A self-adaptive procedure generation method based on case-based reasoning and transfer learning

Abstract

The application provides a self-adaptive procedure generation method based on case-based reasoning and transfer learning, and the application constructs an excellent case knowledge base, structures and makes implicit knowledge, i.e., successful experience of experts verified in history, explicit, and innovatively combines the case-based reasoning and the transfer learning thought, so that the system can not only find the most similar historical case, but also intelligently and self-adaptively correct the scheme in a "transfer" mode according to the difference between the new task and the historical case, thereby automatically generating the engineering procedure which is highly customized, close to the actual situation and has the "design thought" annotation. The application solves the technical problems of low efficiency and poor consistency of the traditional method, and the technical problems of rigidity and lack of learning evolution ability of the existing software system.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of artificial intelligence and engineering applications, specifically involving an adaptive procedure generation method based on case reasoning and transfer learning. Background Technology

[0002] In high-risk and highly complex industrial fields such as mining, it is crucial to develop precise, safe, and efficient operating procedures. Taking mine ventilation procedure design as an example, it is defined as follows: based on a series of input conditions such as the mine's geological conditions, roadway layout, gas emission rate, and equipment parameters, a complete ventilation scheme is designed through complex engineering calculations and safety checks. This scheme includes air volume, air velocity, fan selection, and airflow control to ensure the safety of the underground working environment.

[0003] The existing technologies for solving this type of problem mainly include the following:

[0004] Human expert design: This method relies entirely on experienced engineers for manual design and calculation. It is the most traditional and currently the most mainstream approach. Its drawbacks include: the design process heavily depends on individual "tacit knowledge," which is difficult to quantify, pass on, and replicate on a large scale; design efficiency is low and the cycle is long; it is prone to oversights or misjudgments when facing new situations or scenarios not covered by experience; and the degree of optimization of the solution depends entirely on the individual engineer's skill level.

[0005] Fixed-rule calculation software: Some computer-aided design software exists on the market. After the user inputs parameters, the software performs calculations based on preset physical formulas (such as fluid mechanics and thermodynamics formulas) and industry standards. Its shortcomings are: this type of software is essentially an "advanced calculator," with fixed and rigid models and rules. It cannot handle the uniqueness of each project (such as each mine) and the subtle differences in complex working conditions, lacking adaptability. It cannot learn and evolve from a large number of historical successes and failures, resulting in solutions that are often only "theoretically feasible" rather than "practically optimal."

[0006] Traditional document retrieval systems: Enterprises may have internal document libraries where engineers can search for similar procedures and documents from the past using keywords. The drawback is that this only provides "reference materials," not "solutions." Engineers need to expend significant effort to read, understand, and manually "migrate" usable portions of historical solutions, resulting in an extremely high cognitive load for knowledge application and failing to achieve intelligent processing.

[0007] In summary, existing technologies either rely too heavily on personal experience that is difficult to pass on, or are too rigid and inflexible to adapt to complex and ever-changing real-world working conditions. There is a lack of an effective method to transform historical successful experiences into calculable, adaptive, and dynamic knowledge, and to automatically generate optimal procedures. Summary of the Invention

[0008] The present invention aims to at least partially solve one of the technical problems in the related art.

[0009] Therefore, the first objective of this invention is to propose an adaptive procedure generation method based on case reasoning and transfer learning.

[0010] The second objective of this invention is to propose an adaptive procedure generation device based on case reasoning and transfer learning.

[0011] To achieve the above objectives, a first aspect of the present invention proposes an adaptive procedure generation method based on case reasoning and transfer learning, comprising:

[0012] S1. Construct a structured knowledge base of excellent cases that includes input conditions, process parameters, final solutions and operational effects. Use natural language processing and information extraction technology to parse historical engineering code documents and transform implicit engineering experience into computable explicit knowledge.

[0013] S2, parse the input conditions of the new engineering task, and retrieve historical cases in the knowledge base based on the multi-dimensional similarity calculation model. The multi-dimensional similarity model integrates the normalized distance calculation of numerical features, the precise matching of categorical features, and the semantic vector calculation of text features to generate a comprehensive similarity score and return the reference case with the highest matching degree.

[0014] S3. Based on the difference in input conditions between the new engineering task and the reference case, the solution of the reference case is adaptively adjusted using a hybrid reasoning mechanism driven by parameter dependency graph. The hybrid reasoning mechanism combines the engineering physical model and the preset heuristic rules to perform a chain reaction correction along the parameter dependency relationship network until the adjusted solution with parameter convergence is generated.

[0015] S4 generates an engineering specification document containing the adjusted solution, and adds experience annotations, reference cases, and adjustment basis. The new specification solution that has been verified as excellent in actual application is automatically uploaded back to the knowledge base as a high-quality case, forming a knowledge iteration closed loop.

[0016] In one embodiment of the present invention, S1 includes:

[0017] S11, the input conditions include roadway geometry parameters, gas emission rate, equipment parameters and ventilation method;

[0018] S12, the process parameters include the calculated air volume, the calculated resistance, the selection of the safety factor, and the selection of the fan model.

[0019] In one embodiment of the present invention, S2 includes:

[0020] S21, the formula for calculating the normalized distance of numerical features is:

[0021]

[0022] Where V new,j and V case,j These are the values ​​of the new task and the historical case on feature j, respectively, max(V j ) and min(V j ) represents the maximum and minimum values ​​of this feature across the entire knowledge base;

[0023] S22, the formula for exact matching of categorical features is:

[0024]

[0025] Where m is the total number of categorical features, V new,k V case,k These are the values ​​of the new task and the historical case on the k-th category feature, respectively.

[0026] In one embodiment of the present invention, S3 includes:

[0027] S31, The construction of the parameter dependency graph includes total air demand. Depends on gas outflow volume Number of people Minimum wind speed Required wind pressure Depends on total resistance and ;

[0028] S32, the chain reaction correction uses a breadth-first algorithm to propagate differences along the parameter dependency network, prioritizing the updating of directly related parameters before gradually affecting downstream parameters.

[0029] In one embodiment of the present invention, the method further includes:

[0030] S5 dynamically adjusts the weight parameters of the multi-dimensional similarity model based on the actual operating effect of the newly generated procedures, and optimizes the weight allocation based on feedback data from historical cases through machine learning methods.

[0031] To achieve the above objectives, a second aspect of the present invention provides an adaptive procedure generation device based on case reasoning and transfer learning, comprising:

[0032] The knowledge base construction module is used to build a structured knowledge base of excellent cases that includes input conditions, process parameters, final solutions and operating results. It uses natural language processing and information extraction technology to parse historical engineering code documents and transform implicit engineering experience into computable explicit knowledge.

[0033] The multi-dimensional similarity retrieval module is used to parse the input conditions of new engineering tasks and retrieve historical cases in the knowledge base based on the multi-dimensional similarity calculation model. The multi-dimensional similarity model integrates the normalized distance calculation of numerical features, the precise matching of categorical features, and the semantic vector calculation of text features to generate a comprehensive similarity score and return the reference case with the highest matching degree.

[0034] The parameter dependency graph adjustment module is used to adaptively adjust the solution of the reference case based on the difference in input conditions between the new engineering task and the reference case using a hybrid reasoning mechanism driven by the parameter dependency graph. The hybrid reasoning mechanism combines the engineering physical model and the preset heuristic rules to carry out a chain reaction correction along the parameter dependency relationship network until the adjusted solution with parameter convergence is generated.

[0035] The procedure generation and knowledge iteration module is used to generate engineering procedure documents containing the adjusted solutions, and to add experience annotations, reference cases and adjustment basis. New procedure solutions that have been verified as excellent in actual application are automatically uploaded back to the knowledge base as high-quality cases, forming a knowledge iteration closed loop.

[0036] The embodiments of the present invention have the following significant technical effects in application:

[0037] 1. Explicit Utilization and Inheritance of Tacit Knowledge: A creative method is proposed to activate the valuable and ineffable "tacit knowledge" of senior engineers from dormant historical documents, transforming it into computable, transferable, and reusable structured dynamic knowledge, thus solving the core pain point of the difficulty in inheriting and scaling up expert experience.

[0038] 2. Deep integration of case-based reasoning and transfer learning: Combining the classic case-based reasoning (CBR) paradigm with the transfer learning capabilities of modern AI. This ensures the real-world basis and reliability of the solution (derived from verified successful cases), while endowing the system with intelligent flexibility and adaptability to cope with new situations and problems, far exceeding traditional fixed-rule systems.

[0039] 3. Achieving self-growth and continuous evolution of the knowledge system: By introducing a feedback loop mechanism, each successfully validated new solution is absorbed as new "experience," continuously expanding and optimizing the case knowledge base. This enables the system to continuously learn and self-evolve, adapting to technological developments and changes in operating conditions.

[0040] 4. Significantly improves design efficiency and standardization: The manual design process, which previously took days or even weeks, is reduced to minutes, greatly liberating engineers' productivity. Simultaneously, because the solution generation is based on historical best practices, it ensures high quality and standards in the output procedures, reducing uncertainty caused by individual skill differences.

[0041] 5. Enhance the interpretability and credibility of the decision-making process: The unique "experience annotation" function clearly demonstrates the AI's "thinking process"—which cases it referenced and what adjustments were made to address specific differences. This transparent decision-making path enables engineers to understand, trust, and effectively supervise the results generated by AI.

[0042] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0043] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0044] Figure 1 This is a flowchart of an adaptive procedure generation method based on case reasoning and transfer learning according to an embodiment of the present invention;

[0045] Figure 2 This is an architecture diagram of an adaptive procedure generation method based on case reasoning and transfer learning according to an embodiment of the present invention;

[0046] Figure 3 This is a structural diagram of an adaptive procedure generation device based on case reasoning and transfer learning according to an embodiment of the present invention. Detailed Implementation

[0047] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0048] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.

[0049] The following description, with reference to the accompanying drawings, describes an adaptive procedure generation method and apparatus based on case reasoning and transfer learning according to an embodiment of the present invention.

[0050] Example 1

[0051] Figure 1 This is a flowchart of an adaptive procedure generation method based on case reasoning and transfer learning according to an embodiment of the present invention, such as... Figure 1 As shown, it includes:

[0052] S1. Construct a structured knowledge base of excellent cases that includes input conditions, process parameters, final solutions and operational effects. Use natural language processing and information extraction technology to parse historical engineering code documents and transform implicit engineering experience into computable explicit knowledge.

[0053] S2, parse the input conditions of the new engineering task, and retrieve historical cases in the knowledge base based on the multi-dimensional similarity calculation model. The multi-dimensional similarity model integrates the normalized distance calculation of numerical features, the precise matching of categorical features, and the semantic vector calculation of text features to generate a comprehensive similarity score and return the reference case with the highest matching degree.

[0054] S3. Based on the difference in input conditions between the new engineering task and the reference case, the solution of the reference case is adaptively adjusted using a hybrid reasoning mechanism driven by parameter dependency graph. The hybrid reasoning mechanism combines the engineering physical model and the preset heuristic rules to perform a chain reaction correction along the parameter dependency relationship network until the adjusted solution with parameter convergence is generated.

[0055] S4 generates an engineering specification document containing the adjusted solution, and adds experience annotations, reference cases, and adjustment basis. The new specification solution that has been verified as excellent in actual application is automatically uploaded back to the knowledge base as a high-quality case, forming a knowledge iteration closed loop.

[0056] This invention proposes an adaptive procedure generation method based on case-based reasoning and transfer learning. By constructing a superior case knowledge base, it structures and makes explicit historically validated expert success experiences (tacit knowledge). It uniquely combines case-based reasoning and transfer learning, enabling the system not only to find the most similar historical cases but also to intelligently and adaptively modify the solution based on the differences between the new task and historical cases, thus automatically generating highly customized, practical engineering procedures with "design concept" annotations. This solves the technical problems of low efficiency and poor consistency in traditional methods, as well as the rigidity and lack of learning and evolution capabilities of existing software systems.

[0057] Example 2

[0058] This invention proposes an adaptive procedure generation method based on case reasoning and transfer learning, the overall process of which is as follows: Figure 2 As shown, it mainly includes four core parts: construction of a knowledge base of excellent cases, similarity matching and case retrieval, adaptive adjustment of solutions based on transfer learning, and generation of procedures with experience annotations and iteration of the knowledge base.

[0059] S10. Construction of a knowledge base of excellent cases.

[0060] This stage forms the foundation of the system. First, a large number of historically proven and successful engineering operation procedures (such as mine ventilation design specifications) are collected and input. The system uses Natural Language Processing (NLP) and Information Extraction (IE) technologies to deeply analyze these unstructured or semi-structured documents, automatically extracting key information and organizing it into structured "case studies".

[0061] The specific implementation steps are as follows:

[0062] Document parsing and preprocessing: For Word (.docx) and PDF documents, text processing tools are used to extract the plain text content. The extracted text is then cleaned, removing headers, footers, and special characters, and regular expressions are used to segment the text (e.g., segmenting based on headings such as "I. Overview" and "II. Parameter Calculation").

[0063] Information Extraction (IE): Constructing a deep learning model based on the Transformer model to extract key entities and relationships.

[0064] Define entity labels: including "input conditions" (e.g., LOC_GEO geological parameters, VAL_GAS gas volume) and "process parameters" (e.g., VAL_WIND air volume, EQUIP_FAN fan model).

[0065] Define relationship labels: attribute relationships between entities (e.g., "Gas Emission" Is_Input_Of "Total Air Requirement").

[0066] Model Training: The pre-trained Transformer model is fine-tuned using a labeled historical protocol corpus. Input is a text sequence, output is entity and relation label information.

[0067] Structured storage: The extracted entities and their values ​​are mapped to a predefined JSON template to form structured case data C = {I, P, S, E}, where I is the input set, P is the parameter set, S is the solution set, and E is the effect evaluation.

[0068] Each case study should include at least the following elements: Input conditions are a set of parameters describing the task background, such as geological information, tunnel geometry parameters (length, cross-section), gas emission rate, equipment parameters, ventilation method, etc. Process parameters are key calculated values ​​and decision points in the design process, such as calculated air volume, calculated resistance, selection of safety factors, and selection of fan model, etc. The final solution is the complete and final output procedure. The operational effect is an evaluation of the solution's effectiveness in practical application, such as "excellent," "qualified," or "improved," serving as a quality label for the case study.

[0069] S20. Similarity matching and case retrieval.

[0070] When a new engineering task is received, the system first parses its input conditions. Then, it searches the case library. The similarity calculation here is a multi-dimensional similarity evaluation algorithm, specifically defined as follows:

[0071] Assume the new task is T new Any historical case in the knowledge base is C i Its overall similarity Sim(T) new C i It consists of a weighted sum of different types of features:

[0072]

[0073] Among them, w num , w cat , w text The weights for each type of feature satisfy... The weights are pre-set by domain experts to reflect the degree of influence of different features on the decision. num For numerical feature similarity, Sim cat For categorical feature similarity, Sim text For text-based feature similarity, the definitions are as follows.

[0074] Feature normalization processing:

[0075] Before performing weighted summation, the similarity of each dimension is first normalized to ensure that its value range is [0, 1].

[0076] Numerical Sim num The distance has been restricted to [0,1] by the normalized distance formula.

[0077] Category Sim cat The output is 0 or 1 according to the definition.

[0078] Text-based Sim textCosine similarity was used and normalized. The range of cosine similarity values ​​is [-1, 1], and after normalization, it is mapped to [0, 1].

[0079] Weight setting and initialization:

[0080] Initial setup: AHP (Analytic Hierarchy Process) is adopted. 5-10 domain experts are invited to conduct pairwise comparisons and scores on the importance of "numerical accuracy", "scene similarity (text)" and "process matching degree (category)", construct a judgment matrix, and calculate the eigenvector corresponding to the largest eigenvalue as the initial weight.

[0081] Dynamic adjustment: See the subsequent weight optimization module for details.

[0082] Numerical Feature Similarity (Sim) num For numerical features such as tunnel length and gas emission, normalized distance is used for calculation. For each numerical feature j, its similarity sim... j The calculation is as follows:

[0083]

[0084] Where V new,j and V case,j These are the values ​​of the new task and the historical case on feature j, respectively, max(V j ) and min(V j ) represents the maximum and minimum values ​​of this feature across the entire knowledge base.

[0085] Sim num It is the similarity of all numerical features (sim). j Weighted average:

[0086]

[0087] Where w j Let n be the weights, which are trained and optimized using machine learning methods based on historical data. The sum of all weights is 1.

[0088] Categorical feature similarity (Sim) cat For features such as ventilation methods, exact matching is used. If the feature values ​​of the new task and historical cases are the same, the similarity is 1; otherwise, it is 0. cat It is the average of the similarity of features across all categories. The calculation formula is:

[0089]

[0090] Where m is the total number of categorical features, V new,k Vcase,k These are the values ​​of the new task and the historical case on the k-th category feature, respectively.

[0091] Textual feature similarity (Sim) text For text features such as geological descriptions, semantic similarity calculation based on deep learning is adopted.

[0092] A pre-trained Transformer model is used. Text descriptions of historical cases and new tasks (such as geological overview paragraphs) are input into the pre-trained model (e.g., Qwen3-Embedding) to extract text vectors. The cosine similarity between the two vectors is calculated.

[0093] The system uses a pre-trained language model to convert text descriptions into high-dimensional vectors. The similarity between two texts is the cosine similarity of their normalized corresponding vectors. The calculation formula is:

[0094]

[0095] Among them, V new and V case These are vectors obtained by encoding the text descriptions of new tasks and historical cases using a language model.

[0096] After calculating the comprehensive similarity score between the new task and each historical case using the above formula, the system returns one or more of the most similar reference cases.

[0097] S30. Adaptive adjustment of the scheme based on transfer learning.

[0098] The system does not directly copy the solution from the reference case, but rather uses it as a "starting point" or "template" for intelligent adjustments. A dedicated "adaptive adjustment AI module" performs the following operations:

[0099] Difference analysis: Compare the differences (Δ) of the "input conditions" parameters between the new task and the reference case item by item.

[0100] Definition and calculation of parameter differences:

[0101] The system iterates through new tasks. Compared with reference cases Calculate the difference set based on all input conditions. :

[0102] 1. Differences in numerical parameters: For example, differences in gas emission rates. .

[0103] 2. Categorical parameter differences: If Then mark Changes in ventilation events. For example, the ventilation method changes from "forced ventilation" to "exhaust ventilation".

[0104] The system automatically compares corresponding fields in two structured objects using key-value pair matching to generate a set of differences.

[0105] Migration Correction: This process is based on "model-driven + rule-based reasoning." The AI ​​module calls upon a built-in "migration knowledge base," which integrates fundamental engineering physics models (such as ventilation network calculations and fan performance curve equations) and expert heuristic rules learned from numerous case studies. Based on the difference Δ, the AI ​​intelligently corrects the "process parameters" and "final solutions" of the reference case. These heuristic rules are expert experience embedded in the system, used to handle decision points that cannot be fully described by physical models or require trade-offs. These rules exist in the form of "IF-THEN." Below are some examples of these rules:

[0106] Rule 1: Dynamic adjustment rule for safety factor.

[0107] IF (Gas emission rate Q for new mission) gas new >10 m³ / min and geological complexity rating is "high") OR (historical gas concentration volatility of new mission >20%)

[0108] THEN (Based on the gas volume calculation formula, the safety factor K) gas Increase the value by 0.1-0.3 based on the reference case, or set it to no less than 1.8.

[0109] Note: This rule is used to deal with high-risk and highly uncertain working conditions and to increase safety redundancy.

[0110] Rule 2: Fan selection margin rule.

[0111] IF (The calculated operating point (Q, H) is located at the edge of the high-efficiency region of the reference wind turbine performance curve (e.g., the region where efficiency drops by more than 10% from the point of maximum efficiency)) OR (The calculated total resistance R) total The uncertainty is relatively high, such as the possibility of future tunnel extension plans.

[0112] THEN (Research the fan database again and select a fan model that contains the current operating point near the center of its high-efficiency zone and has a margin of at least 15% in air pressure or air volume).

[0113] Note: This rule ensures that wind turbines operate in an economical and stable range, while allowing for future changes.

[0114] Rule 3: Wind speed limit takes precedence.

[0115] IF (The tunnel is the main transportation tunnel AND there are many dust sources within the tunnel (such as transfer points))

[0116] THEN (Increase the priority of "air volume calculated based on the maximum wind speed limit". Even if other calculated air volumes (such as those based on gas) are larger, it is necessary to verify whether the final air volume will lead to overspeed. If overspeed occurs, solutions such as expanding the cross-section or diverting the flow should be considered).

[0117] Note: This rule balances ventilation requirements with work environment requirements (such as preventing dust and ensuring pedestrian comfort).

[0118] Intelligent correction is an AI module that simulates the complete thought process of an expert adjusting a solution. It doesn't modify a single parameter in isolation, but rather recalculates and makes decisions through a chain reaction of dependencies. An example process is as follows:

[0119] Step 1: Construct a parameter dependency graph. The system pre-constructs a dependency graph containing engineering parameters. For example, the total required air volume Q... total Depends on gas emission Q gas Number of people Q person Minimum wind speed Q speed The airflow resistance R depends on the tunnel length L, cross-section A, and friction coefficient α; the required air pressure H depends on R and Q. total Fan selection model Depends on H and Q total .

[0120] The process of constructing a parameter dependency graph:

[0121] The parameter dependency graph is a directed acyclic graph. Nodes represent engineering parameters (such as gas volume, air volume, and resistance), and edges represent computational dependencies.

[0122] Built using a hybrid approach:

[0123] Physical Formula Layer (Hard Constraints): This layer automatically generates edges based on professional internal standard physical models (such as fluid mechanics), describing the computational relationships between different variables. For example, formulas... Edges from R to H and from Q to H have been established.

[0124] Empirical Rule Layer (Soft Constraints): This layer describes heuristic empirical relationships between different variables through expert input. For example, experts might set the rule "Gas quantity affects safety factor," and input the following rule:

[0125] IF (Delta_Q_gas>0) THEN (Safety_Factor = Safety_Factor_Ref * (1 + 0.1* Delta_Q_gas))

[0126] The system is entered into the database and established. arrive The edge.

[0127] Step 2: Identify the initial difference and initiate the correction chain. The parameter difference Δ output by the "Difference Analysis" module is the starting point of the correction chain. For example, the system identifies the tunnel length L of the new task. new Compared with the reference case L case Different (ΔL).

[0128] Step 3: Propagational correction along the dependency graph. Starting from the initial difference point, the AI ​​performs a breadth-first or depth-first traversal along the parameter dependency graph, correcting all affected downstream parameters.

[0129] Trigger: L new The change in L triggered a recalculation of the wind resistance R. Model invocation: The AI ​​invokes the physical model for wind resistance calculation, substituting L... new Using other relevant parameters (such as section A, perimeter, etc.), the new drag R is calculated. new Chain reaction: R new Changes in [the wind pressure] will trigger a recalculation of the required wind pressure H. The model is invoked again: the AI ​​calls the wind pressure calculation model, substituting R [ into the calculation]. new And the current Q total , obtain a temporary H temp Parallel correction: Meanwhile, if the gas outflow Q of the new task... gas new Unlike the reference case, the AI ​​will initiate another correction chain in parallel. It will invoke the engineering model that calculates airflow based on gas outflow and may trigger "heuristic rule 1" to adjust the safety factor K. gas Thus, a new Q is obtained. gas new calc Decision-making and integration: After all parallel computations are completed, the AI ​​will re-execute the decision-making logic, such as Q. total new = max(Q gas new calc Q person new calc Q speed new calc ), thus obtaining the new total required air volume Q. total new .

[0130] Step 4: Iteration and Convergence. Because Q... total newThe change in the wind pressure (determined in the previous step) will trigger a recalculation of the required wind pressure H. This process will continue until all affected parameters in the dependency graph have been updated and the entire parameter system reaches a new, self-consistent steady state, where no parameters require further updates.

[0131] Algorithm implementation: Breadth-first search is used.

[0132] There will be differences The node is placed into queue Q.

[0133] When Q is not empty, retrieve node u and recalculate the value of u according to the physical model or heuristic rules.

[0134] If the value of u changes, find all child nodes in the dependency graph and add the child nodes to queue Q.

[0135] Convergence condition:

[0136] Parameter stability threshold: The rate of change of all parameters is less than the threshold (e.g., 0.01).

[0137] Maximum number of iterations: To prevent infinite loops, set the maximum propagation depth to 50 or the maximum computation time. If convergence is not achieved after exceeding the limit, the system will issue an alarm and prompt manual intervention.

[0138] Step 5: Final Solution Formation. Once all parameters converge, the system obtains a complete, adaptively adjusted new solution. For example, based on the final Q... total new and H new The system will call "Heuristic Rule 2" to re-select wind turbines and finally output a procedure scheme that adapts to the new task conditions.

[0139] Taking the migration point of "air volume calculation" in ventilation design as an example, the process is explained in detail:

[0140] a. Decision-making logic for tracing the source of the reference case: The AI ​​first analyzes the final air volume Q in the reference case. case How was it determined? Q was recorded in the structured case data. case It is max(Q) gas Q person Q speed The result. For example, in the reference case, the air volume Q calculated based on gas emission. gas =450m³ / min, calculated based on the number of people. person =200m³ / min, calculate Q based on the minimum wind speed requirement. speed =300 m³ / min. Therefore, the final air volume is determined by the gas emission, Q case = 450m³ / min.

[0141] b. Identifying key parameter differences: AI identifies the gas emission rate Q for the new task. gas new Q in the reference case gas case Different (ΔQ) gas ).

[0142] c. Invoke migration rules and recalculate: The AI ​​invokes the engineering model / rule of "calculate air volume based on gas outflow" to recalculate the air volume Q required for the new task. gas new Q person new and Q speed new .

[0143] d. Forming a new decision: The AI ​​ultimately determines the total required air volume Q for the new task. new = max(Q gas new Q person new Q speed new This process does not involve scaling up the original air volume, but rather replicates the expert's complete decision-making chain and re-reasons by incorporating new conditions, thereby ensuring the engineering rationality of the result.

[0144] The migration of other parameters follows a similar logic: the adjustment of one parameter will serve as input, triggering a chain of recalculations of downstream parameters, forming a complete and self-consistent scheme adjustment chain.

[0145] S40. Generate procedures and knowledge base iterations with "experience annotations".

[0146] Generation of procedural documents and experience-based annotations:

[0147] Template-based document generation: The system provides multiple pre-set procedure document templates (XML / Word XML format). Each template contains fixed text and dynamic slots. The system then fills the calculated final parameters into the corresponding slots.

[0148] Annotation generation using templates:

[0149] 1. The system records the trigger chain during the reasoning process.

[0150] 2. The corresponding explanation template is invoked: "Because the [input parameter name] has an increased [difference value] compared to the reference case, the system automatically adjusts the [modified parameter name] to the [new value] according to the [rule name]."

[0151] 3. The final generated sentence is: "Since the gas emission rate increased by 2 m³ / min compared to the reference case, the system automatically adjusts the safety factor to 1.8 according to the gas safety factor adjustment rules."

[0152] Ultimately, the system generates not only a procedure document containing calculation results and solution details, but more importantly, it automatically adds AI's "design thinking" or "experience annotations." This interpretability greatly enhances engineers' trust in the AI-generated solutions.

[0153] Furthermore, the weight parameters of the multi-dimensional similarity model are dynamically adjusted based on the actual operational performance of the newly generated procedures, and the weight allocation is optimized using machine learning methods based on feedback data from historical cases. Once the newly generated procedures are successfully verified in practical applications, the solution, along with its complete parameters and results, will be absorbed into the excellent case knowledge base as a new high-quality case. This forms a virtuous cycle and self-growing mechanism of knowledge, making the system "smarter with use."

[0154] Example 3

[0155] The following example, using a specific scenario of selecting a local ventilation fan and generating ventilation procedures in a mine, will be used to illustrate the implementation of the present invention in detail.

[0156] Scenario: A coal mine needs to design a local ventilation scheme for the new "11302 tunneling face".

[0157] S100: Task parameter input.

[0158] The engineer inputs the key parameters for "Working Face 11302" into the system interface: the estimated total tunneling length L = 1200 meters, the tunnel cross-section S = 16 m², and the predicted gas emission Q from the geological report. gas =22 m³ / min, requiring forced ventilation.

[0159] S200: Case retrieval and matching.

[0160] After receiving the task parameters, the system searches the "Excellent Case Knowledge Base." Using a multi-dimensional similarity algorithm, the system finds that "Case No. C-87" (corresponding to the successful solution for the "9105 tunneling face") has the highest similarity to the current task (overall score 0.93). The core parameters of "Case C-87" are: L=1000 meters, S=16 m², Q... gas =18 m³ / min, and finally selected the FBCZ-8-No.25 model fan.

[0161] S300: The solution is adaptively adjusted.

[0162] The system activates the "Adaptive Adjustment AI Module" and uses "Case C-87" as a template for modification:

[0163] 1. Difference Recognition: The AI ​​identified the main differences between the old and new tasks: ΔL = +200 meters, ΔQgas = +4 m³ / min.

[0164] 2. Airflow Migration Adjustment: The AI ​​calls the "Gas Emission Airflow Calculation Rule" in the rule base. The rule states that the required airflow must be calculated based on the gas emission rate and safety factor. Due to Q gas With this addition, AI increases the required airflow from 480 m³ / min in the reference case to 560 m³ / min.

[0165] 3. Wind resistance shift adjustment: The AI ​​invokes the "ventilation resistance calculation rule" (based on Atkinson's formula). Due to the increase in both the tunnel length L and the air volume Q, the AI ​​calculates that the total system resistance will increase significantly from 2800Pa in the reference case to 3750Pa.

[0166] 4. Fan Selection Migration and Adjustment: AI analysis of the performance curve of the FBCZ-8-No.25 fan in the reference case revealed that under operating conditions of 3750Pa resistance and 560m³ / min airflow, the fan was already on the edge of the stall zone, with extremely low efficiency and unstable operation. AI then re-matched the fan in its built-in equipment library, recommending the FBCZ-8-No.28 model fan with higher pressure and greater flow rate, and verified that its operating point was in the high-efficiency zone (efficiency > 75%).

[0167] S400: Generates annotated specifications.

[0168] Based on the above reasoning, a complete "Design Specification for Local Ventilation of 11302 Tunneling Face" was automatically generated using the specification template system. In addition to including all calculation results, fan parameters, ventilation duct specifications, and installation requirements, a special "Design Notes" section was also generated.

[0169] "

Reference Basis

[0170] "

Adjustment 1: Air volume

[0171] "

Adjustment Item 2: Fan Selection

[0172] S500: Knowledge base iteration.

[0173] Six months after the solution was adopted and successfully implemented, the field engineer marked it as "performing well" in the system. The system automatically stored the complete design process and results of "Workface 11302" as a new high-quality case (e.g., case number C-152) in the knowledge base for reference and learning in similar tasks in the future.

[0174] Example 4

[0175] To achieve the above embodiments, such as Figure 3 As shown, this embodiment also provides an adaptive procedure generation device 10 based on case reasoning and transfer learning. The device 10 includes a knowledge base construction module 100, a multi-dimensional similarity retrieval module 200, a parameter dependency graph adjustment module 300, and a procedure generation and knowledge iteration module 400.

[0176] The knowledge base construction module 100 is used to build a structured knowledge base of excellent cases containing input conditions, process parameters, final solutions and operating effects. It uses natural language processing and information extraction technology to parse historical engineering code documents and transform implicit engineering experience into computable explicit knowledge.

[0177] The multi-dimensional similarity retrieval module 200 is used to parse the input conditions of the new engineering task and retrieve historical cases in the knowledge base based on the multi-dimensional similarity calculation model. The multi-dimensional similarity model integrates the normalized distance calculation of numerical features, the precise matching of categorical features, and the semantic vector calculation of text features to generate a comprehensive similarity score and return the reference case with the highest matching degree.

[0178] The parameter dependency graph adjustment module 300 is used to adaptively adjust the solution of the reference case based on the difference in input conditions between the new engineering task and the reference case using a hybrid reasoning mechanism driven by the parameter dependency graph. The hybrid reasoning mechanism combines the engineering physical model and the preset heuristic rules to carry out a chain reaction correction along the parameter dependency relationship network until the adjusted solution with parameter convergence is generated.

[0179] The procedure generation and knowledge iteration module 400 is used to generate engineering procedure documents containing adjusted solutions, and add experience annotations, reference cases and adjustment basis. New procedure solutions that have been verified as excellent in actual application are automatically uploaded back to the knowledge base as high-quality cases, forming a knowledge iteration closed loop.

[0180] Furthermore, the aforementioned knowledge base construction module is also used for:

[0181] Input conditions include roadway geometry parameters, gas emission rate, equipment parameters, and ventilation method;

[0182] The process parameters include calculated air volume, calculated resistance, selection of safety factor, and selection of fan model.

[0183] Furthermore, the aforementioned multi-dimensional similarity retrieval module is also used for:

[0184] The formula for calculating the normalized distance of the numerical features is as follows:

[0185]

[0186] Where Vnew,j and Vcase,j are the values ​​of the new task and the historical case on feature j, respectively, and max(Vj) and min(Vj) are the maximum and minimum values ​​of this feature in the entire knowledge base;

[0187] The formula for exact matching of the categorical features is:

[0188]

[0189] Where m is the total number of categorical features, and Vnew,k and Vcase,k are the values ​​of the new task and the historical case on the k-th categorical feature, respectively.

[0190] Furthermore, the parameter dependency graph adjustment module mentioned above is also used for:

[0191] The construction of the parameter dependency graph includes total air demand. Depends on gas outflow volume Number of people Minimum wind speed Required wind pressure Depends on total resistance and ;

[0192] The chain reaction correction uses a breadth-first algorithm to propagate differences along the parameter dependency network, prioritizing the updating of directly related parameters before gradually affecting downstream parameters.

[0193] Furthermore, the device 10 also includes:

[0194] The weight adjustment module is used to dynamically adjust the weight parameters of the multi-dimensional similarity model based on the actual running effect of the newly generated procedure, and optimize the weight allocation based on feedback data from historical cases through machine learning methods.

[0195] Weighting adjustment module:

[0196] A supervised learning method based on logistic regression is used to optimize the weight parameters of multi-dimensional similarity. The specific steps are as follows:

[0197] 1. Construction of the training dataset:

[0198] The system extracts "query-case" pairs from historical logs to construct positive and negative sample sets:

[0199] Positive samples: Marked as y=1. Selecting excellent cases that engineers ultimately adopted in historical tasks. Combined with task T, calculate its eigenvectors.

[0200]

[0201] Negative samples: Marked as y=0. For the same task T, several other cases that were not adopted are randomly selected. Calculate its eigenvectors.

[0202]

[0203] 2. Model training algorithm:

[0204] A logistic regression model was used for fitting. The model assumes that the probability P of a case being adopted follows a sigmoid function relationship with the similarity feature:

[0205]

[0206] in, These correspond to numerical values, categories, and text similarity, respectively. The regression coefficients are to be learned.

[0207] The model is trained using the stochastic gradient descent (SGD) algorithm to minimize the log loss function, thereby obtaining the optimal coefficients.

[0208] 3. Weight extraction and normalization:

[0209] After training, the regression coefficients learned by the model This directly reflects the contribution (i.e., importance) of each dimension's similarity to the "case adoption" decision. To satisfy the constraint that the sum of the weights is 1, the system performs Softmax normalization on the coefficients, obtaining the final numerical, category, and text similarity weights used for similarity calculation:

[0210]

[0211] 4. Update:

[0212] When the newly accumulated case data reaches a preset threshold (e.g., every 100 new cases), the model is retrained and the weights are updated.

[0213] This invention discloses an adaptive procedure generation device based on case-based reasoning and transfer learning. By constructing a superior case knowledge base, it structures and makes explicit historically validated expert success experiences (tacit knowledge). It uniquely combines case-based reasoning and transfer learning, enabling the system not only to find the most similar historical cases but also to intelligently and adaptively modify the solution based on the differences between the new task and historical cases, thereby automatically generating highly customized, practical engineering procedures with "design concept" annotations. This invention solves the technical problems of low efficiency and poor consistency in traditional methods, as well as the rigidity and lack of learning and evolution capabilities of existing software systems.

[0214] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0215] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

Claims

1. An adaptive procedure generation method based on case-based reasoning and transfer learning, characterized in that, include: S1. Construct a structured knowledge base of excellent cases that includes input conditions, process parameters, final solutions and operational effects. Use natural language processing and information extraction technology to parse historical engineering code documents and transform implicit engineering experience into computable explicit knowledge. S2, parse the input conditions of the new engineering task, and retrieve historical cases in the knowledge base based on the multi-dimensional similarity calculation model. The multi-dimensional similarity model integrates the normalized distance calculation of numerical features, the precise matching of categorical features, and the semantic vector calculation of text features to generate a comprehensive similarity score and return the reference case with the highest matching degree. S3. Based on the difference in input conditions between the new engineering task and the reference case, the solution of the reference case is adaptively adjusted using a hybrid reasoning mechanism driven by parameter dependency graph. The hybrid reasoning mechanism combines the engineering physical model and the preset heuristic rules to carry out a chain reaction correction along the parameter dependency relationship network until the adjusted solution with converged parameters is generated. S4 generates an engineering specification document containing the adjusted scheme, and adds experience annotations, reference cases and adjustment basis. The new specification scheme that has been verified as excellent in actual application is automatically uploaded back to the knowledge base as a high-quality case, forming a knowledge iteration closed loop. S1 includes: S11, the input conditions include roadway geometric parameters, gas emission rate, equipment parameters, and ventilation method; S12, the process parameters include calculated air volume, calculated resistance, safety factor selection, and fan model selection; The S3 includes: S32, the chain reaction correction uses a breadth-first algorithm to propagate differences along the parameter dependency network, prioritizing the updating of directly related parameters before gradually affecting downstream parameters.

2. The method as described in claim 1, characterized in that, S2 includes: S21, the formula for calculating the normalized distance of numerical features is: Where V new,j and V case,j These are the values ​​of the new task and the historical case on feature j, respectively, max(V j ) and min(V j ) represents the maximum and minimum values ​​of this feature across the entire knowledge base; S22, the formula for exact matching of categorical features is: Where m is the total number of categorical features, V new,k V case,k These are the values ​​of the new task and the historical case on the k-th category feature, respectively.

3. The method as described in claim 1, characterized in that, The S3 further includes: S31, The construction of the parameter dependency graph includes total air demand. Depends on gas outflow volume Number of people Minimum wind speed Required wind pressure Depends on total resistance and .

4. The method as described in claim 1, characterized in that, The method further includes: S5 dynamically adjusts the weight parameters of the multi-dimensional similarity model based on the actual operating effect of the newly generated procedures, and optimizes the weight allocation based on feedback data from historical cases through machine learning methods.

5. An adaptive procedure generation device based on case-based reasoning and transfer learning, characterized in that, include: The knowledge base construction module is used to build a structured knowledge base of excellent cases that includes input conditions, process parameters, final solutions and operating results. It uses natural language processing and information extraction technology to parse historical engineering code documents and transform implicit engineering experience into computable explicit knowledge. The multi-dimensional similarity retrieval module is used to parse the input conditions of new engineering tasks and retrieve historical cases in the knowledge base based on the multi-dimensional similarity calculation model. The multi-dimensional similarity model integrates the normalized distance calculation of numerical features, the precise matching of categorical features, and the semantic vector calculation of text features to generate a comprehensive similarity score and return the reference case with the highest matching degree. The parameter dependency graph adjustment module is used to adaptively adjust the solution of the reference case based on the difference in input conditions between the new engineering task and the reference case using a hybrid reasoning mechanism driven by the parameter dependency graph. The hybrid reasoning mechanism combines the engineering physical model and the preset heuristic rules to carry out a chain reaction correction along the parameter dependency relationship network until the adjusted solution with parameter convergence is generated. The procedure generation and knowledge iteration module is used to generate engineering procedure documents containing the adjusted solutions, and add experience annotations, reference cases and adjustment basis. New procedure solutions that have been verified as excellent in actual application are automatically uploaded back to the knowledge base as high-quality cases, forming a knowledge iteration closed loop. The knowledge base construction module is also used for the input conditions, including roadway geometric parameters, gas emission rate, equipment parameters, and ventilation method. The process parameters include calculated air volume, calculated resistance, safety factor selection, and fan model selection. The parameter dependency graph adjustment module is also used for chain reaction correction. It uses a breadth-first search algorithm to propagate differences along the parameter dependency network, prioritizing the updating of directly related parameters and then gradually affecting downstream parameters.

6. The apparatus as claimed in claim 5, characterized in that, The multi-dimensional similarity retrieval module is also used for: The formula for calculating the normalized distance of numerical features is: Where V new,j and V case,j These are the values ​​of the new task and the historical case on feature j, respectively, max(V j ) and min(V j ) represents the maximum and minimum values ​​of this feature across the entire knowledge base; The formula for exact matching of categorical features is: Where m is the total number of categorical features, V new,k V case,k These are the values ​​of the new task and the historical case on the k-th category feature, respectively.

7. The apparatus as claimed in claim 5, characterized in that, The parameter dependency graph adjustment module is also used for: The construction of the parameter dependency graph includes total air demand. Depends on gas outflow volume Number of people Minimum wind speed Required wind pressure Depends on total resistance and .

8. The apparatus as claimed in claim 5, characterized in that, Also includes: The weight adjustment module is used to dynamically adjust the weight parameters of the multi-dimensional similarity model based on the actual running effect of the newly generated procedure, and optimize the weight allocation based on feedback data from historical cases through machine learning methods.

Images