detected, the system first locates the baseline knowledge base where the inference process is incomplete, and calculates the logical interval (e.g., difference in inference steps) between the actual inference node and the abnormal node as the real-time deviation. Simultaneously, the system backtracks the abnormal node, extracting historical deviations from associated historical nodes, until the historical deviation is less than or equal to a preset fault tolerance threshold (e.g., 2 steps), at which point backtracking stops. A deviation calculation function (e.g., the mean squared error function) is called to combine the real-time deviation and historical deviation to obtain the inference deviation. Based on the deviation, path correction is performed: if the number of historical backtracking iterations is greater than or equal to an adjustment threshold (e.g., 3 times), the abnormal node is logically offset, generating a corrected inference node (e.g., skipping the current conflicting rule and enabling a backup rule), and inference continues under the corrected node; if the number of backtracking iterations is less than the adjustment threshold, the system continues to collect deviation data until the threshold is reached, triggering correction.

[0089] The present invention will be further described below with reference to Examples 1 to 5:

[0090] Example 1:

[0091] In the process of acquiring diverse knowledge data on design consulting needs and performing knowledge preprocessing to generate a benchmark knowledge base, the system specifically performs the following operations:

[0092] To facilitate the collection of diverse knowledge data, the system utilizes multi-source data interfaces to automatically collect design requirement documents, historical case data, and industry standard data. Design requirement documents contain text-based descriptions of user needs, such as product function definitions, performance requirements, and design constraints. The system uses Natural Language Processing (NLP) technology to parse the document content and extract key parameters and entity information using regular expressions, such as extracting key elements like "load capacity" and "500kg" from "design an industrial shelf with a load capacity ≥ 500kg". Historical case data originates from an internal case database and includes structured case numbers, project backgrounds, design process records, and final solution parameters. The system directly reads this data via database connection protocols (such as JDBC) to ensure the integrity and consistency of the case data. Industry standard data is integrated with external authoritative standards platforms (such as the National Standards Full-Text Public Disclosure System), with real-time updates of standard clauses via API interfaces. This covers design standards, safety regulations, and material selection rules, such as fire protection codes in architectural design and tolerance standards in mechanical design.

[0093] When integrating various types of data into independent knowledge units, the system assigns a unique identifier to each knowledge unit. The identifier uses a hierarchical coding structure, such as "D-20250528-001-01," where "D" represents the design requirement category, "20250528" is the data collection date, "001" is the sequence number under the major category, and "01" is the sub-item number of this knowledge unit within the major category. This coding rule facilitates rapid location, retrieval, and management of knowledge units, ensuring that data from different sources has a unique identifier in the baseline knowledge base.

[0094] After the initial integration of knowledge units, the system performs redundancy filtering on the data within each independent knowledge unit. Redundancy filtering consists of two stages: the first stage is vertical filtering, which uses a hash value comparison algorithm to detect duplicate records within the same knowledge unit. For example, a design requirements document may contain multiple identical performance indicator descriptions. The system calculates the hash value of each record and identifies and deletes records with completely identical hash values ​​as duplicates. The second stage is horizontal filtering, which cross-validates data from different knowledge units to identify conflicting entries. For example, if a design parameter in a historical case is set to "10mm," while industry standards require the parameter to be in the range of "15-20mm," the system marks the conflicting entry as pending and triggers a manual review process. Design consultants then confirm and delete the erroneous entry or adjust the parameter value.

[0095] In the semantic standardization process, the system unifies the data structure of filtered independent knowledge units based on a predefined knowledge meta-model. The knowledge meta-model adopts an Entity-Attribute-Value (EAV) architecture, abstracting various types of knowledge into entity objects. Each entity contains several attributes, each with a corresponding value. For example, "steel" is defined as an entity, with attributes including "grade," "yield strength," and "elongation," and attribute values ​​are standardized based on the data source. For unstructured text data in design requirement documents, Named Entity Recognition (NER) technology is used to extract entities and attributes. For example, from "select Q235B steel, yield strength not less than 235MPa," the entity "Q235B steel," the attribute "yield strength," and its value "≥235MPa" are identified and converted into the standard format of the EAV model. For structured data from historical cases, data mapping rules map the original fields to corresponding attributes in the knowledge meta-model. For example, the "design cycle days" field in the case is mapped to the "time dimension - cycle" attribute.

[0096] During semantic standardization, the system also needs to address terminology differences from different data sources. For example, design requirements may use "floor area ratio," while industry standards use "gross floor area ratio." The system establishes a terminology mapping table to associate synonyms or near-synonyms with unified standard terms, ensuring semantic consistency between different knowledge units. The terminology mapping table adopts a bidirectional mapping structure, supporting forward queries (from business terms to standard terms) and reverse queries (from standard terms to business terms), facilitating semantic conversion and association during reasoning.

[0097] After completing the above processing, the system outputs the unified, independent knowledge units as the baseline knowledge base. The baseline knowledge base is stored in a graph database (such as Neo4j), with knowledge units acting as nodes in the graph. These nodes form a knowledge network through semantic relationships (such as "belongs to," "associated with," and "constraints"). For example, design requirement nodes and industry standard nodes are connected through a "complies with" relationship, and historical case nodes and design parameter nodes are connected through a "contains" relationship. This storage structure facilitates the subsequent extraction of knowledge association features and the construction of reasoning paths.

[0098] During data acquisition and preprocessing, the system also features data quality monitoring. By setting monitoring indicators such as data integrity, consistency, and accuracy, it detects anomalies in the data acquisition and processing process in real time. For example, if the missing rate of key attribute values ​​for a knowledge unit exceeds a preset threshold (e.g., 20%), the system automatically triggers a supplementary data acquisition process, reminding relevant personnel to supplement the data via email or push notification. If the system detects that the semantically standardized knowledge unit data structure does not conform to the knowledge element model definition, it generates an error log and locates the problem node for technical personnel to troubleshoot and repair.

[0099] Furthermore, the system supports version management of the baseline knowledge base. Each time the knowledge base is updated (e.g., adding knowledge units, modifying attribute values, deleting conflicting entries), a new version number is automatically generated. The version number uses the format "major version number.minor version number.revision number" (e.g., V1.0.1), and records the update time, update content, and operator information. The version management function ensures that the evolution of knowledge can be traced during reasoning, and also supports rolling back to historical versions, facilitating the handling of reasoning anomalies caused by data updates.

[0100] Example 2:

[0101] In the process of extracting knowledge association features from various benchmark knowledge bases, the system analyzes the strength of knowledge node associations and determines feature types through the following steps:

[0102] The system performs structured parsing on each benchmark knowledge base, identifying knowledge nodes within it. The type of knowledge node is determined based on the content attributes of the knowledge base. For example, in a benchmark knowledge base for architectural design consulting, knowledge nodes may include specific concepts or parameters such as "fire separation distance," "sunlight standard," "concrete strength grade," and "unit area." For design requirement-based knowledge bases, nodes may involve functional requirements proposed by users (such as "seismic fortification intensity" and "intelligent system integration requirements"); for historical case knowledge bases, nodes contain key design elements from the cases (such as "structural form," "material selection," and "project cycle"); and nodes in industry standard knowledge bases are typically core indicators from standard clauses (such as "load values," "safety factor," and "environmental protection requirements").

[0103] The system calculates and sorts the association strength between each knowledge node. The association strength calculation is based on a semantic analysis model, using Word2Vec word vector technology to map each knowledge node into a vector in a high-dimensional space. The cosine similarity formula is used to calculate the cosine of the angle between any two node vectors, which serves as a quantitative indicator of association strength. For example, the vector for node A "fire separation distance" is [0.2, 0.5, 0.3, ...], and the vector for node B "building height" is [0.3, 0.4, 0.3, ...]. Their cosine similarity is 0.82, indicating a strong association. However, the cosine similarity between node C "apartment area" and node D "steel yield strength" is 0.35, indicating a weak association. After calculating the association strength of all node pairs, the system generates a list of node association strengths from high to low according to the cosine similarity value, clearly indicating the order of the closeness between nodes.

[0104] When extracting the semantic correlation between adjacent knowledge nodes, the system uses a correlation strength list as a basis to filter adjacent node pairs (i.e., nodes that are directly adjacent in the sorting) and further analyzes their semantic correlation attributes. The calculation of semantic correlation is combined with the domain ontology model, using conceptual relationships defined in the ontology (such as "subclass relationship," "attribute relationship," and "constraint relationship") to semantically annotate node pairs. For example, in the building code ontology, "fire separation distance" and "building height" are related through a "positive correlation," with a semantic correlation of 0.8; while "apartment area" and "room layout" are related through an "inclusion relationship," with a semantic correlation of 0.65. These semantic correlation values ​​serve as classification condition parameters for subsequent feature analysis.

[0105] The system presets a correlation threshold (e.g., 0.6) to divide classification condition parameters into intervals. The correlation threshold is set based on the experience of domain knowledge experts and can be dynamically adjusted according to different industries or design consulting types. Using 0.6 as the dividing point, the classification condition parameters are divided into high correlation intervals (≥0.8), medium correlation intervals (0.6-0.8), and low correlation intervals (<0.6). For example, if the semantic correlation values ​​of adjacent node pairs in a benchmark knowledge base are 0.85, 0.72, 0.58, and 0.35 respectively, they would be classified into the high correlation interval, medium correlation interval, low correlation interval, and low correlation interval, respectively.

[0106] The system counts the number of classification condition parameters within each association interval and generates feature distribution parameters. These parameters are presented as a percentage of the total number of parameters in each interval. For example, in the above example, the high-association interval has 25% (1 / 4) of parameters, the medium-association interval has 25% (1 / 4) of parameters, and the low-association interval has 50% (2 / 4) of parameters. These parameters reflect the distribution characteristics of the degree of association between knowledge nodes in the benchmark knowledge base and are a key basis for determining the type of knowledge association features.

[0107] When determining the type of knowledge association features, the system first obtains the feature distribution parameters of each benchmark knowledge base. It then sorts the feature distribution parameters within the same knowledge base from highest to lowest value and calculates the weight of the feature distribution parameter with the highest value. For example, if a knowledge base has feature distribution parameters of 40% for high-association intervals, 30% for medium-association intervals, and 30% for low-association intervals, the highest value after sorting is 40% of the high-association interval, and its weight is 40%. The system presets a weight threshold (e.g., 50%) and compares the weight of the highest value with the weight threshold: if the weight is greater than the weight threshold, the benchmark knowledge base is determined to have an aggregation characteristic, indicating that its knowledge nodes are mainly concentrated in high-association or medium-high-association intervals, with a compact knowledge structure that facilitates hierarchical reasoning; if the weight is less than or equal to the weight threshold, it is determined to have a dispersion characteristic, indicating that the knowledge nodes are relatively dispersed and the degree of association is low, requiring supplementary reasoning rules.

[0108] Taking a benchmark knowledge base for architectural design consulting as an example, suppose this knowledge base contains 100 pairs of adjacent knowledge nodes, of which 65 are highly correlated interval parameters (65%), 20 are moderately correlated interval parameters (20%), and 15 are lowly correlated interval parameters (15%). If the feature distribution parameters are sorted by value as 65%, 20%, and 15%, with the highest value being 65% and having a weight of 65%, which is greater than the preset weight threshold of 50%, then this knowledge base is determined to have an aggregation characteristic. Conversely, if a mechanical design knowledge base has a high-correlation interval parameter ratio of 40%, a moderately correlated interval ratio of 35%, and a low-correlation interval ratio of 25%, and the highest value of 40% is less than the weight threshold of 50%, then it is determined to have a dispersion characteristic.

[0109] Regarding the determination of reasoning priority, the baseline knowledge base with aggregated features is given a higher reasoning priority due to its tightly linked knowledge nodes and clear reasoning paths. The system uses the maximum value of the correlation interval corresponding to the aggregated feature as the hierarchical reasoning interval. For example, the maximum value of the high correlation interval in the aforementioned architectural design knowledge base is 0.85, and the hierarchical reasoning interval is set to 0.85 units (the unit can be defined as milliseconds, seconds, etc., according to the system's calculation cycle). The knowledge base with dispersed features has a lower reasoning priority than the knowledge base with aggregated features because of its loosely linked nodes and higher uncertainty in reasoning paths. It is usually started as a supplementary step after the aggregated reasoning is completed to provide multi-faceted reasoning support or verification.

[0110] Furthermore, the system supports dynamic updates to knowledge association features. When the content of the baseline knowledge base changes (such as adding knowledge units or modifying node attributes), it automatically triggers a recalculation of association strength and semantic relevance, updates feature distribution parameters and feature type determination results, and ensures that knowledge reasoning rules are always based on the latest knowledge base structure. This dynamic mechanism enables the system to adapt to the rapid updates and evolution of knowledge in the design consulting field, maintaining the accuracy and effectiveness of the reasoning process.

[0111] Example 3:

[0112] When determining inference rules for different knowledge bases based on knowledge association characteristics, the system dynamically generates and optimizes inference rules through the following process:

[0113] The system acquires the knowledge association features of each benchmark knowledge base and distinguishes between aggregated and dispersed features. For benchmark knowledge bases with aggregated features, the knowledge nodes are closely related and suitable for hierarchical reasoning rules, i.e., reasoning is performed sequentially according to the strength of the association between knowledge nodes. For knowledge bases with dispersed features, since the node associations are relatively loose, they are usually used as a supplementary reasoning step and are initiated after the aggregated reasoning is completed.

[0114] For a benchmark knowledge base with aggregated features, the system uses the maximum value of its corresponding association interval as the hierarchical inference interval. The maximum value of the association interval reflects the strongest association degree of knowledge nodes in that knowledge base. For example, the maximum value of the high association interval of a certain aggregated knowledge base is... Then the hierarchical reasoning interval is set to ,in For time unit conversion factors (e.g., when converting correlation values ​​to seconds), (milliseconds / unit correlation). This setting makes the inference interval positively correlated with the correlation strength between knowledge nodes; the stronger the correlation between nodes, the shorter the inference interval and the more compact the inference process.

[0115] Next, the system summarizes the hierarchical inference intervals of the benchmark knowledge base under all aggregated features. ( For knowledge base indexing, The initial inference priority sequence is generated by arranging the interval values ​​from smallest to largest. (in For example, the hierarchical reasoning intervals of the three aggregated knowledge bases are respectively... Second, Second, seconds, initial sequence is The corresponding reasoning order is Knowledge Base A → Knowledge Base B → Knowledge Base C.

[0116] To optimize the inference priority sequence, the system introduces a dynamic adjustment mechanism based on historical data. First, it retrieves historical adjustment records of hierarchical inference intervals from the baseline knowledge base under aggregated features. Each record includes a timestamp and the original interval value. Adjusted interval value and actual effective duration Then, extract the actual effective duration of the hierarchical reasoning interval in the historical reasoning path. And calculate its time with the preset duration. deviation ratio The calculation formula is:

[0117] ;

[0118] in, Indicates the deviation ratio (dimensionless). The actual effective duration (in seconds). The preset duration (in seconds) is used to determine the initial inference interval. As The default value.

[0119] Based on deviation ratio The system generates dynamic adjustment factors. The adjustment factor is calculated using a linear mapping method, for example, setting when... hour (No adjustment required) When hour (Maximum adjustment range), the intermediate value is determined by linear interpolation, that is (when (Time). This factor is used to weight and correct the inference interval of the current level; the corrected interval value... for:

[0120] ;

[0121] in, The interval for hierarchical reasoning before correction (unit: seconds). It is a dynamic adjustment factor (dimensionless).

[0122] Taking knowledge base X as an example, its initial hierarchical reasoning interval seconds, preset duration Seconds, the actual duration of a specific effective event in historical records. Seconds, then the deviation ratio Dynamic adjustment factor Corrected interval Second.

[0123] After correcting the hierarchical reasoning intervals for all aggregated knowledge bases, the system summarizes the weighted hierarchical reasoning intervals. Rearrange the reasoning priority sequence The sorting rule remains that the interval value is from smallest to largest. For example, the original initial sequence was... The interval of knowledge base X after correction becomes seconds, the initial interval of another knowledge base Y is seconds, after correction , , If the time is seconds, then the new sequence is The corresponding reasoning order is adjusted to Knowledge Base A → Knowledge Base Y → Knowledge Base X.

[0124] Adjusted inference priority sequence This will be synchronized into the path construction of subsequent inference nodes, ensuring that the entire inference process is executed based on real-time optimized rules. For example, when constructing the next inference node, the system prioritizes calling the knowledge base at the beginning of the sequence for inference, in order to improve inference efficiency and accuracy.

[0125] For a baseline knowledge base with distributed features, the inference rule is to initiate inference in an append-only manner after the inference of the knowledge base with aggregated features is completed. Append-only inference employs a parallel processing mode, simultaneously calling multiple distributed knowledge bases to cross-validate or supplement the current inference results. For example, after the aggregated inference of architectural design is completed, distributed knowledge bases such as structural safety codes and energy-saving design standards are initiated to conduct parallel evaluations of the safety and environmental friendliness of the design scheme, generating multi-dimensional inference results.

[0126] The system also features a conflict detection mechanism when executing inference rules. When inference results from different knowledge bases contradict each other (e.g., an aggregated knowledge base recommends a certain material model, while a decentralized knowledge base suggests another model), the system automatically triggers an evidence weight comparison. It assigns a weight value to each inference result based on factors such as the authority and timeliness of the knowledge base, prioritizing results with higher weight values. If the weight values ​​are the same, a conflict report is generated, prompting design consultants to intervene with manual decision-making.

[0127] Furthermore, the entire process of dynamically adjusting the inference rules is logged. The system records information such as the time of each adjustment, the object being adjusted, the interval before and after the adjustment, the deviation ratio, and the adjustment factor, forming an audit log. The audit log can be used to trace the evolution of the inference rules, facilitating subsequent optimization and debugging of the inference model.

[0128] Example 4:

[0129] In the process of validating the validity of each inference node and calculating the deviation under abnormal conditions, the system takes the design inference node of a certain mechanical component as an example to carry out specific operations. The task of this inference node is to perform inference based on the user's requirement of "designing the gear parameters of a certain transmission system", calling the material knowledge base, design specification knowledge base and historical case knowledge base, and outputting the result "the gear material is selected as 45 steel, module m=3, number of teeth z=20, and tooth width b=60mm".

[0130] First, a vector mapping transformation is performed on the reasoning results. The system extracts semantic entities and relational attributes from the reasoning results: semantic entities include "45 steel", "module m=3", "number of teeth z=20", and "tooth width b=60mm"; relational attributes include "material-performance matching" (the matching of 45 steel with gear load requirements), "parameter-design specification association" (the conformity of module and number of teeth with mechanical design manual standards), and "parameter-case experience association" (the comparison relationship between tooth width values ​​and similar historical cases). Based on these entities and attributes, a multi-dimensional feature vector is constructed, with vector dimensions corresponding to material type, module, number of teeth, tooth width, and various association relationships, for example, represented as [material, module, number of teeth, tooth width, material performance matching degree, parameter specification conformity, parameter case association degree].

[0131] Dimensionality reduction is performed on the multidimensional feature vectors. Since high-dimensional vectors may contain redundant information, the system employs Principal Component Analysis (PCA) to map the data from the original high-dimensional space to a low-dimensional vector space through linear transformation, while preserving the main feature information. Assuming the original vectors are 7-dimensional, after dimensionality reduction and mapping to a 2-dimensional space, a standardized mapping result of (0.6, 0.7) is obtained. This result reflects the comprehensive performance of the inference results across the key feature dimensions.

[0132] The standardized mapping results are weighted according to a pre-defined semantic weight matrix. This semantic weight matrix is ​​pre-set based on the importance of knowledge within the design consulting domain; for example, material performance matching has a weight of 40%, parameter specification compliance has a weight of 30%, and parameter case relevance has a weight of 30%. The dimensionality-reduced mapping result (0.6, 0.7) is mapped to the principal components of the weight matrix. Assuming the first principal component corresponds to the combined characteristics of material performance and parameter specifications (70% weight), and the second principal component corresponds to parameter case characteristics (30% weight), the weighted calculation process is as follows: Validation vector value = 0.6 × 0.7 + 0.7 × 0.3 = 0.63. This value serves as the final vector value input to the validation function.

[0133] The system calls a preset verification function to evaluate the verification vector. This function uses a Support Vector Machine (SVM) classification model, trained on historical inference data, capable of distinguishing between normal and abnormal inference results. Inputting the verification vector (0.63) into the function outputs a verification parameter of probability 0.6, representing the probability that the inference result is within the normal range. The system's preset verification threshold is 0.7. Since 0.6 < 0.7, the inference node is determined to be in a normal inference state, allowing subsequent inference to continue in hierarchical order, such as triggering the gear strength verification knowledge base to perform mechanical analysis on the current parameters.

[0134] If an inference node outputs an abnormal value, taking another mechanical design node as an example, its inference result is "gear material selected as 20 steel, module m=2, number of teeth z=15, tooth width b=30mm". The semantic entities "20 steel", "module m=2", "number of teeth z=15", and "tooth width b=30mm" are extracted, along with the relational attributes "material-performance matching" (20 steel has low strength and may not meet load-bearing requirements) and "parameter-design specification association" (module 2 is less than the industry recommended minimum value of 2.5). After constructing a multi-dimensional feature vector, a standardized mapping result (0.4, 0.9) is obtained through dimensionality reduction. Combined with the semantic weight matrix (material performance weight 40%, specification compliance weight 30%), the verification vector value is calculated to be 0.4×0.7+0.9×0.3=0.55. After inputting into the SVM verification function, the output verification parameter is 0.8, which is greater than the preset threshold of 0.7, indicating an abnormal inference state.

[0135] Upon entering the abnormal state handling process, the system first calculates the inference deviation of the baseline knowledge base under the abnormal inference node. The knowledge bases with incomplete correlation inference are identified as the material knowledge base and the design specification knowledge base. The actual inference node is the current abnormal node (step 3). According to the inference rules, the inference of the material knowledge base should be completed in step 1, and the inference of the design specification knowledge base should be completed in step 2. Therefore, the logical intervals between the actual inference node and the abnormal node are: material knowledge base interval = 3 - 1 = 2 steps, design specification knowledge base interval = 3 - 2 = 1 step. The maximum value of 2 is taken as the real-time deviation.

[0136] For abnormal reasoning nodes, historical backtracking is performed to extract the historical deviation amount of their associated historical nodes. Assuming that the historical record shows that the deviation amount of similar material selection questions in the previous two reasonings was 1 step and 1.5 steps respectively, and the system's preset fault tolerance threshold is 2 steps, since 1≤2 and 1.5≤2, the backtracking continues to earlier historical nodes. If the deviation amount of the previous one is found to be 2.5 steps (>2), the backtracking stops. The final extracted historical deviation amounts are 1 step and 1.5 steps for the two most recent ones.

[0137] Call the deviation calculation function (e.g., take the arithmetic mean), input the real-time deviation 2 and the historical deviations 1 and 1.5 into the function, and calculate the inference deviation = (2 + 1 + 1.5) / 3 ≈ 1.5. This deviation reflects the degree of deviation of the abnormal node from the inference path, providing a quantitative basis for subsequent path correction.

[0138] Throughout the verification and bias calculation process, the system stores the input and output data for each step in a structured manner, including the original inference results, multi-dimensional feature vectors, dimensionality-reduced mapping results, verification vector values, verification parameters, and the bias calculation process. This data can be used for subsequent training and optimization of the inference model. For example, by analyzing the feature vector distribution of abnormal nodes, the semantic weight matrix or the parameters of the verification function can be adjusted to improve the system's accuracy in identifying abnormal states.

[0139] In addition, the system supports dynamic updates to the validation function. After accumulating a certain number of anomalous inference cases, the model retraining process is automatically triggered, using new data to retrain the SVM classifier, update the classification boundaries, and ensure that the validation function can adapt to new inference patterns or changes in data distribution.

[0140] Example 5:

[0141] During the path correction process for abnormal inference nodes, the system takes a specific inference scenario of a building structure design as an example. Assume that in the inference node for the design of the column cross-section dimensions of a frame structure, the system initially calls the building load code knowledge base, the concrete structure design code knowledge base, and the historical case knowledge base, outputting the result "column cross-section dimensions 400mm×400mm, concrete strength grade C30". However, after validity verification, it is determined to be an abnormal inference state (verification parameter 0.85 ≥ preset threshold 0.7).

[0142] The system first counts the historical backtracking count of the abnormal inference node. Assuming this node is the first time the anomaly has been triggered, the historical backtracking count is 0, and the adjusted reference value record is 0. The system presets an adjustment threshold of 3 times. Since 0 < 3, it enters the continuous collection phase of inference deviation. At this point, the system recalculates the real-time deviation: the knowledge base for incomplete association inference is located as the seismic design code knowledge base (because the seismic fortification intensity of this region is 7 degrees, the column axial compression ratio needs to be verified), the actual inference node is the current abnormal node (step 4), and the seismic design code inference should be performed after the load code and concrete code, and before the case inference (i.e., step 3). Therefore, the logical interval is 4-3=1 step, which is used as the real-time deviation. Simultaneously, historical nodes are backtracked, and it is found that this type of anomaly has not appeared in historical inference, so the historical deviation is 0. Therefore, the inference deviation equals the real-time deviation of 1. Since the adjusted reference value has not reached the threshold, the system does not immediately trigger path correction, but continues to monitor the subsequent inference process, accumulating deviation data.

[0143] When the abnormal inference node triggers an anomaly for the second time, the historical backtracking count is updated to 1. Since the adjusted reference value is still less than the adjustment threshold of 3, the system continues to collect deviation data. Assuming the second real-time deviation is 1.5 steps (not called due to the newly added wind load code knowledge base), and the accumulated historical deviation is 1 step from the previous one, the inference deviation is calculated as (1+1.5) / 2=1.25. At this point, the system still does not perform a correction but instead stores the two deviation data in the historical record for analyzing anomaly patterns.

[0144] When the abnormal inference node triggers an anomaly for the third time, the historical backtracking count reaches 3 times (equal to the adjustment threshold), and the system triggers the path correction process. At this time, the real-time deviation is 2 steps (due to the disordered order of calling multiple knowledge bases caused by incorrect structural type judgment), the historical deviation is 1 step and 1.5 steps of the previous two, and the inference deviation is (2+1+1.5) / 3≈1.5. The system performs logical offset on the abnormal inference node based on the inference deviation: First, it analyzes the source of the deviation and finds that it is due to the failure to prioritize the call to the seismic design code knowledge base, resulting in missing parameters. Therefore, a corrected inference node is generated, and the inference order is adjusted to: Building Load Code Knowledge Base → Seismic Design Code Knowledge Base → Concrete Structure Design Code Knowledge Base → Historical Case Knowledge Base. Under the corrected inference node, the associated inference is re-executed: Based on the seismic fortification intensity of 7 degrees, the seismic design code knowledge base is called to calculate the column axial compression ratio limit. Combined with the axial force design value obtained from the load code, the column cross-section size is recalculated to be 500mm×500mm, and the concrete strength grade is maintained at C30. The validity was verified again. The semantic entities “500mm×500mm” and “C30” and the relational attributes “seismic calculation compliance” and “axial compression ratio rationality” were extracted. After constructing a multi-dimensional feature vector and reducing its dimensionality, the verification vector value was calculated to be 0.68. The input verification function output parameter was 0.65 < 0.7, which was determined to be a normal reasoning state. The subsequent reasoning steps, such as the floor stiffness ratio calculation, continued.

[0145] If, before the adjusted reference value reaches the threshold, the system detects a significant change in the deviation pattern of the abnormal inference node (e.g., from a parameter missing anomaly to a rule conflict anomaly), it automatically resets the historical backtracking count to avoid invalid corrections due to differences in anomaly types. For example, in a mechanical design anomaly node, the first two anomalies were triggered by material parameter conflicts (historical deviations were 1.2 and 1.4 respectively). The third anomaly was caused by a failure to synchronize user requirement changes, resulting in a deviation in the inference target (real-time deviation of 2.0). In this case, the system identifies it as a new anomaly type, clears the historical backtracking count to zero, and re-accumulates and adjusts the reference value.

[0146] During the path correction process, the system's logical offset operation is based on a preset backup rule base. This backup rule base contains alternative inference paths or supplementary rules corresponding to each benchmark knowledge base. For example, when the original design specification rules conflict with case experience, it can switch to industry-recommended empirical formula rules or expert consensus rules. Taking bridge design as an example, if the original inference rule triggers an anomaly due to the bridge span parameters exceeding the applicable scope of the specification, the logical offset can call the rules in the special design guidelines for long-span bridges to generate a corrected inference node: "Use finite element analysis to replace simplified calculations."

[0147] The system meticulously records each path correction operation, including correction time, anomaly node identifier, inference deviation, adjusted reference value, and a comparison of the inference path before and after correction, forming a correction log. This log can be used by design consultants to trace the inference process, analyze the causes of anomalies, and provide data support for optimizing the structure of the baseline knowledge base or inference rules. For example, by statistically analyzing the frequency of a certain type of anomaly, it can be determined whether there are missing knowledge entries or rule conflicts in the baseline knowledge base, thereby triggering the knowledge base update process.

[0148] Furthermore, path correction supports manual intervention. When multiple possible solutions exist for the correction inference node generated by the system (e.g., different backup rules lead to different correction paths), a prompt box will pop up for design consultants to choose from, ensuring that the correction meets the actual engineering needs. For example, in the correction of abnormalities in building energy-saving design, the system may provide two correction options: "adjust the thickness of the external wall insulation material" or "change the airtightness level of the window type," allowing personnel to make a decision based on project cost and schedule requirements.

[0149] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0150] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A knowledge reasoning method for design consulting, characterized in that: include: Acquire diverse knowledge data related to design consulting needs, perform knowledge preprocessing, and generate a benchmark knowledge base; Extract the knowledge association features of each of the benchmark knowledge bases, and determine the reasoning rules of different knowledge bases based on the knowledge association features; Based on the inference rules, inference paths are constructed for each benchmark knowledge base to form multiple inference chains and inference nodes for each inference chain. The validity of each inference node is verified to obtain the inference state of each node, wherein the inference state includes normal inference state and abnormal inference state. Under the normal reasoning state, the baseline knowledge base under each reasoning node is associated and reasoned in hierarchical order. In the abnormal reasoning state, abnormal reasoning nodes are identified, and the reasoning deviation of the benchmark knowledge base under the abnormal reasoning node is calculated. Based on the inference deviation, the abnormal inference nodes are corrected according to their paths until all the benchmark knowledge bases have completed the inference in hierarchical order. The extraction of knowledge association features from each of the benchmark knowledge bases includes: Obtain the knowledge nodes under each of the aforementioned benchmark knowledge bases and sort them according to their association strength; Extract the semantic correlation between adjacent knowledge nodes and label it as a classification condition parameter; Obtain a preset association threshold, and divide each classification condition parameter into intervals based on the association threshold to generate multiple association intervals; The number of classification condition parameters within each of the associated intervals is counted and recorded as feature distribution parameters. Based on the feature distribution parameters, the knowledge association features of each benchmark knowledge base are determined. The knowledge association features include aggregate features and dispersed features, and the reasoning priority of the dispersed features is lower than that of the aggregate features. The process of acquiring diverse knowledge data on design consulting needs, performing knowledge preprocessing, and generating a benchmark knowledge base includes: Collect design requirements documents, historical case data, and industry standard data, integrate them into independent knowledge units, and assign a unique identifier to each knowledge unit; The design requirements document covers a text-based description of user requirements, including product function definitions, performance requirements, and design constraints. The historical case data comes from an internal case database, which contains structured case numbers, project backgrounds, design process records, and final solution parameters. The data is read directly through a database connection protocol. The industry standard data is updated in real time via API interface, and the standard clauses are fire protection standards in the field of building design. Redundancy filtering is performed on the data within each independent knowledge unit to remove duplicate or conflicting knowledge entries; The filtered independent knowledge units are semantically standardized to unify the data structure of the knowledge units, and the independent knowledge units with unified structure are output as the benchmark knowledge base. The benchmark knowledge base is stored in the form of a graph database, with knowledge units serving as nodes in the graph; The process of determining the knowledge association features of each benchmark knowledge base based on the feature distribution parameters includes: Obtain the feature distribution parameters under each of the aforementioned benchmark knowledge bases; The feature distribution parameters under the same benchmark knowledge base are sorted from high to low according to their values, and the proportion weight of the feature distribution parameter with the highest value is calculated. Obtain the weight threshold and compare the weight threshold with the proportion weight of the feature distribution parameter with the highest value; If the weight percentage is greater than the weight threshold, then the benchmark knowledge base corresponding to the feature distribution parameter is determined to have an aggregation feature. If the percentage weight is less than or equal to the weight threshold, then the benchmark knowledge base corresponding to the feature distribution parameter is determined to have a dispersion characteristic.

2. The knowledge reasoning method for design consulting according to claim 1, characterized in that: The step of determining the reasoning rules for different knowledge bases based on the knowledge association features includes: Obtain the knowledge association features of each of the aforementioned benchmark knowledge bases; The maximum value of the association interval corresponding to the aggregated feature is used as the hierarchical inference interval; Summarize the hierarchical reasoning intervals of the benchmark knowledge base under all the aforementioned aggregated features, and arrange them in ascending order of interval value to generate a reasoning priority sequence; Hierarchical reasoning is performed on each of the benchmark knowledge bases according to the reasoning priority sequence, and after the reasoning of the benchmark knowledge base under the aggregation feature is completed, the reasoning of the benchmark knowledge base under the dispersion feature is added. The generation of the inference priority sequence includes: Obtain the historical adjustment records of the hierarchical reasoning interval of the benchmark knowledge base under the aforementioned aggregation feature; Extract the actual effective duration of the hierarchical reasoning interval in the historical reasoning path and calculate the deviation ratio between it and the preset duration; A dynamic adjustment factor is generated based on the deviation ratio, and the current level inference interval is weighted and corrected according to the adjustment factor. The weighted hierarchical inference intervals are aggregated, the inference priority sequence is rearranged, and the adjusted sequence is synchronized to the path construction of subsequent inference nodes.

3. The knowledge reasoning method for design consulting according to claim 1, characterized in that: The validity verification of each of the inference nodes includes: Obtain the reasoning results of the baseline knowledge base under each reasoning node, and perform vector mapping transformation to generate multiple verification vectors; Call the preset verification function, input the verification vector into the verification function, and record the output value as the verification parameter; Obtain the verification threshold and compare the verification parameters with the verification threshold; If the verification parameter is less than the verification threshold, the inference node is determined to be in a normal inference state. If the verification parameter is greater than or equal to the verification threshold, the inference node is determined to be in an abnormal inference state.

4. The knowledge reasoning method for design consulting according to claim 1, characterized in that: The calculation of the inference bias of the baseline knowledge base under the abnormal inference node includes: Obtain the actual reasoning nodes of the baseline knowledge base that have not completed the associated reasoning under the abnormal reasoning node; Calculate the logical interval between the actual inference node and the abnormal inference node, and calibrate it as the real-time deviation. The abnormal inference nodes are backtracked to extract the historical deviation of their associated historical nodes; Obtain a preset fault tolerance threshold, and terminate the backtracking operation when the historical deviation is less than or equal to the fault tolerance threshold; Call the deviation calculation function, input the real-time deviation and historical deviation into the calculation function, and record the output result as the inference deviation.

5. The knowledge reasoning method for design consulting according to claim 4, characterized in that: The step of correcting the path of the abnormal inference node based on the inference deviation includes: The historical backtracking count of the abnormal inference nodes is counted and recorded as adjustment reference values; Obtain the adjustment threshold and compare the adjustment reference value with the adjustment threshold; When the adjustment reference value is greater than or equal to the adjustment threshold, the abnormal inference node is logically offset according to the inference deviation, a corrected inference node is generated, and the associated inference is performed under the corrected inference node. When the adjustment reference value is less than the adjustment threshold, the inference deviation is continuously collected until the adjustment reference value reaches or exceeds the adjustment threshold, at which point path correction is triggered.

6. The knowledge reasoning method for design consulting according to claim 3, characterized in that: The vector mapping transformation includes: Extract semantic entities and relational attributes from the reasoning results and construct multidimensional feature vectors; The multidimensional feature vectors are subjected to dimensionality reduction processing to generate a standardized mapping result in a low-dimensional vector space; Based on the preset semantic weight matrix, the standardized mapping results are weighted and calculated to generate the verification vector, which is then input into the verification function.

7. A knowledge reasoning system for design consulting, characterized in that: The knowledge reasoning method for design consulting as described in any one of claims 1 to 6 includes: The data acquisition module is used to acquire diverse knowledge data related to design consulting needs, perform knowledge preprocessing, and generate a benchmark knowledge base. The rule output module is used to extract the knowledge association features of each of the benchmark knowledge bases and determine the reasoning rules of different knowledge bases based on the knowledge association features. The reasoning generation module is used to construct reasoning paths for each benchmark knowledge base according to the reasoning rules, forming multiple sets of reasoning chains, and reasoning nodes for each reasoning chain. A state evaluation module is used to verify the validity of each inference node and obtain the inference state of each node. Under the normal reasoning state, the baseline knowledge base under each reasoning node is associated and reasoned in hierarchical order. In the abnormal reasoning state, abnormal reasoning nodes are identified, and the reasoning deviation of the benchmark knowledge base under the abnormal reasoning node is calculated. The reasoning optimization module is used to correct the path of the abnormal reasoning node based on the reasoning deviation, until all the benchmark knowledge bases have completed the reasoning in hierarchical order and then the process terminates.

Images