Research and development system based on multi-objective optimization and knowledge mining

Abstract

The invention relates to the field of research and development assistance systems, in particular to a research and development system based on multi-objective optimization and knowledge mining, which is characterized in that a demand task generation module matches an objective function and a constraint condition according to a design demand to form an initial task book; the case retrieval module retrieves a plurality of historical cases with high similarity in the knowledge base to generate a historical case set; the sensitivity analysis module calculates the sensitivity of each adjustment value under the constraint of the constraint condition of the design demand, and generates a sensitivity report; the multi-algorithm fusion optimization module uses multiple algorithms to carry out parallel iterative analysis and fusion communication according to the initial task book, the historical case set and the sensitivity report to obtain an optimal scheme set, and then the digital twinborn verification module carries out multi-physics field simulation analysis to generate a simulation report; the final review module performs comprehensive review on the initial task book, the optimal scheme set and the simulation report, and outputs a review scheme and a modification suggestion; the problem of low research and development efficiency caused by multi-person cooperation in the research and development process is solved.

Description

Technical Field

[0001] This invention relates to the field of research and development assistance systems, and in particular to a research and development system based on multi-objective optimization and knowledge mining. Background Technology

[0002] Currently, most companies' product development processes rely on a series of discrete software tools. For example, in transformer development, the development task is broken down into multiple sub-tasks and distributed to different engineers. Two-dimensional or three-dimensional software (such as AutoCAD, Solidworks, etc.) is used for structural design, finite element analysis software (such as FEA) is used for electromagnetic field analysis and loss calculation, and computational fluid dynamics software (such as CFD) is used for thermal simulation, etc. Then, the R&D project manager integrates the results of the sub-tasks to conduct performance tests and feedback modifications, which requires a lengthy R&D process. More importantly, for complex products like transformers that involve multi-physics coupling, multiple iterations are often required to meet all performance indicators, resulting in long development cycles and high costs, which greatly affects overall R&D efficiency. Summary of the Invention

[0003] To address the aforementioned shortcomings, the present invention aims to propose a research and development system based on multi-objective optimization and knowledge mining, which solves the problem of low research and development efficiency caused by multi-person collaboration during the research and development process.

[0004] To achieve this objective, the present invention adopts the following technical solution: A research and development system based on multi-objective optimization and knowledge mining includes: Requirements and Task Generation Module: Receives design requirements as input, matches objective functions and constraints according to the design requirements, and forms an initial task specification from the design requirements, objective functions, and constraints; Case retrieval module: Based on the design requirements in the initial task book, retrieve multiple historical cases with high similarity from the knowledge base and generate a historical case set; Sensitivity Analysis Module: Calculates the sensitivity of each adjustment value of the design requirements in the initial task statement under the constraints, and generates a sensitivity report; Multi-algorithm fusion optimization module: Based on the initial task description, historical case set and sensitivity report, multiple algorithms are first used in parallel iterative analysis, and then fused and communicated to obtain the best solution set; Digital twin verification module: Performs multiphysics simulation analysis on each solution in the optimal solution set and generates a simulation report; Review module: The review model comprehensively reviews the initial task description, the best solution set, and the simulation report, and outputs the review plan and modification suggestions.

[0005] Furthermore, the review module is also used to generate review constraints based on modification suggestions when such suggestions exist, and feed them back to the requirement task generation module, adding them to the initial task book.

[0006] Furthermore, it also includes a knowledge management module: used to count the number of iterations of each review scheme given by the review module in the multi-algorithm fusion optimization module; also used to count the number of modification suggestions for each review scheme given by the review module; comparing the number of iterations and the number of modification suggestions with preset number of iterations and preset number of suggestions respectively; when the number of iterations and the number of modification suggestions are less than the preset number of iterations and preset number of suggestions respectively, the scheme is added to the knowledge base; or when the scheme is confirmed for use by the R&D personnel, the scheme is added to the knowledge base.

[0007] Furthermore, the requirement task generation module also includes: after receiving the design requirement input, combining all the design parameters in the design requirement into a design vector.

[0008] Furthermore, the case retrieval module performs the following steps: A1: Based on the design requirements in the initial task book, use the KNN algorithm to retrieve K historical cases in the knowledge base whose similarity is within the preset similarity range; A2: For each of the K historical cases, a knowledge graph algorithm is used to traverse the nodes related to the historical case in the graph to obtain a list of suggestions for the historical case. A3: The historical case set is composed of the K historical cases in step A1 and the suggestion list in step A2.

[0009] Furthermore, the sensitivity analysis module employs the Sobol sequence algorithm.

[0010] Furthermore, in the multi-algorithm fusion optimization module, the NSGA-III algorithm, MOEA-D algorithm and PSO algorithm are used in parallel iteration, and population migration and elite retention strategies are adopted for fusion and communication during each Gming generation, so as to finally obtain the Pareto optimal solution set as the best solution set.

[0011] Furthermore, in the review module, the review model adopts an NLP model.

[0012] The technical solution provided by this invention can include the following beneficial effects: By replacing the R&D experience (i.e., knowledge mining ability) of various engineers in different roles through a requirement task generation module, a case retrieval module, and a sensitivity analysis module, multiple feasible preliminary solutions can be quickly generated; then, by replacing the R&D design, communication, collaboration, and iterative modification work of various engineers in the R&D process (i.e., multi-objective optimization ability), the R&D cycle that originally required weeks or even months is shortened to hours or even minutes; subsequently, by replacing the actual testing work of testers on samples through a digital twin verification module, preliminary sample testing is quickly completed, providing a basis for judging the best solution; finally, the review module replaces the decision-maker to select the best solution. In summary, this R&D system replaces the entire process of an R&D team, reducing the workload, communication time, and repetitive work of the R&D team, and greatly improving R&D efficiency. Attached Figure Description

[0013] Figure 1 This is a schematic diagram of a research and development system based on multi-objective optimization and knowledge mining, which is one embodiment of the present invention. Detailed Implementation

[0014] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0015] In the description of embodiments of the present invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of embodiments of the present invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0016] In the description of the embodiments of the present invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in the embodiments of the present invention according to the specific circumstances.

[0017] The following is combined Figure 1This invention describes a research and development system based on multi-objective optimization and knowledge mining, according to an embodiment of the present invention.

[0018] A research and development system based on multi-objective optimization and knowledge mining includes: Requirements and Task Generation Module: Receives design requirements as input, matches objective functions and constraints according to the design requirements, and forms an initial task specification from the design requirements, objective functions, and constraints; Case retrieval module: Based on the design requirements in the initial task book, retrieve multiple historical cases with high similarity from the knowledge base and generate a historical case set; Sensitivity Analysis Module: Calculates the sensitivity of each adjustment value of the design requirements in the initial task statement under the constraints, and generates a sensitivity report; Multi-algorithm fusion optimization module: Based on the initial task description, historical case set and sensitivity report, multiple algorithms are first used in parallel iterative analysis, and then fused and communicated to obtain the best solution set; Digital twin verification module: Performs multiphysics simulation analysis on each solution in the optimal solution set and generates a simulation report; Review module: The review model comprehensively reviews the initial task description, the best solution set, and the simulation report, and outputs the review plan and modification suggestions.

[0019] In a preferred embodiment of a research and development system based on multi-objective optimization and knowledge mining, this invention proposes, for example... Figure 1 As shown, the system replaces the R&D experience (i.e., knowledge mining ability) of individual engineers by using a requirement task generation module, a case retrieval module, and a sensitivity analysis module, enabling the rapid generation of multiple feasible preliminary solutions. Then, a multi-algorithm fusion optimization module replaces the R&D design, communication, collaboration, and iterative modification work of individual engineers during the R&D process (i.e., multi-objective optimization capability), shortening the R&D cycle from weeks or even months to hours or even minutes. Subsequently, a digital twin verification module replaces the actual testing of samples by testers, quickly completing preliminary sample testing and providing a basis for judging the optimal solution. Finally, a review module replaces the decision-maker to select the best solution. In summary, this R&D system replaces the entire workflow of an R&D team, reducing the team's workload, communication time, and repetitive work, significantly improving R&D efficiency.

[0020] It should be noted that the multiphysics simulation analysis performed on each scheme in the optimal scheme set in the digital twin verification module is essentially a physical field simulation formed by digital twins of various physical formulas to achieve verification. This can be achieved through existing digital twin technology, and will not be elaborated on here.

[0021] Furthermore, the review module is also used to generate review constraints based on modification suggestions when such suggestions exist, and then feed these constraints back to the requirement task generation module, adding them to the initial task book.

[0022] In this embodiment, after the R&D system completes one round of the R&D process and generates a review plan and modification suggestions, if the R&D personnel do not want to manually modify the plan, they can add the modification suggestions as new constraints to the initial task book and re-execute the R&D process until the review module finds a plan without modification suggestions.

[0023] Furthermore, it also includes a knowledge management module: used to count the number of iterations of each review solution in the multi-algorithm fusion optimization module, as well as the number of modification suggestions for the solution; then compare the number of iterations and the number of modification suggestions with the preset number of iterations and the preset number of suggestions, respectively. When the number of iterations and the number of modification suggestions are less than the preset number of iterations and the preset number of suggestions, the solution is added to the knowledge base (this is an automatic deposition method); or when the solution is confirmed for use by the R&D personnel (for example, by clicking confirmation on the operation interface), the solution is added to the knowledge base (this is a manual deposition method).

[0024] In this embodiment, if the multi-algorithm fusion optimization module wants to replace manual research and development with knowledge mining to obtain the product design solution that is closest to reality, it needs to rely on a knowledge base with a sufficiently deep knowledge accumulation. Therefore, it is preferable to put successful cases into the knowledge base after each research and development is completed in the research and development system to improve the research and development capability of the research and development system.

[0025] It should be noted that when the multi-algorithm fusion optimization module has a large number of iterations, it usually indicates a performance bottleneck and insufficient knowledge base resources. Therefore, the solution obtained at this time should not be included in the knowledge base, as it may lead to the accumulation of bad cases in the knowledge base. It is recommended to modify the number of cases accordingly.

[0026] Furthermore, the requirement task generation module also includes: after receiving the design requirement input, combining all the design parameters in the design requirement into a design vector.

[0027] In this embodiment, since the design requirements involve numerous design parameters and require subsequent algorithm calculations based on objective functions and constraints, the design requirements are preferably grouped into a design vector to facilitate subsequent calculations.

[0028] For example, design vector X=(D, N) lv B m J,d oil η cool ) T , representing a set of transformer design parameters; where D represents the core diameter, N lv Indicates the number of turns in the low-voltage winding, B mLet J represent magnetic flux density, and d represent current density. oil Indicates the oil passage spacing, η cool This represents the cooling efficiency coefficient; thus, the design requirements are structured and defined as the objectives and constraints of the optimization problem. The values ​​and boundaries of each design parameter are subsequently adjusted through the objective function and constraints matched with numerous design parameters. The multiple objective functions and constraints can be represented as follows: Objective function:

[0029] Where: f1(X) is the overall objective function, where P cu For copper loss, P fe For iron loss, P aux For auxiliary system losses; f2(X) is the total cost objective function, including material cost C. material and manufacturing cost C manufacturing f3(X) is the volume objective function, V tank Let f be the volume of the fuel tank; f4(X) be the objective function for temperature rise. This is due to the temperature rise of the winding hot spot.

[0030] Constraints:

[0031] Where: g1(X) is the voltage regulation constraint, ΔU% is the voltage regulation percentage, and ΔUmax is the maximum allowable voltage regulation; g2(X) is the temperature rise constraint, t wind For winding temperature rise, t max The maximum allowable temperature rise; g3(X) is the insulation strength constraint, E max For the maximum electric field strength, E breakdow n is the oil gap breakdown field strength; g4(X) is the mechanical strength constraint. For the von Mises equivalent effect, g5(X) represents the allowable stress of the material; g5(X) represents the noise constraint. For transformer noise level, This is the noise limit.

[0032] It should be noted that the above are only some of the objective functions and constraints, which are derived from existing transformer design specifications and physical theory formulas, etc., to provide a basis for the multi-algorithm fusion optimization module. The scope of the objective functions and constraints is not limited here.

[0033] Furthermore, the case retrieval module execution steps include: A1: Based on the design requirements in the initial task book, use the KNN algorithm to retrieve K historical cases in the knowledge base whose similarity is within the preset similarity range; A2: For each of the K historical cases, a knowledge graph algorithm is used to traverse the nodes related to the historical case in the graph to obtain a list of suggestions for the historical case. A3: The historical case set is composed of the K historical cases in step A1 and the suggestion list in step A2.

[0034] In this embodiment, design vectors are used based on design requirements, and feature vectors are also used for historical cases in the knowledge base. Therefore, the KNN algorithm, i.e., k-Nearest Neighbor, is used. It is a classification algorithm based on feature space similarity and is one of the simplest algorithms in machine learning. It can quickly find K historical cases with high similarity.

[0035] Furthermore, the sensitivity analysis module employs the Sobol sequence algorithm.

[0036] In this embodiment, the design requirements are subject to constraints (e.g., the core diameter must be a certain value, and its range must be within which there are multiple possible values). Therefore, the sensitivity analysis module receives the design vector X and its value range (i.e., the constraints), and then generates N sampling points in the design space using the Sobol sequence algorithm. The model is run to calculate the objective function value of each sampling point, and the sensitivity of the sampling point is calculated using variance to quantify its importance. This information is then used to determine the sensitivity of each adjusted value during the subsequent multi-algorithm fusion optimization module's process of generating the optimal solution. A sensitivity report is generated to guide the multi-algorithm fusion optimization module in focusing its efforts.

[0037] Furthermore, in the multi-algorithm fusion optimization module, the NSGA-III algorithm, MOEA-D algorithm and PSO algorithm are used in parallel iteration, and population migration and elite preservation strategies are adopted for fusion and communication during each Gming generation, and finally the Pareto optimal solution set is obtained as the best solution set.

[0038] In this embodiment, NSGA-III (the main optimizer or main algorithm) is responsible for maintaining the diversity and breadth of the solution set; it is particularly good at handling more than three optimization objectives (such as simultaneously optimizing loss, cost, volume, and weight), ensuring that the final solution can cover a variety of different performance trade-off combinations.

[0039] The MOEA / D algorithm (from the optimizer) decomposes a complex objective problem into many simple single-objective subproblems, and then performs a local fine-grained search in parallel and efficiently; it converges very quickly and can find the optimal solution in each small range rapidly.

[0040] The PSO algorithm (from the optimizer) simulates flock behavior and has a strong global exploration capability; it tends to search in unknown areas, which helps to escape local optima and avoids the algorithm from prematurely converging to a suboptimal solution set.

[0041] These three algorithms each generate and maintain their own populations (solution sets), but they do not work completely independently; instead, they cooperate through population migration and elite preservation strategies. (1) Parallel operation: The three algorithms start optimizing simultaneously, each evolving its own population based on its own rules (selection, crossover and mutation in genetic algorithms, speed update of particle swarms, etc.).

[0042] (2) Regular communication (transfer): At a preset interval (i.e., Gming generation, such as every 10 generations), the system will organize a "communication meeting" where each algorithm shares the best performing "elite individuals" (i.e. excellent design schemes) in its population with other algorithms.

[0043] (3) Complementary advantages: The elite individuals obtained by the NSGA-III algorithm can guide the MOEA / D algorithm and the PSO algorithm to explore more diverse regions. The local gems found by the MOEA / D algorithm can help the NSGA-III algorithm and the PSO algorithm refine their search more quickly. The potential new regions discovered by the PSO algorithm can provide new search directions for the NSGA-III algorithm and the MOEA / D algorithm, preventing them from getting "stuck" in local optima.

[0044] For example: Design requirements (or design vector X): such as adjustable parameters like core diameter and number of winding turns.

[0045] Objective function: The metric that needs to be minimized or maximized (e.g., f1(X) = total loss, f2(X) = total cost).

[0046] Constraints: Limits that the scheme must meet (e.g., g1(X) = temperature rise ≤ 65K).

[0047] Using the above input, the three algorithms continuously search for a design scheme that satisfies all constraints and achieves an optimal balance among multiple objective functions through an iterative process of "generating a population - evaluating the merits - generating a new population". After fusion optimization, the final output is not a single optimal solution, but a set of solutions, namely the "Pareto optimal solution set".

[0048] Characteristics: In this Pareto optimal solution set, no single solution is better than all other solutions on all objectives. For example: Option A: Lowest loss, but highest cost.

[0049] Option B: Lowest cost, but largest size.

[0050] Option C: Smallest size, but higher loss.

[0051] Value: This solution set provides decision-makers (designers) with clear trade-offs. Designers can make the most informed final choice from this high-quality set of solutions, based on the specific priorities of the project (whether efficiency or cost is the priority).

[0052] Furthermore, in the review module, the review model adopts an NLP model.

[0053] In this embodiment, the review module uses an AI agent to analyze text data (i.e., the initial task description, the optimal solution set, and the simulation report) using an NLP model. Its classification layer outputs the probability distribution of the review solutions. Then, based on the BERT architecture deep learning within the NLP model, intelligent document analysis and risk identification are achieved. This allows for the instant extraction of key information from the current design solution, automatic comparison with text rules, and the use of a weighted voting mechanism to derive the final review solution. The trained NLP model can: Entity identification: Key concepts such as "high altitude area", "external insulation distance", and "Appendix B" were identified.

[0054] Understanding the relationship: Understand that "high altitude" is the reason for "increasing the external insulation distance".

[0055] Logical verification: Automatically links to the standard original text in the knowledge base to verify whether the revised clause has been correctly applied.

[0056] For example, through knowledge graph links and semantic similarity calculations (i.e., suggestion lists), NLP models can automatically associate the description in the current design (such as "using silicon steel sheet type ABC") with the fault reports recorded in the knowledge base (such as "type ABC has increased brittleness at low temperatures") and issue a warning to the reviewer: "The current material selection is similar to historical fault cases, please pay attention to its low-temperature toughness." Therefore, the NLP model was chosen because the core basis for transformer design review is hidden in a vast amount of unstructured text (standards, specifications, reports, etc.), and NLP technology can understand and process this text on a large scale, automatically and intelligently, to achieve a high level of knowledge-driven automatic review.

[0057] Other components and operations of the R&D system based on multi-objective optimization and knowledge mining according to embodiments of the present invention are known to those skilled in the art and will not be described in detail here.

[0058] In the description of this specification, references to terms such as "embodiment," "example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, 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.

[0059] Although embodiments of the invention have been shown and described, those skilled in the art will understand 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 claims and their equivalents.

Claims

1. A research and development system based on multi-objective optimization and knowledge mining, characterized in that: include: Requirements and Task Generation Module: Receives design requirements as input, matches objective functions and constraints according to the design requirements, and forms an initial task specification from the design requirements, objective functions, and constraints; Case retrieval module: Based on the design requirements in the initial task book, retrieve multiple historical cases with high similarity from the knowledge base and generate a historical case set; Sensitivity Analysis Module: Calculates the sensitivity of each adjustment value of the design requirements in the initial task statement under the constraints, and generates a sensitivity report; Multi-algorithm fusion optimization module: Based on the initial task description, historical case set and sensitivity report, multiple algorithms are first used in parallel iterative analysis, and then fused and communicated to obtain the best solution set; Digital twin verification module: Performs multiphysics simulation analysis on each solution in the optimal solution set and generates a simulation report; Review module: The review model comprehensively reviews the initial task description, the best solution set, and the simulation report, and outputs the review plan and modification suggestions.

2. The R&D system based on multi-objective optimization and knowledge mining according to claim 1, characterized in that: The review module is also used to generate review constraints based on modification suggestions when such suggestions exist, and then feed them back to the requirement task generation module and add them to the initial task book.

3. The R&D system based on multi-objective optimization and knowledge mining according to claim 1, characterized in that: It also includes a knowledge management module: used to count the number of iterations of each review scheme given by the review module in the multi-algorithm fusion optimization module; and to count the number of modification suggestions for each review scheme given by the review module; comparing the number of iterations and the number of modification suggestions with preset number of iterations and preset number of suggestions respectively, and when the number of iterations and the number of modification suggestions are less than the preset number of iterations and preset number of suggestions respectively, the scheme is added to the knowledge base; Alternatively, once the solution is confirmed for use by the R&D personnel, it can be added to the knowledge base.

4. The R&D system based on multi-objective optimization and knowledge mining according to claim 1, characterized in that: The requirement task generation module further includes: after receiving the design requirement input, combining all the design parameters in the design requirement into a design vector.

5. The R&D system based on multi-objective optimization and knowledge mining according to claim 1, characterized in that: The execution steps of the case retrieval module include: A1: Based on the design requirements in the initial task book, use the KNN algorithm to retrieve K historical cases in the knowledge base whose similarity is within the preset similarity range; A2: For each of the K historical cases, a knowledge graph algorithm is used to traverse the nodes related to the historical case in the graph to obtain a list of suggestions for the historical case. A3: The historical case set is composed of the K historical cases in step A1 and the suggestion list in step A2.

6. The R&D system based on multi-objective optimization and knowledge mining according to claim 1, characterized in that: The sensitivity analysis module uses the Sobol sequence algorithm.

7. The R&D system based on multi-objective optimization and knowledge mining according to claim 1, characterized in that: In the multi-algorithm fusion optimization module, the NSGA-III algorithm, MOEA-D algorithm and PSO algorithm are used in parallel iteration. During each Gming generation, population migration and elite retention strategies are used for fusion and communication. Finally, the Pareto optimal solution set is obtained and used as the best solution set.

8. The R&D system based on multi-objective optimization and knowledge mining according to claim 1, characterized in that: In the review module, the review model adopts an NLP model.

Images