An engineering education system and method fusing explainable multi-objective intelligent optimization

Abstract

The application provides an engineering education system and method fusing explainable multi-objective intelligent optimization, and relates to the technical fields of intelligent education and artificial intelligence optimization. The method comprises the following steps: obtaining an initial engineering design scheme of a user, and obtaining a standardized multi-dimensional index vector through physical simulation analysis and logical analysis; configuring weights with structured reasons for each index, identifying critical weight values leading to decision reversal through sensitivity analysis, and then determining optimization objectives and constraint conditions; performing multi-objective optimization search using an explainable multi-objective optimizer, generating a Pareto optimal solution set and outputting explainable data; simultaneously generating a behavior interaction sequence of the user; generating a capability evaluation report of the user according to the behavior interaction sequence; and adjusting a preset teaching logic sequence according to the capability evaluation report to form a closed-loop teaching process. The application is suitable for cultivating and evaluating the decision-making ability of students in balancing technical performance and environmental sustainability in engineering design.

Description

Technical Field

[0001] This invention relates to the fields of intelligent education and artificial intelligence optimization technology, and in particular to an engineering education system and method that integrates interpretable multi-objective intelligent optimization. Background Technology

[0002] In response to the demands of global climate change and the green transformation of manufacturing, cultivating engineers capable of systematically balancing performance, cost, and the environmental impact throughout the product lifecycle in product design has become one of the core objectives of modern engineering education. However, existing technological tools and teaching methods have the following technical shortcomings in supporting the cultivation of this comprehensive ability:

[0003] First, there is a significant disconnect between data and workflow. Current engineering practices rely on multiple independent software toolchains: Computer-aided design (CAD) and engineering simulation (CAE) software primarily handle geometric modeling and physical performance (such as structural, fluid, and thermal) analysis, and their data models typically do not include environmental load parameters. Life cycle assessment (LCA) and carbon footprint calculation tools, on the other hand, often operate as independent evaluation modules after design completion, and their calculation interfaces lack deep compatibility with mainstream design software. Therefore, sustainability indicators such as material carbon emissions and processing energy consumption cannot be directly embedded as native variables into the design iteration cycle, forcing the workflow to follow a sequential "performance design first, environmental assessment later" model. This model hinders students from developing a design mindset that integrates technological, environmental, and economic factors.

[0004] Secondly, there is a lack of transparency in the decision support mechanism. When using multi-objective optimization algorithms to automatically optimize designs, such as using multi-objective evolutionary algorithms to search for the Pareto front, the algorithm itself is a "black box" model. It can output a series of optimal trade-off solutions, but it cannot reveal the decision-making logic that leads to these solutions. For example, it cannot clearly answer how much a change in a design variable contributes to a specific performance or environmental goal; nor can it clearly quantify the degree of environmental compromise required to improve a certain performance. In a teaching context, students can only passively accept the results given by the algorithm, unable to understand its complex multi-objective trade-off mechanism, thus making it difficult to internalize the core engineering thinking ability of trade-offs through the use of tools.

[0005] Finally, there are problems with the lack of dimensions and high subjectivity in the assessment of teaching effectiveness and competence. Traditional teaching evaluation relies heavily on teachers' subjective interpretation and scoring of the final static design report. Dynamic behavioral data such as the evolution of students' decision preferences throughout the design exploration process, their search paths in the multidimensional solution space, and their comparison logic of different solutions cannot be effectively captured and structured by existing technical systems. Competency assessment lacks objective and quantitative process data support, and the evaluation results are greatly influenced by teachers' personal experience, resulting in high volatility and difficulty in achieving accurate and consistent feedback, thus restricting the large-scale and efficient cultivation of engineering innovation talents.

[0006] Analysis of existing patents and technical literature reveals that most solutions focus on single-dimensional improvements: for example, enhancing the visualization and interaction of the simulation process, optimizing collaborative workflows in engineering project management, or improving the search performance and convergence speed of multi-objective optimization algorithms. Currently, no systematic and integrated technical solution has emerged to address the three interconnected problems of data fragmentation, black-box optimization, and subjective evaluation. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention proposes an engineering education system and method that integrates interpretable multi-objective intelligent optimization by designing an architecture with deep collaboration among functional modules and a closed-loop data flow. This system integrates an interpretable multi-objective optimization model with a quantitative evaluation mechanism for the decision-making process. The aim is to construct a closed-loop teaching environment that encompasses the entire process from design parameter input, integrated feature quantification, transparent optimization exploration, to quantitative evaluation of the decision-making process and adaptive feedback. This environment is suitable for cultivating and evaluating students' decision-making ability to balance technical performance and environmental sustainability in engineering design.

[0008] On the one hand, this invention proposes an engineering education system that integrates interpretable multi-objective intelligent optimization, the system comprising:

[0009] The sustainability quantification and multi-source feature fusion module is used to obtain the initial engineering design scheme provided by student users and extract the set of design variables for that scheme. ; respectively for the design variable set Physical simulation analysis and logical analysis based on design features and process knowledge are performed to obtain a standardized multi-dimensional index vector of the initial engineering design scheme. ;

[0010] The structured display and decision-making process simulation module is used to standardize multi-dimensional indicator vectors. Each indicator in the algorithm is assigned a weight with a structured rationale, and a weighted comprehensive utility value is calculated. Sensitivity analysis is used to identify the critical weight values ​​that lead to decision reversal, thereby assisting student users in determining optimization objectives and constraints.

[0011] An interpretable multi-objective collaborative optimization module is used to optimize the set of design variables using an interpretable multi-objective optimizer based on the optimization objective and constraints. Perform multi-objective optimization search to generate a Pareto optimal solution set and simultaneously output interpretable data; at the same time, in response to the student user's selection operation on the Pareto optimal solution set, generate a behavioral interaction sequence.

[0012] The decision-making process tracking and dynamic cognitive evaluation module is used to calculate the multi-dimensional ability indicators of student users based on the behavioral interaction sequence and using a reinforcement learning-based evaluation model, and generate an ability evaluation report for student users.

[0013] The teaching process intelligent controller is used to schedule and drive the startup, execution and data exchange of the sustainability quantification and multi-source feature fusion module, the interpretable multi-objective collaborative optimization module, and the decision process tracking and dynamic cognitive evaluation module according to the teaching logic sequence preset by the teacher user, so as to execute the serialized teaching tasks, and adjust the teaching logic sequence in real time according to the Pareto optimal solution set, interpretable data and ability evaluation report.

[0014] The intelligent controller for the teaching process is electrically connected to the sustainability quantification and multi-source feature fusion module, the structured display and decision-making process simulation module, the interpretable multi-objective collaborative optimization module, and the decision-making process tracking and dynamic cognitive evaluation module.

[0015] Furthermore, the respective sets of design variables... Physical simulation analysis and logical analysis based on design features and process knowledge are performed to obtain a standardized multi-dimensional index vector of the initial engineering design scheme. The specific content is as follows:

[0016] By calling the physics simulation engine, the set of design variables is... Physical simulation analysis is performed to obtain the technical performance index vector of the initial engineering design scheme. ;

[0017] Based on the design feature parsing rule base and material-process knowledge graph, the design variable set... Logical analysis based on design features and process knowledge is performed to obtain the sustainability index vector of the initial engineering design scheme. ;

[0018] according to and The standardized multi-dimensional index vector of the initial engineering design scheme is obtained. .

[0019] Furthermore, the design variable set is analyzed based on the design feature parsing rule base and the material-process knowledge graph. Logical analysis based on design features and process knowledge is performed to obtain the sustainability index vector of the initial engineering design scheme. The specific content is as follows:

[0020] Utilize predefined feature recognition rules to construct a design feature parsing rule base;

[0021] Using the aforementioned design feature parsing rule base, the design variable set is... Perform process semantic recognition to obtain the manufacturing features and assembly relationships of the initial engineering design scheme, and map the obtained manufacturing features and assembly relationships into a list of manufacturing activities of the initial engineering design scheme;

[0022] Construct a material-process knowledge graph to store material entities, processing technology entities, and environmental load factors associated with each entity;

[0023] Based on the list of manufacturing activities, the environmental load factor corresponding to each manufacturing activity is obtained by querying the material-process knowledge graph;

[0024] Based on the set of design variables Calculate the geometric parameters of the initial engineering design scheme, and establish a mapping model between the geometric parameters and the environmental load factors based on the preset environmental load factor mapping relationship library;

[0025] Based on the mapping model, the sustainability index of the initial engineering design scheme is calculated. .

[0026] Furthermore, the specific content of the structured display and decision-making process simulation module is as follows:

[0027] For standardized multidimensional indicator vectors For any indicator, assign a weight to that indicator;

[0028] Based on the pre-set structured reason knowledge base, a structured reason is associated with the weighting behavior of the indicator, generating triple data containing the indicator, weight, and structured reason;

[0029] Based on the triplet data, calculate the weighted comprehensive utility value of the current engineering design scheme;

[0030] Based on the weighted comprehensive utility value, sensitivity analysis is used to identify the critical weight value that leads to a decision reversal in the current engineering design scheme.

[0031] Student users, based on the aforementioned critical weight values, select from a standardized multi-dimensional indicator vector. Several indicators are selected as optimization objectives, and all unselected indicators are used as constraints.

[0032] Furthermore, the specific content of the interpretable multi-objective collaborative optimization module is as follows:

[0033] Under the condition that the constraints are met, and using the optimization objective as the standard, an interpretable multi-objective optimizer is employed to optimize the set of design variables. Perform multi-objective optimization search to generate a Pareto optimal solution set, and simultaneously output the interpretability data corresponding to each solution in the Pareto optimal solution set;

[0034] The interpretable multi-objective optimizer mentioned above is implemented using an interpretable optimizer based on the Transformer neural network architecture;

[0035] The generated Pareto optimal solution set is provided to student users, who can then select to view interpretable data and standardized multi-dimensional index vectors corresponding to any solution in the Pareto optimal solution set. A solution is selected from the Pareto optimal solution set as the final engineering design scheme, and a structured demonstration report is generated.

[0036] During the process of student users making selections and viewings, a sequence of behavioral interactions recorded in chronological order is generated based on the student users' selections.

[0037] Furthermore, the method for constructing the interpretable optimizer based on the Transformer neural network architecture is as follows:

[0038] Construct a Transformer model, including a decoder and an encoder; wherein the input to the encoder includes: constraints, optimization objective, and a set of design variables. The encoder is used to map the encoder input to a high-dimensional semantic feature space using a multi-head attention mechanism to obtain the encoder output, and to extract the attention weight matrix generated by the multi-head attention mechanism. The decoder is used to decode the output of the encoder and generate a sequence of candidate engineering design schemes.

[0039] A multi-task loss function with an interpretability regularization term is used to train the Transformer model, resulting in a well-trained Transformer model, which is then used as an interpretable optimizer based on the Transformer neural network architecture.

[0040] The multi-task loss function is a weighted sum of a performance loss term, a diversity loss term, and an interpretability regularization term. The performance loss term drives the Pareto solution set generated by the interpretable optimizer to converge to the true Pareto front. The diversity loss term maximizes the coverage of the Pareto solution set generated by the interpretable optimizer in the target space. The interpretability regularization term applies structured constraints to the attention weight matrix, including at least one of sparsity constraints, block structure constraints, or alignment constraints with the engineering prior knowledge matrix.

[0041] Furthermore, the interpretability data includes: an attention weight matrix. Variable-objective relationship heatmap, list of key decision factors, and text describing trade-offs;

[0042] The attention weight matrix , used to characterize the relationship between design variables and optimization objectives;

[0043] For the attention weight matrix Normalization and visualization mapping are performed to generate a variable-target correlation heatmap; the variable-target correlation heatmap is used to display the correlation between the design variables and the optimization target corresponding to the solution currently selected by the student user;

[0044] The list of key decision factors is as follows: Based on the variable-objective correlation heatmap, the influence of each design variable on the current Pareto optimal solution set or a specific solution is obtained, and the design variables are sorted in descending order of influence to obtain the sorting result;

[0045] The trade-off relationship description text is generated by using a predefined text template or a lightweight natural language generation model, based on the variable-objective correlation heatmap and the objective function value of the current solution, to describe the trade-off relationship between design variables and optimization objectives in the current engineering design scheme.

[0046] Furthermore, the specific content of the decision-making process tracking and dynamic cognitive evaluation module is as follows:

[0047] Define a Markov Decision Process (MDP) as a quintuple. ;in For state space; For action space; The state transition probability; For the reward function; As a discount factor, and ;

[0048] The state space Used for storage State vector at time step The state vector mentioned above This includes: the current standardized multi-dimensional indicator vector Weights of each indicator The coordinates of the Pareto solution that the current student user is viewing or selecting. and historical behavior summary information;

[0049] The action space Used for storage Moment of action The actions described therein For student users The interactive behavior at any given moment is obtained from the sequence of behavioral interactions;

[0050] The state transition probability , used to describe the state Next action Afterwards, transition to the new state. The probability of;

[0051] The reward function Used to monitor student users' status The following actions Conduct immediate value assessment;

[0052] The reward logic defined in the reward function includes at least one of the following behaviors:

[0053] Value clarification rewards are used to positively reward actions that increase the weight of sustainability indicators and provide reasonable justifications.

[0054] The depth exploration reward is used to give a positive reward when a student user continuously selects or views a Pareto solution and the Euclidean distance in the target space is greater than a preset threshold.

[0055] The causal inquiry reward is used to give positive rewards to student users who actively trigger the viewing of the interpretable data;

[0056] Logical consistency reward is used to give positive rewards to behaviors that are consistent with the value preferences and exploration trajectory stated in the early stage when choosing the final engineering design scheme.

[0057] Based on the quintuple The behavioral interaction sequence is modeled as the trajectory of a Markov Decision Process (MDP), and the cumulative discount reward for student users is calculated.

[0058] The cumulative discount rewards are decomposed into sources, and multi-dimensional ability indicators of student users are calculated based on the decomposition results; wherein the multi-dimensional ability indicators include at least one of the following: decision consistency index, exploration completeness, sustainability value sensitivity, and causal analysis depth.

[0059] Based on a pre-set competency evaluation report template, a competency evaluation report for student users is generated according to the multi-dimensional competency indicators of student users.

[0060] Furthermore, the engineering teaching system integrating interpretable multi-objective intelligent optimization also includes: a user interface; the user interface is electrically connected to the teaching process intelligent controller, and is used to display the output data of each module in the sustainability quantification and multi-source feature fusion module, the interpretable multi-objective collaborative optimization module, and the decision process tracking and dynamic cognitive evaluation module pushed by the teaching process intelligent controller to student users, and to receive interactive operation instructions from student users.

[0061] On the other hand, this invention proposes an engineering education method that integrates interpretable multi-objective intelligent optimization, implemented using the aforementioned engineering education system that integrates interpretable multi-objective intelligent optimization, comprising the following processes:

[0062] Based on the teaching logic sequence preset by the teacher user, deploy and start the engineering education system that integrates interpretability and multi-objective optimization;

[0063] Obtain the initial engineering design scheme provided by student users and extract the set of design variables for that scheme;

[0064] Physical simulation analysis and logical analysis based on design features and process knowledge are performed on the set of design variables respectively to obtain a standardized multi-dimensional index vector of the initial engineering design scheme.

[0065] Assign weights with structured rationale to each indicator in the standardized multidimensional indicator vector, and calculate the weighted comprehensive utility value;

[0066] Based on the weighted comprehensive utility value, sensitivity analysis is used to identify the critical weight values ​​that lead to decision reversal, thereby assisting student users in determining optimization objectives and constraints.

[0067] Based on the optimization objective and constraints, an interpretable multi-objective optimizer is used to perform multi-objective optimization search on the set of design variables, generate a Pareto optimal solution set, and simultaneously output interpretable data.

[0068] Student users can select to view interpretability data corresponding to any solution in the Pareto optimal solution set, choose the final engineering design scheme from the Pareto optimal solution set, and generate a sequence of behavioral interactions based on the student user's selection.

[0069] Based on the behavioral interaction sequence, a reinforcement learning-based evaluation model is used to calculate the multi-dimensional ability indicators of student users and generate an ability evaluation report for student users.

[0070] Based on student users' competency evaluation reports, the pre-set teaching logic sequence for teachers is adjusted, and then the engineering education system is redeployed to form a closed-loop teaching process.

[0071] The beneficial effects of adopting the above technical solution are as follows:

[0072] This invention proposes to map design features into calculable sustainability indicators in real time, achieving synchronous generation with technical performance indicators; it integrates a multi-objective optimization engine with internal causal explanation capabilities, making the optimization process transparent; and it achieves quantitative evaluation of the decision-making process by tracking and analyzing the interaction behavior sequences of student users. Based on this, this invention, by constructing a system architecture with deep collaboration among functional modules and a closed-loop data flow, produces the following significant technical effects, specifically:

[0073] (1) This invention realizes the integration and parallel computing of the data sources of engineering design parameters and full life cycle environmental indicators, and solves the problem of data and process separation.

[0074] In current engineering design and teaching practices, the technical performance indicators and sustainability indicators of a product are usually calculated and analyzed separately by independent software tools. This separation prevents the integrated consideration and simultaneous optimization of technical and sustainability indicators during the design phase.

[0075] This invention establishes a real-time computational pathway for generating sustainability indicators directly and in real-time from original design features by introducing an automated sustainability quantification engine comprised of a design feature parsing rule base and a material-process knowledge graph. Specifically, using preset parsing rules, it automatically identifies manufacturing features (such as holes, slots, and stiffeners) and assembly relationships in the input design model, generating a series of structured manufacturing activity descriptions. Simultaneously, by querying the material-process knowledge graph, it obtains the environmental load factor corresponding to each manufacturing activity. Based on the automated sustainability quantification engine, the original design parameter set can be automatically and in real-time transformed into a standardized multi-dimensional feature vector. This vector simultaneously includes technical performance indicator sub-vectors (such as strength, weight, and natural frequency) and sustainability indicator sub-vectors (such as implicit carbon emissions and manufacturing energy consumption). This process changes the traditional sequential model of sustainability assessment as an independent post-hoc analysis in engineering practice, seamlessly embedding environmental considerations into the starting point of the design iteration cycle. From a technical perspective, this not only enables the synchronous calculation of technical and environmental attributes from the data source, eliminating information silos and laying a data foundation for subsequent integrated considerations, but also forces a synchronous thinking mode at the process level, providing a basic and integrated technical environment for the integrated consideration of multiple factors such as technology, environment, and economy in the early stages of design.

[0076] (2) This invention realizes the transformation of the multi-objective optimization process from black box result output to white box causal mechanism exploration, and solves the problem of opaque optimization decision-making.

[0077] Traditional multi-objective optimization algorithms (such as multi-objective evolutionary algorithms) do not present their internal decision-making logic to the user during the solution process. While these algorithms can output a set of Pareto optimal solutions, they cannot explain why each solution represents an optimal trade-off, nor can they clarify how each design variable specifically affects conflicting objectives. This makes it difficult for students to understand the trade-off principles of multi-dimensional objectives behind the optimization results, hindering the construction of deeper cognitive understanding.

[0078] This invention employs an improved optimizer based on the Transformer neural network architecture to construct a transparent multi-objective collaborative optimization environment that outputs an internal causal explanation. Specifically, during the optimizer's training phase, this invention introduces a specially designed multi-task loss function. Besides a performance loss term that approximates the true Pareto front by driving the solution set, this loss function innovatively integrates an interpretability regularization term. This regularization term imposes specific structured constraints (e.g., encouraging sparsity, block structure, or alignment with engineering prior knowledge) on the attention weight matrix generated by the optimizer's internal self-attention mechanism, forcing the model to learn and generate easily interpretable internal representations while searching for optimal solutions. Thus, while efficiently searching for Pareto optimal solutions, the optimizer's internal self-attention mechanism is constrained to learn and generate a clearly structured and easily interpretable attention weight matrix.

[0079] This invention maps the attention weight matrix into a variable-objective correlation heatmap and deeply integrates it with an interactive Pareto front visualization interface for real-time interaction. This allows student users to instantly observe the key design variables and their influence strength as they explore different solution points on the frontier. The heatmap intuitively reveals the relative influence or contribution of each design variable to different optimization objectives (including conflicting technical and sustainability objectives), thus transforming the numerical computation process within the optimizer into causal logical relationships that student users can observe and understand. This technical feature transforms the implicit, abstract trade-off logic of traditional optimization algorithms into a visualized causal chain that student users can observe and interactively explore in real time. The effect is that the optimizer transforms from a "black box" tool that only provides answers into a "white box" exploration tool that can reveal the intrinsic relationships between design variables and conflicting objectives, greatly lowering the cognitive threshold for understanding the trade-offs in complex systems and realizing the substantial application of artificial intelligence interpretability technology in engineering education.

[0080] (3) This invention constructs a high-level engineering decision-making ability quantitative assessment and adaptive feedback closed loop based on user interaction behavior sequence tracking, which solves the problem of subjective assessment of ability.

[0081] Current assessments of students' abilities in sustainable engineering design primarily rely on qualitative analysis and subjective grading by instructors of the final submitted design reports. There is a lack of objective, quantitative methods for recording and analyzing students' decision-making processes, making it impossible to systematically evaluate their exploration strategies, preference evolution, and logical consistency within the solution space.

[0082] This invention establishes a quantitative evaluation system for engineering decision-making ability based on user interaction behavior sequence tracking by constructing a dynamic cognitive evaluation model based on a reinforcement learning evaluation framework. Specifically, this invention formally models the sequence of student user interactions (such as adjusting weights, selecting solution points, and viewing heatmaps) in a complete training task as a Markov Decision Process (MDP), and performs real-time value evaluation on advanced cognitive behaviors such as representational depth exploration, causal inquiry, and logical consistency based on a pre-set reward function based on cognitive science principles. By calculating the cumulative discounted reward of the user behavior sequence and further decomposing it into multi-dimensional quantitative indicators such as decision consistency index, exploration completeness, and causal analysis depth, an objective digital profile of cognitive ability is automatically generated, thereby transforming teaching evaluation from relying on subjective experience judgment to analysis based on objective behavioral data and computational models.

[0083] This invention overcomes the limitations of traditional subjective qualitative evaluations that rely on final reports for decision-making ability, achieving objective and process-oriented assessment based on full-process behavioral data. Furthermore, the intelligent controller for the teaching process, based on this quantitative evaluation result, triggers personalized learning path recommendations and adaptive teaching interventions (e.g., guiding exploration or reinforcing explanations for weak areas), forming a precise teaching closed loop of "assessment-feedback-intervention." This effect advances the cultivation of engineering decision-making talent from a static, experience-based model to a new stage of dynamic, precise navigation driven by data intelligence. Attached Figure Description

[0084] Figure 1 This is a schematic diagram of the overall architecture of an engineering education system that integrates interpretability and multi-objective optimization in this embodiment;

[0085] Figure 2 This is a schematic diagram comparing traditional black-box optimization with the interpretable optimization strategy in this embodiment;

[0086] Figure 3 This is a schematic diagram illustrating the internal workflow of the multi-objective optimizer in this embodiment;

[0087] Figure 4 This is a flowchart of the quantitative evaluation calculation for decision evaluation based on Markov decision process in this embodiment;

[0088] Figure 5 This is a flowchart of the adaptive intervention process for teaching based on the quantitative evaluation of sustainable decision-making ability in this embodiment;

[0089] Figure 6 This is a flowchart of an engineering education method that integrates interpretability and multi-objective optimization in this embodiment;

[0090] Figure 7 This is a flowchart illustrating the implementation process of a sustainable design training case based on the optimization of a single-degree-of-freedom stamping vibration reduction system in this implementation. Detailed Implementation

[0091] To facilitate understanding of this application, specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and embodiments. The following embodiments are illustrative of the invention but are not intended to limit its scope. Rather, these embodiments are provided to provide a more thorough and complete understanding of the disclosure of this application.

[0092] Example 1:

[0093] This embodiment presents an engineering education system that integrates interpretable multi-objective optimization, such as... Figure 1 As shown, the system includes: a sustainability quantification and multi-source feature fusion module, a structured display and decision-making process simulation module, an interpretable multi-objective collaborative optimization module, a decision-making process tracking and dynamic cognitive evaluation module, and a teaching process intelligent controller.

[0094] In this embodiment, the entire system can be constructed using a layered and modular software architecture. Its core lies in encapsulating four core functions—sustainability quantification and multi-source feature fusion, interpretable multi-objective collaborative optimization, decision process tracking and dynamic cognitive evaluation, and intelligent control of the teaching process—into relatively independent services or components that communicate through well-defined interfaces.

[0095] The sustainability quantification and multi-source feature fusion module is used to obtain the initial engineering design scheme provided by student users and extract the set of design variables for the scheme. ; respectively for the design variable set Physical simulation analysis and logical analysis based on design features and process knowledge are performed to obtain a standardized multi-dimensional index vector of the initial engineering design scheme. .

[0096] In this embodiment, the sustainability quantification and multi-source feature fusion module is the core of the system for achieving the fusion of engineering design parameters and full life-cycle environmental data sources. This module integrates two key knowledge components: a design feature parsing rule base and a materials-process knowledge graph. Simultaneously, this module calls integrated or external physical simulation engines to perform technical performance analysis on the same design model and calculate technical performance indicators.

[0097] The respective design variable sets Physical simulation analysis and logical analysis based on design features and process knowledge are performed to obtain a standardized multi-dimensional index vector of the initial engineering design scheme. The specific content is as follows:

[0098] By calling the physics simulation engine, the set of design variables is... Physical simulation analysis is performed to obtain the technical performance index vector of the initial engineering design scheme. .

[0099] In this embodiment, the physical simulation engine can be an integrated open-source solver (such as CalculiX for FEA) or a commercial software kernel (such as ANSYS or Abaqus) called through an application programming interface (API). The key is that it can automatically perform simulation calculations based on the input design parameters without the need for manual preprocessing.

[0100] Based on the design feature parsing rule base and material-process knowledge graph, the design variable set... Logical analysis based on design features and process knowledge is performed to obtain the sustainability index vector of the initial engineering design scheme. .

[0101] The design variable set is based on a design feature parsing rule base and a material-process knowledge graph. Logical analysis based on design features and process knowledge is performed to obtain the sustainability index vector of the initial engineering design scheme. The specific content is as follows:

[0102] A design feature parsing rule base is constructed using predefined feature recognition rules.

[0103] In this embodiment, 3D software such as UG is used to identify the dimensional parameters of the design object, such as radius, length, width, and thickness, as well as material parameters and external load information, thereby obtaining a set of variable parameters used to describe the input and output relationship of the calculation model, which serves as a predefined feature recognition rule.

[0104] Using the aforementioned design feature parsing rule base, the design variable set is... Perform process semantic recognition to obtain the manufacturing features and assembly relationships of the initial engineering design scheme, and map the obtained manufacturing features and assembly relationships into a list of manufacturing activities of the initial engineering design scheme.

[0105] The manufacturing activity list is used to structurally describe the processing technology and assembly relationships involved in the initial engineering design scheme, as well as the resource consumption corresponding to each processing technology and assembly relationship.

[0106] In this embodiment, the design feature parsing rule base contains a series of predefined feature recognition rules for automatically identifying and analyzing the input 3D design model or set of design variables. These rules can identify manufacturing features (such as holes, slots, chamfers, bosses, stiffeners, thin walls, etc.) and assembly relationships (such as mating, connection, welding, etc.) in the input data, and map this geometric and structural information into a quantifiable list of manufacturing and assembly activities. For example, identifying a "through hole" feature can map to a "drilling" machining activity; identifying a "thin wall" feature may map to a "milling" or "sheet metal forming" activity; identifying a "welding assembly" relationship maps to the corresponding "welding" activity and the required welding materials and energy consumption.

[0107] Construct a material-process knowledge graph to store material entities, processing technology entities, and environmental load factors associated with each entity.

[0108] In this embodiment, the materials-process knowledge graph is constructed in the form of a graph database to store material entities (such as steel grades and plastic types), processing technology entities (such as turning, milling, and injection molding), and the relationships between each entity and environmental load factors (such as implicit carbon emission factors per unit mass and energy consumption factors per unit processing time). In the materials-process knowledge graph, nodes represent entities or environmental load factors associated with each entity, and edges represent relationships (such as "material - can be used in - process", "process - has - unit energy consumption factor", "material - has - unit carbon emission factor"). This graph encapsulates environmental load data for material production and processing. Its data sources can be built and maintained based on publicly available life cycle assessment databases (such as Ecoinvent), engineering materials handbooks, and industry energy consumption statistics, and the timeliness and authority of the data are ensured through a version update mechanism.

[0109] Based on the list of manufacturing activities, the environmental load factor corresponding to each manufacturing activity is obtained by querying the material-process knowledge graph.

[0110] In this embodiment, an efficient query interface is provided to quickly return the corresponding environmental impact data based on the input bill of materials and process route (i.e., list of manufacturing activities). Specifically, based on the type of material selected in the design, the material-process knowledge graph is queried to obtain the environmental load factor (such as carbon emission factor and energy consumption factor) corresponding to each manufacturing activity.

[0111] Based on the set of design variables The geometric parameters of the initial engineering design scheme are calculated, and a mapping model between the geometric parameters and the environmental load factors is established based on the preset environmental load factor mapping relationship library.

[0112] In this embodiment, sustainability indicators are typically functions of geometric parameters. For example, when manufacturing similar parts, larger volumes result in higher energy consumption and carbon emissions. The mapping model is usually characterized as a nonlinear functional relationship of geometric dimensions; more dimensions lead to greater complexity in processing and manufacturing, resulting in higher resource consumption.

[0113] Based on the mapping model, the sustainability index of the initial engineering design scheme is calculated. .

[0114] In this embodiment, based on the mapping model, a set of sustainability indicators is obtained by weighting geometric parameters (such as volume, area, and length). (such as total implicit carbon emissions, total manufacturing energy consumption, material utilization rate, etc.).

[0115] according to and The standardized multi-dimensional index vector of the initial engineering design scheme is obtained. .

[0116] It should be noted that, in actual use, before student users submit their initial engineering design schemes, the engineering teaching system has pre-set the engineering logic framework of the project, such as the vibration reduction design of a stamping press, within the sustainability quantification and multi-source feature fusion module. This framework includes: design objectives, design variables, constraints on the variables, and models or functions describing the mapping relationship between the variables and the design objectives. When student users submit their initial engineering design schemes to the sustainability quantification and multi-source feature fusion module, the module extracts the corresponding design variables and their specific values ​​from the scheme based on the pre-set engineering logic framework. Then, through physical simulation analysis and logical parsing based on design features and process knowledge, it obtains a standardized multi-dimensional index vector for the initial engineering design scheme. .

[0117] In this embodiment, through the aforementioned parallel processing, the module achieves the transformation from the original design parameters to a standardized multidimensional feature vector. This transformation process can be formally described as a feature fusion mapping function. ,Right now:

[0118]

[0119] in, This represents the predefined feature recognition rules in the feature parsing rule base. It represents the constraints and relationships contained in the material-process knowledge graph. This module solves the problem of environmental load data and physical performance data being separated at the source and unable to be calculated and analyzed synchronously in traditional engineering processes by outputting a standardized initial scheme multi-dimensional indicator vector that integrates technical and sustainability dimensions.

[0120] The structured display and decision-making process simulation module is used to provide standardized multi-dimensional indicator vectors. Each indicator in the algorithm is assigned a weight with a structured rationale, and a weighted comprehensive utility value is calculated. Sensitivity analysis is used to identify the critical weight values ​​that lead to decision reversal, thereby assisting student users in determining optimization objectives and constraints.

[0121] The specific content of the structured display and decision-making process simulation module is as follows:

[0122] For standardized multidimensional indicator vectors For any indicator, assign a weight to that indicator.

[0123] In this embodiment, the system guides student users to use standardized multi-dimensional indicator vectors. Conduct an importance assessment, which involves assigning a weight to each indicator. The initial weight allocation is based on, but is not limited to, the student's understanding of the current engineering problem, the priority set by the teaching project itself, and at least one of the industry norms and standards.

[0124] Based on a pre-defined knowledge base of structured reasons, a structured reason is associated with the weighting behavior of the indicator, generating a triplet data containing the indicator, weight, and structured reason.

[0125] In this embodiment, to promote deep thinking and make the decision-making process traceable, the system mandates that student users associate each weight assignment with a structured reason (selected from a preset structured reason knowledge base or entered according to specifications). The system persistently stores the resulting (indicator, weight, structured reason) triple data.

[0126] The structured reasoning knowledge base is used to provide predefined structured reasoning options for student users to select or reference when performing indicator weighting.

[0127] Based on the triplet data, calculate the weighted comprehensive utility value of the current engineering design scheme.

[0128] In this embodiment, the system is based on the weight distribution currently set by the student user. , To determine the number of indicators, a multi-criteria decision analysis is performed on the current engineering design scheme to calculate its weighted comprehensive utility value. The comprehensive utility value is a weighted sum of technical and sustainability indicators, which are assigned their respective weights and then superimposed to obtain the comprehensive indicator. For example, technical indicators might include the lowest vibration amplitude or the highest natural frequency, while sustainability indicators might include the lowest carbon emissions. These two indicators are usually not directly related. For instance, increasing the natural frequency means increasing structural stiffness for the same mass, such as by adding stiffeners. However, this might maximize the energy consumption of structural manufacturing, thus increasing carbon emissions. Therefore, changes in weights will lead to different design schemes. This section identifies the critical weight values ​​that cause significant changes in the design scheme.

[0129] Based on the weighted comprehensive utility value, sensitivity analysis is used to identify the critical weight value that leads to a decision reversal in the current engineering design scheme.

[0130] In this embodiment, the system uses a sensitivity analysis tool to visually demonstrate to student users through a user interface how the overall utility value or the ranking of options fluctuates when the weight of a certain indicator is fine-tuned, and identifies the critical weight value that leads to a reversal in decision-making. This process aims to encourage student users to reflect on and clarify their own value judgment logic.

[0131] Student users, based on the aforementioned critical weight values, select from a standardized multi-dimensional indicator vector. Several indicators are selected as optimization objectives, and all unselected indicators are used as constraints.

[0132] In this embodiment, after seeing the critical weight value for decision reversal in the current engineering design scheme, the student user, guided by the system, proceeds from... Several indicators are selected as optimization objectives, such as minimizing "vibration transmissibility" and minimizing "implicit carbon emissions," and other indicators are defined as constraints, thus formally defining a multi-objective optimization problem. The intelligent controller then calls the interpretable multi-objective collaborative optimization module to perform the calculation.

[0133] The interpretable multi-objective collaborative optimization module is used to optimize the set of design variables using an interpretable multi-objective optimizer based on the optimization objective and constraints. A multi-objective optimization search is performed to generate a Pareto optimal solution set, and interpretable data is output simultaneously. At the same time, in response to the student user's selection operation on the Pareto optimal solution set, a behavioral interaction sequence is generated.

[0134] The specific content of the explainable multi-objective collaborative optimization module is as follows:

[0135] like Figure 2As shown, under the constraints, and using the optimization objective as the standard, an interpretable multi-objective optimizer is employed to optimize the set of design variables. Perform multi-objective optimization search to generate a Pareto optimal solution set, and simultaneously output the interpretability data corresponding to each solution in the Pareto optimal solution set.

[0136] The interpretable multi-objective optimizer mentioned above is implemented using an interpretable optimizer based on the Transformer neural network architecture;

[0137] In this embodiment, the core of the interpretable multi-objective collaborative optimization module is an interpretable multi-objective optimizer based on the Transformer neural network architecture and specially trained. The optimizer adopts the Transformer architecture because its core multi-head self-attention mechanism can effectively model the dependency relationship between any two elements in the input sequence. In multi-objective optimization problems, there are often complex and non-linear interactions between design variables and optimization objectives. The self-attention mechanism, by calculating the attention weights between variables and objectives, can capture and quantify these influences, providing a data foundation for subsequently generating interpretable relationships.

[0138] The method for constructing the interpretable optimizer based on the Transformer neural network architecture is as follows:

[0139] Construct a Transformer model, including a decoder and an encoder; wherein the input to the encoder includes: constraints, optimization objective, and a set of design variables. The encoder is used to map the encoder input to a high-dimensional semantic feature space using a multi-head attention mechanism to obtain the encoder output, and to extract the attention weight matrix generated by the multi-head attention mechanism. The decoder is used to decode the output of the encoder and generate a sequence of candidate engineering design schemes.

[0140] In this embodiment, the encoder-decoder structure of the Transformer model is used as the basic architecture. The encoder maps the input sequence of design variables to a high-dimensional semantic feature space. The decoder, in an autoregressive manner, progressively generates a sequence of candidate engineering design schemes based on the encoder's output and the currently generated partial sequences. The core component of this architecture, the multi-head self-attention mechanism, provides the physical basis for generating interpretable data. During the optimization iteration process, the self-attention layer dynamically calculates the correlation strength between design variables and between design variables and the optimization objective. This calculation process can be represented as follows:

[0141]

[0142] in, For query matrix; The key matrix; It is a value matrix; , , All are obtained by linear transformation of the input design variable sequence and target information; Let be the dimension of the key vector. Extract the generated attention weight matrix. In other words, the result of softmax quantifies the correlation between different parts of the input.

[0143] A multi-task loss function with an interpretability regularization term is used to train the Transformer model, resulting in a well-trained Transformer model, which is then used as an interpretable optimizer based on the Transformer neural network architecture.

[0144] The multi-task loss function is a weighted sum of a performance loss term, a diversity loss term, and an interpretability regularization term. The performance loss term drives the Pareto solution set generated by the interpretable optimizer to converge to the true Pareto front. The diversity loss term maximizes the coverage of the Pareto solution set generated by the interpretable optimizer in the target space. The interpretability regularization term applies structured constraints to the attention weight matrix, including at least one of sparsity constraints, block structure constraints, or alignment constraints with the engineering prior knowledge matrix.

[0145] To enable the optimizer to generate easily interpretable attention patterns after training, this embodiment employs a multi-task loss function with interpretability constraints. Train it. Multi-task loss function. It is a weighted sum of multiple sub-loss terms, expressed as:

[0146]

[0147] in, This is a performance loss term, whose main function is to drive the solution set generated by the interpretable optimizer to approximate the true Pareto front as closely as possible, thereby ensuring the quality of the solution (such as convergence). This is a diversity loss term used to ensure that the generated Pareto solutions are uniformly distributed in the target space to cover different trade-off regions; This is the interpretability regularization term, which is also the key to this method. This term is applied to the attention weight matrix generated by the self-attention layer inside the interpretable optimizer. Imposing structured constraints, such as: (1) Sparsity constraints: encouraging attention weight matrix Most elements are close to zero, forcing the model to focus on a few key design variables and their relationship with the objective, which is easier to interpret. (2) Prior knowledge alignment constraints: using known engineering knowledge (e.g., spring stiffness) (It mainly affects the system's natural frequency, while having a small direct impact on carbon emissions), so a priori weight matrix is ​​constructed. Then minimize and The differences between them (such as KL divergence or mean squared error) guide the model to learn logic that conforms to human cognition; and These are all hyperparameters used to balance the importance of various losses. Their values ​​are usually between 0.1 and 0.5, and the specific values ​​can be determined through grid search or based on historical training experience.

[0148] Using an interpretable multi-objective optimizer, the set of design variables is... During the multi-objective optimization search, through multiple rounds of iterative optimization, Pareto optimal solutions that are not dominated by other solutions are identified from all candidate engineering design schemes, forming the final Pareto optimal solution set.

[0149] The interpretable data includes: attention weight matrix Variable-objective relationship heatmap, list of key decision factors, and text describing trade-offs.

[0150] The attention weight matrix It is used to characterize the relationship between design variables and optimization objectives.

[0151] For the attention weight matrix Normalization and visualization mapping are performed to generate a variable-target correlation heatmap; the variable-target correlation heatmap is used to display the correlation between the design variables and the optimization target corresponding to the solution currently selected by the student user.

[0152] In this embodiment, the variable-target correlation heatmap is presented in matrix form, with rows representing different design variables and columns representing different optimization targets. The color intensity of the cells intuitively encodes the... The design variable for the first The heatmap shows the relative importance of each optimization objective. Simultaneously, it dynamically displays the strength of the influence of the design variables corresponding to the current focus solution on the optimization objective by linking in real-time with the interactive Pareto front view in the user interface.

[0153] The list of key decision factors is as follows: based on the variable-objective correlation heatmap, the influence of each design variable on the current Pareto optimal solution set or a specific solution is obtained, and the design variables are sorted in descending order of influence to obtain the sorting result.

[0154] In this embodiment, based on the variable-target correlation heatmap, the design variables that play a decisive role in the current Pareto optimal solution set or a specific solution are automatically identified and sorted and output according to their influence.

[0155] The trade-off description text is generated by using a predefined text template or a lightweight natural language generation model, based on the variable-objective correlation heatmap and the optimization objective value corresponding to the current solution, to describe the trade-off between design variables and optimization objectives in the current engineering design scheme.

[0156] In this embodiment, based on the variable-objective correlation heatmap data and the objective function value of the current solution (i.e., the optimization objective value), a descriptive text is automatically generated using a predefined text template or a lightweight natural language generation model. This text summarizes the core trade-offs of the current engineering design scheme, for example: "The current scheme significantly increases the damping coefficient..." Vibration transmissibility was achieved Extremely low levels, but this comes at the cost of hidden carbon emissions. "At the cost of a significant rise."

[0157] During the inference phase, when the interpretable optimizer has completed training and is ready for use, this module performs two tasks: first, it searches for and generates a Pareto optimal solution set; second, it simultaneously extracts and processes the attention weight matrix corresponding to the optimal solution set. To generate easily understandable variable-target correlation heatmaps, the matrix... The following processing is performed: First, for the currently focused solution or solution set, extract its corresponding attention weight submatrix; then, normalize this submatrix by row (design variables) or by column (optimization objective) (e.g., using...). Method normalization to (Intervals); Finally, the normalized values ​​are mapped to color shades to generate a heatmap. In the heatmap, rows represent design variables, columns represent optimization objectives, and cell colors visually encode the relative strength of the variable's influence on the objective.

[0158] The generated Pareto optimal solution set is provided to student users, who can then select to view interpretable data and standardized multi-dimensional index vectors corresponding to any solution in the Pareto optimal solution set. A solution is selected from the Pareto optimal solution set as the final engineering design scheme, and a structured demonstration report is generated.

[0159] Upon completion of a full optimization run or in-depth analysis of a specific solution, the interpretable optimizer synchronously outputs a structured interpretable data package. This data package typically includes: an attention weight matrix. The system includes a variable-objective correlation heatmap, a list of key decision factors, and descriptive text of trade-offs. The generated interpretable data package is deeply integrated with the interactive Pareto front visualization in the user interface. When student users explore the Pareto front (e.g., hovering the mouse or clicking on a solution point), the interface performs real-time actions: displaying the specific technical and sustainability indicators for that focus solution; dynamically drawing and highlighting the variable-objective correlation heatmap corresponding to that focus solution in another linked view area; and allowing student users to immediately identify which design variables dominate the performance of the current solution by observing the distribution of colors in the heatmap. For example, at a solution point aiming for ultimate vibration isolation performance, the heatmap might display the "damping coefficient..." "Vibration transmissibility" "There is a very strong positive correlation (darker); however, at the solution point in the pursuit of the lowest carbon emissions, the heatmap may show 'quality'." "and material type" for "hidden carbon" "It has a significant impact."

[0160] After guiding student users through exploration, the system guides them to select a final solution from the Pareto front. Specifically, students analyze and view the results, choosing a solution from the Pareto optimal set as the final engineering design. They are then required to complete a structured justification report. This report template mandates that students cite the weights of each current indicator and the corresponding rationale, combining this with observed Pareto front analysis results and findings from correlation heatmaps to demonstrate the rationality of their final choice.

[0161] During the process of student users making selections and viewings, a sequence of behavioral interactions recorded in chronological order is generated based on the student users' selections.

[0162] In this embodiment, the user's selection operation typically includes interactive actions such as the solution selected by the user, the variable-target correlation heatmap viewed by the user, and the solution marked by the user.

[0163] like Figure 3 As shown, through the aforementioned real-time visualized causal feedback mechanism, the abstract multi-objective trade-off logic is transformed into a concrete cognitive experience that learners can actively operate, observe in real time, and verify firsthand, thus realizing a closed loop of teaching experience from "black box operation" to "white box interaction." Through this interactive exploration, students can intuitively perceive how the dominant design variables change when moving from the "high-performance" region to the "low-emission" region on the Pareto front, thereby transforming the abstract "trade-off" relationship into a visible and perceptible causal chain.

[0164] The decision-making process tracking and dynamic cognitive evaluation module is used to calculate the multi-dimensional ability indicators of student users based on the behavioral interaction sequence and using a reinforcement learning-based evaluation model, and generate an ability evaluation report for student users.

[0165] In this embodiment, as Figure 4 As shown, the core function of the decision-making process tracking and dynamic cognitive evaluation module is to construct a digital profile of student users' decision-making behavior and to dynamically quantify their decision-making cognitive abilities. It incorporates an evaluation model based on the Markov Decision Process (MDP) framework, automatically calculating quantitative ability indicators for this training session (such as decision consistency index, exploration completeness, etc.) and generating a sustainable decision-making cognitive maturity report. This evaluation model models a single training session as an MDP, collecting behavioral sequences, calculating immediate rewards and cumulative discount rewards, and ultimately decomposing and outputting multiple quantitative ability indicators throughout the entire process. The aim is to map discrete user interaction sequences into quantitative indicators representing higher-order cognitive abilities.

[0166] The specific content of the decision-making process tracking and dynamic cognitive evaluation module is as follows:

[0167] Define a Markov Decision Process (MDP) as a quintuple. ;in For state space; For action space; The state transition probability; For the reward function; As a discount factor, and .

[0168] The state space Used for storage State vector at time step The state vector mentioned above This includes: the current standardized multi-dimensional indicator vector Weights of each indicator The coordinates of the Pareto solution that the current student user is viewing or selecting. And historical behavior summary information.

[0169] In this embodiment, the state vector Used to indicate student users The context in which the system interacts at any given moment. This historical behavior summary information typically includes: the variance of weight adjustments over a past period, the proportion of the area covered by explored solutions on the Pareto front, etc., for example, past... The variance of weight adjustments during the step operation, the proportion of the area occupied by the explored solution points on the Pareto front, and the trade-off characteristics of the current focus solution obtained from the optimization module are described.

[0170] The action space Used for storage Moment of action The actions described therein For student users The interactive behavior at any given moment is obtained from the sequence of behavioral interactions.

[0171] In this embodiment, the action Used to describe user-executable interactive commands, such as: Indicates the first The weights of each indicator are adjusted as follows: ; Indicates the selection of the first One solution. Indicates the first The weights of the coordinates of the solution points for each solution; Indicates viewing the One solution; Indicates the mark of the first Each solution, etc.

[0172] The state transition probability , used to describe the state Next action Afterwards, transition to the new state. The probability of is expressed as In deterministic systems, this transition probability is uniquely determined by the system interface logic and the backend computational model.

[0173] The reward function Used to monitor student users' status The following actions Perform real-time value assessment, expressed as Its reward value The design is based on cognitive science and teaching principles.

[0174] The reward logic defined in the reward function includes at least one of the following behaviors:

[0175] Value clarification rewards will be given to actions that increase the weight of sustainability indicators and provide reasonable justifications.

[0176] In-depth exploration rewards are given to those who continuously explore solution points in different regions of the Pareto front.

[0177] A causal investigation reward will be given to those who actively trigger the viewing of the interpretable data;

[0178] A logical consistency reward will be given to behaviors that align the final solution selection with the previously stated value preferences and exploration trajectory.

[0179] In this embodiment, the reward function The design directly reflects the guidance of expected teaching behaviors. The system is based on a series of pre-set cognitive behavior evaluation rules, and the specific reward logic and its setting basis are as follows:

[0180] Value Clarification Reward: A positive reward is given when a student user increases a previously low-weighted sustainability indicator (such as "implicit carbon emissions") and associates it with a reasonable justification (such as "to meet environmental regulations") selected from the "Structured Justification Knowledge Base" through the user interface. This knowledge base is pre-built based on engineering education syllabi, sustainable design specifications, and industry standards, and allows teacher users to expand it according to teaching needs.

[0181] In-depth exploration reward: When a student user continuously selects or views a Pareto solution, the Euclidean distance in the target space is greater than a preset threshold. This indicates that student users are actively exploring different areas of the trade-off frontier, and positive rewards are given accordingly. Among these, the threshold... The setting is dynamically calculated based on the current distribution range of the Pareto front in the target space. For example, it can be taken as 10% of the distance between the two farthest points on the front to encourage cross-regional exploration.

[0182] Causal Inquiry Reward: When student users view solutions with significant conflicting target values ​​(such as excellent performance but extremely high carbon emissions), they are given a positive reward for actively triggering the viewing of the corresponding variable-target correlation heatmap. This indicates that the student users are intentionally exploring the reasons behind the conflict.

[0183] Logical consistency reward: At the end of the task, the feature vector of the engineering design scheme finally selected by the student user is compared with its value preferences (weights and reasons) declared earlier in the task, as well as the tendency vector of the entire exploration trajectory. The cosine similarity between these vectors is calculated; if the average similarity is higher than a preset threshold, a reward is given. ,For example If the reward is positive, a reward will be given. The decision consistency index is calculated based on the proportion of this reward in the total reward.

[0184] The discount factor It is used to calculate the current value of future rewards and balance the impact of near-term and long-term rewards on the overall evaluation.

[0185] Based on the quintuple The behavioral interaction sequence is modeled as the trajectory of a Markov Decision Process (MDP), and the cumulative discount reward for student users is calculated.

[0186] In this embodiment, after completing a decision training session, the model uses the recorded user action sequence... Given the corresponding state sequence, calculate the cumulative discount reward earned by the student user, expressed as:

[0187]

[0188] in, This represents the cumulative discount reward, which serves as a quantitative total score for the student user's overall performance in this task. express The reward function value at time step; express Discount factor of time; Indicates duration.

[0189] The cumulative discount rewards are decomposed into sources, and multi-dimensional capability indicators of student users are calculated based on the decomposition results.

[0190] The multidimensional capability indicators include at least one of the following: decision consistency index, exploration completeness, sustainability value sensitivity, and causal analysis depth.

[0191] like Figure 5 As shown, to provide further refined diagnostics, cumulative discount rewards... Source decomposition is performed by breaking down the contributions of various rewards to calculate multi-dimensional capability indicators, including...

[0192] (1) Decision Consistency Index: Calculate the logical consistency reward in the cumulative discount reward The proportion of students' decisions in the data is used to determine the student user decision consistency index. A higher proportion indicates that the user's final decision is more consistent with their thought process.

[0193] (2) Completeness of exploration: The ratio of the cumulative value of the deep exploration reward obtained by the student user to the maximum value of the deep exploration reward that can be obtained theoretically under the task situation is used as the completeness of the student user's exploration. The higher the ratio, the more comprehensive the user's exploration of the solution space.

[0194] (3) Sustainability value sensitivity: Based on the frequency of value clarification rewards being triggered and the magnitude of each reward, the degree of attention and importance that users attach to sustainability factors is calculated, i.e., the sustainability value sensitivity of student users.

[0195] (4) Depth of causal analysis: A comprehensive score is given based on the number of times student users trigger the viewing of the heatmap, their behavior of comparing different solutions on the heatmap, and whether they make any related parameter adjustments after viewing. By calculating the cumulative value of the causal exploration reward, the student users' sustainability value sensitivity is obtained, which reflects the user's willingness and depth to actively explore the causal relationship between the design variables and the goal.

[0196] Based on a pre-set competency evaluation report template, a competency evaluation report for student users is generated according to the multi-dimensional competency indicators of student users.

[0197] In this embodiment, multiple ability indicators together constitute a digital profile of cognitive ability, which is output in the form of a quantitative evaluation report.

[0198] It should be noted that before the system is put into use, the evaluation model needs to be cross-validated across different user groups (such as students of different grades and majors) and different types of engineering optimization problems (such as structural design and thermal design) to ensure the stability and generalization ability of its evaluation results.

[0199] The intelligent controller for the teaching process is used to schedule and drive the activation, execution, and data exchange of the sustainability quantification and multi-source feature fusion module, the interpretable multi-objective collaborative optimization module, and the decision process tracking and dynamic cognitive evaluation module according to the teaching logic sequence preset by the teacher user, so as to execute the serialized teaching tasks, and to adjust the teaching logic sequence in real time according to the Pareto optimal solution set, interpretable data, and ability evaluation report.

[0200] The intelligent controller for the teaching process is electrically connected to the sustainability quantification and multi-source feature fusion module, the structured display and decision-making process simulation module, the interpretable multi-objective collaborative optimization module, and the decision-making process tracking and dynamic cognitive evaluation module.

[0201] In this embodiment, the teaching process intelligent controller serves as the central scheduling unit of the system, coordinating and driving the startup, execution, and data exchange of other modules based on a preset teaching logic sequence.

[0202] The engineering teaching system integrating interpretable multi-objective intelligent optimization further includes: a user interface; the user interface is electrically connected to the teaching process intelligent controller, and is used to display the output data of each module in the sustainability quantification and multi-source feature fusion module, the interpretable multi-objective collaborative optimization module, and the decision process tracking and dynamic cognitive evaluation module pushed by the teaching process intelligent controller to student users, and to receive interactive operation instructions from student users.

[0203] In this embodiment, the user interface is implemented through a graphical user interface (GUI) accessing the system via a client (e.g., a web browser or a dedicated desktop client). This interface layer is responsible for: Visual presentation: displaying design models, interactive Pareto front plots, variable-objective correlation heatmaps, weight sensitivity analysis plots (such as tornado plots), etc., in two-dimensional / three-dimensional graphical formats. Interaction capture: receiving and responding to all user operation commands, such as inputting design parameters, dragging and adjusting indicator weights, selecting or hovering over solutions on the Pareto front plot, triggering interpretability views, etc. Data tracking and reporting: recording key user interaction events (i.e., "tracking points") with fine granularity, including operation type, operation object, timestamp, context state, etc., and sending these behavioral event data to the backend service in real time or near real time.

[0204] In this embodiment, the specific scheduling process executed by the engineering teaching system is as follows:

[0205] The teaching process intelligent controller starts the teaching process, guides students to create or import an engineering design scheme (such as a 3D CAD model), and extracts the set of key design variables from it.

[0206] The controller calls the sustainability quantification and multi-source feature fusion module, which serves as the system's input processing unit. This module is responsible for converting the original design parameters from the external design environment or student users into unified technical performance indicators and sustainability indicators in parallel.

[0207] The structured display and decision-making process simulation module provides standardized multi-dimensional indicator vectors. Each indicator in the algorithm is assigned a weight with a structured rationale, and a weighted comprehensive utility value is calculated. Sensitivity analysis is used to identify the critical weight values ​​that lead to decision reversal, thereby assisting student users in determining optimization objectives and constraints.

[0208] The interpretable multi-objective collaborative optimization module receives the above optimization objectives and constraints, performs multi-objective optimization search, and simultaneously generates internal data to interpret the optimization logic.

[0209] The decision-making process tracking and dynamic cognitive evaluation module collects real-time interactive behavior data generated by student users through the user interface. Combined with interpretable data output by the optimization module, it performs quantitative modeling and ability assessment of student users' decision-making process.

[0210] The controller receives real-time quantitative indicators of user capabilities from the decision-making process tracking and dynamic cognitive evaluation modules. Based on this objective data and according to preset rules, it dynamically adjusts the difficulty and content of subsequent teaching tasks or provides personalized guidance. For example, if the exploration completeness is below a preset threshold, the controller will guide students to consciously explore areas not yet covered on the Pareto front when they perform similar tasks in the future, through interface highlighting, pop-up guidance boxes, or adjustments to task requirements. If the causal analysis depth is below a preset threshold, the controller may automatically display a correlation heatmap when a student selects a solution point, or force the student to complete a sub-task of "comparing the heatmaps of two different solutions points and explaining the differences." This assessment-feedback-intervention closed loop based on objective evaluation indicators achieves precise mapping and personalized guidance of learners' ability development status.

[0211] All modules provide learners with a unified, visual, and interactive learning experience through a user interface, forming a complete teaching and assessment loop. The system integrates with external engineering design software environments (such as CAD / CAE) to form a complete engineering teaching solution.

[0212] In this embodiment, the system adopts a teaching process engine service with embedded complete teaching state machine logic. It is responsible for calling other services in sequence (such as triggering feature parsing and starting optimization calculation), and dynamically determining subsequent process branches (such as whether to jump to the review stage or push specific prompts) based on the real-time capability indicators fed back by the decision process analysis service, so as to achieve adaptive teaching guidance.

[0213] Furthermore, various database technologies may be employed to adapt to different data characteristics for the storage and management of different types of data. For example: Time-series databases: used for efficient storage and querying of massive user interaction event streams generated in chronological order. Relational databases: used to store relational data such as user accounts, project metadata, structured reports, and final quantitative capability evaluation results. Graph databases: serving as the underlying storage for knowledge graph query services, supporting rapid traversal and querying of material-process-environment attribute association networks. Vector databases: used to store multi-dimensional feature vectors of historical successful design cases to support similarity-based case retrieval and recommendation functions.

[0214] Example 2:

[0215] This embodiment of an engineering education method integrating interpretable multi-objective optimization is implemented using the aforementioned engineering education system integrating interpretable multi-objective optimization, such as... Figure 6 As shown, the method includes the following steps:

[0216] Based on the teaching logic sequence preset by the teacher user, deploy and start the engineering education system that integrates interpretability and multi-objective optimization;

[0217] Obtain the initial engineering design scheme provided by student users and extract the set of design variables for that scheme;

[0218] Physical simulation analysis and logical analysis based on design features and process knowledge are performed on the set of design variables respectively to obtain a standardized multi-dimensional index vector of the initial engineering design scheme.

[0219] Assign weights with structured rationale to each indicator in the standardized multidimensional indicator vector, and calculate the weighted comprehensive utility value;

[0220] Based on the weighted comprehensive utility value, sensitivity analysis is used to identify the critical weight values ​​that lead to decision reversal, thereby assisting student users in determining optimization objectives and constraints.

[0221] Based on the optimization objective and constraints, an interpretable multi-objective optimizer is used to perform multi-objective optimization search on the set of design variables, generate a Pareto optimal solution set, and simultaneously output interpretable data.

[0222] Student users can select to view interpretability data corresponding to any solution in the Pareto optimal solution set, choose the final engineering design scheme from the Pareto optimal solution set, and generate a sequence of behavioral interactions based on the student user's selection.

[0223] Based on the behavioral interaction sequence, a reinforcement learning-based evaluation model is used to calculate the multi-dimensional ability indicators of student users and generate an ability evaluation report for student users.

[0224] Based on student users' competency evaluation reports, the pre-set teaching logic sequence for teachers is adjusted, and then the engineering education system is redeployed to form a closed-loop teaching process.

[0225] Example 3:

[0226] This embodiment takes the design of a single-degree-of-freedom stamping vibration damping system as an example, and provides an engineering education system that integrates interpretable multi-objective optimization to achieve multi-objective collaborative design of the single-degree-of-freedom stamping vibration damping system, such as... Figure 7As shown, the specific implementation method and application effect of the technical solution are further described in detail. This embodiment selects the classic teaching content in the course of mechanical system dynamics, namely the design of a single-degree-of-freedom stamping vibration reduction system, as the teaching project, and fully demonstrates the closed-loop process of the system of the present invention from design parameter input, dual-dimensional index quantification, interpretable multi-objective optimization exploration to quantitative evaluation and feedback of the decision-making process. The implementation process of this embodiment strictly follows the step sequence defined in the claims of the present invention, and the execution of each module is uniformly scheduled and coordinated by the teaching process intelligent controller. The specific process is as follows:

[0227] The initial teaching task requires students to design a vibration isolation system for a precision stamping machine. This physical system is modeled as a single-degree-of-freedom dynamic model, with the core design variable being the mass of the mass block. Its unit is kilogram (kg) Spring stiffness coefficient Its unit is Newton per meter (N / m). ), and the damping coefficient of the damper Its unit is Newton-second per meter (N / m). Students input or confirm a set of initial design parameters through the system interface, such as setting the mass. stiffness coefficient Damping coefficient .

[0228] Upon receiving the original design variables, the system immediately initiates the sustainability quantification and multi-source feature fusion module, performing the following two types of calculations in parallel.

[0229] The first type of calculation involves the automated simulation and quantification of technical performance indicators: Utilizing the module's embedded physical simulation engine, frequency domain analysis is automatically performed based on the motion differential equations of a single-degree-of-freedom system. This process eliminates the need for manual mesh generation, load setting, and other complex preprocessing; the system autonomously calls the solver based on the model type. The calculation outputs three key technical performance indicators:

[0230] The first indicator is vibration transmissibility. The curve of the ratio of the mass block response amplitude to the basic excitation amplitude under basic excitation as a function of frequency is calculated, and its maximum value is extracted as the index value. This index value needs to be minimized to obtain the best vibration isolation effect.

[0231] The second indicator is the system's natural frequency. From the formula Direct calculation shows that this frequency needs to be avoided in the operating frequency band of the equipment to prevent resonance.

[0232] The third metric is the maximum vibration displacement response under a given typical load spectrum. This value must meet preset safety constraints, such as being less than 1 mm. The system automatically compares the calculated index values ​​with the preset constraints and clearly marks the constraint satisfaction status on the user interface.

[0233] The second type of calculation is the automated calculation of sustainability indicators: this process relies on the collaborative work of two core components within the module: the design feature parsing rule base and the materials-process knowledge graph.

[0234] First, the design feature parsing rule base is designed for the input design variables. , , Logical parsing is performed. This rule base predefines mapping rules from engineering parameters to manufacturing characteristics. Specifically, the rule base will... The analysis suggests that a specific volume of metal material needs to be processed and shaped to dampen the material. The analysis suggests that a certain amount of viscoelastic damping material is needed for molding and assembly to increase stiffness. The analysis indicates that a spring component is required, and its size and manufacturing complexity are similar to... Value-related. The rule base transforms these parsed results into a series of structured descriptions of manufacturing activities.

[0235] Subsequently, these manufacturing activity descriptions are sent to a materials-process knowledge graph query service. This knowledge graph stores the relationships between material entities, processing technology entities, and their environmental impact factors in a graph structure. The system performs a graph query operation; for example, based on the material being Q235 steel and the process being turning, it queries to obtain the implicit carbon emission factor per unit mass of that material. Its unit is kilogram carbon dioxide equivalent per kilogram, and the processing energy consumption factor per unit volume of the process. The unit is megajoules per cubic centimeter. For natural rubber materials and vulcanization molding processes, relevant environmental data can be obtained by querying the database.

[0236] Combined with material usage derived from design variables Processing volume Geometric parameters are used to automatically calculate three sustainability indicators.

[0237] The first indicator is material utilization efficiency. This indicator is defined as the current design quality. Achieving the same natural frequency Required theoretical minimum mass The ratio, i.e. It is used to measure the rationality of material use.

[0238] The second indicator is implicit carbon emissions. The calculation is derived by summing up the carbon emissions from all component material production and processing stages. The calculation formula is as follows:

[0239]

[0240] in, Indicates component index; Indicates the first The quality of each component; Indicates the first Carbon emission factor of materials in each component; Indicates the index of processing activities; Indicates the first Material usage for each processing activity; Indicates the first The carbon emission factor of materials in each processing activity; and All are integers greater than 0.

[0241] The third indicator is the manufacturing complexity score. Based on the information on process complexity and number of assembly steps stored in the knowledge graph, it is calculated comprehensively through a predefined scoring model.

[0242] The user interface clearly and side-by-side displays all the above quantitative results, including technical performance indicator vectors. and sustainability indicator vector .in Include , , and its constraint states, Include , as well as The original set of design variables Automatically convert into a unified, computable, standardized multi-dimensional indicator vector. This provides a consistent and synchronized data foundation for subsequent processes.

[0243] The system guides students to assign relative importance weights to the above six indicators. ,in The value range is 1 to 6. To ensure traceability and evaluability of decision-making, the system mandates that users associate a reason with each weighted action. Users must select or input compliant custom reasons from the system's structured reason knowledge base, and cannot arbitrarily assign weights and skip ahead. This knowledge base includes pre-designed, instructional options, such as for vibration transmissibility. When assigning high weight, possible reasons include ensuring stamping process accuracy or improving equipment reliability; and implicit carbon emissions. When assigning high weights, possible reasons include responding to carbon neutrality strategies or reducing the environmental costs of products. The system persistently records the resulting triplet data of indicators, weights, and reasons, and adds a timestamp to each triplet, thus forming a serialized record of the evolution of user value preferences.

[0244] At the same time, the system is based on the currently set weight distribution set. The system initiates multi-criteria decision simulation analysis in real time. It standardizes the technical performance and sustainability indicators of the current design scheme and calculates its weighted comprehensive utility value. The calculation formula is: ,in For the first The system displays normalized values ​​for each indicator. In the user interface, it dynamically displays a preliminary ranking of the design scheme points, including the current design scheme point and several neighboring variation points. More importantly, the system integrates sensitivity analysis functionality. This function uses numerical methods to fine-tune the weight of each indicator. Observe the overall utility value The system can also analyze the degree of change in the ranking of alternatives, presenting the results in visual formats such as tornado diagrams. Based on the analysis results, the system will generate insightful prompts, such as indicating the current damping coefficient. The affected indicators are most sensitive to the overall score because the user is also the vibration transmissibility. and implied carbon emissions It was given a higher weight, while This has a significant impact on both objectives. This immediate feedback forces students to think deeply about the intrinsic connections and conflicts between different indicators, thereby driving their decision-making logic to deepen from isolated indicator assignment to a systematic and comprehensive trade-off.

[0245] Students need to formally define a multi-objective optimization problem. For example, minimizing the vibration transmissibility. Minimize hidden carbon emissions Set as a pair of conflicting optimization objectives, while also considering the system's inherent frequency. Specific ranges and maximum vibration displacement responses need to be avoided. The value less than the limit is used as a constraint. Design variables. , , Search within its feasible domain.

[0246] After students confirm the problem definition, the system invokes the interpretable multi-objective collaborative optimization module. The core of this module is an optimizer based on the Transformer neural network architecture and specially trained. Its innovation lies in the loss function used. In addition to the performance loss term that aims to approximate the true Pareto front of the solution set, the middle part is also included. and the diversity loss term to ensure uniform distribution of solution sets. In addition, an interpretability regularization term was forcibly added. This term relates to the attention weight matrix generated by the optimizer's internal self-attention mechanism. Imposing structured constraints, such as encouraging sparsity through L1 norm penalties, or by minimizing the prior engineering knowledge matrix. The differences prompt it to align with physical logic, thereby making the internal representation of the model easier to translate into human-understandable interpretations.

[0247] During optimizer runtime, it encodes and searches the sequence of design variables, dynamically generating a Pareto optimal solution set that approximates the real-world situation. The user interface then presents two core, real-time interactive visualization areas. Dynamically plotting the vibration transmissivity With hidden carbon emissions The Pareto front is constructed in a two-dimensional objective space, where each point on the front represents an optimal trade-off under given constraints. Region Initially, it is an interpretable view area.

[0248] When the user hovers the mouse cursor over the area At any point on the Pareto front, the region The page refreshes instantly, displaying a variable-target correlation heatmap specifically generated for that point. This heatmap is directly derived from the normalized attention weight matrix calculated internally by the optimizer for this candidate solution. The rows of the heatmap matrix represent design variables. , , , column represents optimization objective , The color intensity of each cell in the matrix intuitively encodes the relative importance of the design variable represented by the corresponding row to the optimization objective represented by the corresponding column.

[0249] Specifically, when the cursor hovers over the leftmost point of the Pareto front, this point corresponds to the vibration transmissibility. The lowest possible solution. At this point in the region... The heatmap is likely to show that, for the target The column it belongs to, the damping coefficient The darkest color in the current row indicates that the damping coefficient should be increased in the current solution. This is the dominant factor in achieving optimal vibration isolation. However, at the same time, the heat map is also important for the target... In the column where, The cell in the same row may also be displayed in a darker color, indicating that this strategy also results in hidden carbon emissions. The system may simultaneously provide an automatically generated text description summarizing the key trade-offs embodied at that point.

[0250] As the user moves the cursor to the right along the Pareto front, they can explore hidden carbon emissions. In lower regions, the region The heatmap will continuously evolve. At the far right of the Pareto front, the heatmap pattern may change, displaying quality. To reduce Their contributions became the most prominent, but at the same time right The negative impacts have also deepened significantly. Through this proactive and continuous interactive exploration, students can observe in real time and visually how the dominant design variables and their impact patterns on different objectives dynamically change as they move across the performance-environment trade-off spectrum. This interactive mechanism, based on model-native interpretable data, transforms traditional multi-objective optimization from a black-box tool that only outputs results into a white-box exploration tool capable of revealing the intrinsic causal relationships of complex systems.

[0251] After interactive exploration, students are required to select a final solution on the Pareto front and submit a structured justification report. This report template mandates that students cite the weighting assignments and rationale for the current records, and, in conjunction with the analysis of the Pareto front and the observed variable-objective relationship heatmap, demonstrate how their final choice reflects their stated value preferences and reconciles conflicts between different objectives. For example, students need to explain why they chose a specific combination of damping coefficient and mass, rather than an extreme choice that pursues a single performance or environmental objective.

[0252] Meanwhile, the system's decision-making process tracking and dynamic cognitive evaluation module runs entirely in the background. Starting with the allocation of indicator weights, this module captures and structures all fine-grained user interaction events in real time, forming a complete behavioral sequence log. The module's built-in evaluation model is constructed based on the Markov Decision Process (MDP) framework. Its core is a predefined reward function. Based on the principles of cognitive science, this function assigns immediate rewards to specific actions that represent good learning behavior. The reward rules include, but are not limited to, rewarding users for improving a sustainability metric that was previously given a lower weight (such as...). When a user provides a valid reason for their actions, the system awards a value clarification reward; when a user continuously explores the target space and the geometric distance exceeds a preset threshold... When a Pareto solution is found, the system awards a reward for in-depth exploration; when the user actively triggers the viewing of targets with significant conflicts (such as T and...), the reward is given. When a detailed heatmap of the solution points (all at extreme values) is generated, the system awards a reward for causal exploration; when the task ends, the characteristics of the solution finally selected by the user are highly consistent with their previously stated value preferences and the characteristics of the entire exploration trajectory, i.e., the similarity is higher than a preset threshold. When this happens, the system awards a reward for logical consistency.

[0253] After a single training session, the model analyzes the user's complete behavioral sequence and calculates the cumulative discount reward received. The calculation formula is: ,in A discount factor between 0 and 1, used to weigh the importance of near-term and long-term rewards. Scalar This score itself is a quantifiable score reflecting the user's overall performance in this task. The system further analyzes the various rewards within the total reward... The contribution ratio is broken down into multiple quantitative ability indicators, thereby automatically generating an objective digital profile report of cognitive ability. The report primarily includes a decision consistency index, which is based on the logical consistency reward in the total reward... The percentage of the reward is calculated as follows: 0.88 in this embodiment; exploration completeness is calculated as the ratio of the accumulated value of the deep exploration reward actually obtained by the user to the maximum value of the deep exploration reward that can be theoretically obtained under the current task situation, which is 0.92 in this embodiment; causal analysis depth is calculated based on the accumulated value of causal exploration reward, which is 0.75 in this embodiment; and sustainability value sensitivity is calculated as the combination of the frequency of value clarification reward being triggered and the average reward magnitude, which is 0.72 in this embodiment.

[0254] This fully automated quantitative evaluation report can be provided to both teachers and students, realizing a paradigm shift in teaching evaluation from relying on teachers' subjective experience to being based on objective process behavior data. More importantly, the intelligent controller for the teaching process receives this evaluation result. The controller has pre-set rules for teaching intervention, which can automatically trigger personalized learning interventions based on the values ​​of various quantitative indicators. For example, if a student's performance on the causal analysis depth indicator is insufficient, the system can automatically add a mandatory requirement to subsequent similar design challenge tasks: a detailed comparison of the variable-objective correlation heatmaps between the high-damping and lightweight solutions, and a written explanation of the underlying physical mechanisms. This constitutes a complete "learning-assessment-feedback-intervention" closed loop, enabling the system to provide precise teaching guidance based on objective data, thereby promoting the systematic development of learners' higher-order engineering decision-making abilities.

[0255] This embodiment, through the specific and operable engineering teaching case of designing a single-degree-of-freedom stamping vibration reduction system, fully and meticulously elucidates the collaborative workflow of each module of the system and the specific implementation methods of each step in the claims. It clearly demonstrates the technical feasibility of achieving data consistency between technical and environmental indicators through the sustainability quantification and multi-source feature fusion module, achieving white-box and exploratory optimization of the decision-making process through the interpretable multi-objective collaborative optimization module, and achieving quantitative assessment of decision-making ability and personalized teaching intervention through the decision-making process tracking and dynamic cognitive evaluation module. This embodiment provides sufficient and implementable support for the technical solutions defined in this invention.

Claims

1. An engineering education system integrating interpretable multi-objective intelligent optimization, characterized in that, The system includes: The sustainability quantification and multi-source feature fusion module is used to obtain the initial engineering design scheme provided by student users and extract the set of design variables for that scheme. ; respectively for the design variable set Physical simulation analysis and logical analysis based on design features and process knowledge are performed to obtain a standardized multi-dimensional index vector of the initial engineering design scheme. ; The structured display and decision-making process simulation module is used to standardize multi-dimensional indicator vectors. Each indicator in the algorithm is assigned a weight with a structured rationale, and a weighted comprehensive utility value is calculated. Sensitivity analysis is used to identify the critical weight values ​​that lead to decision reversal, thereby assisting student users in determining optimization objectives and constraints. An interpretable multi-objective collaborative optimization module is used to optimize the set of design variables using an interpretable multi-objective optimizer based on the optimization objective and constraints. Perform multi-objective optimization search to generate a Pareto optimal solution set and simultaneously output interpretable data; at the same time, in response to the student user's selection operation on the Pareto optimal solution set, generate a behavioral interaction sequence. The decision-making process tracking and dynamic cognitive evaluation module is used to calculate the multi-dimensional ability indicators of student users based on the behavioral interaction sequence and using a reinforcement learning-based evaluation model, and generate an ability evaluation report for student users. The teaching process intelligent controller is used to schedule and drive the startup, execution and data exchange of the sustainability quantification and multi-source feature fusion module, the interpretable multi-objective collaborative optimization module, and the decision process tracking and dynamic cognitive evaluation module according to the teaching logic sequence preset by the teacher user, so as to execute the serialized teaching tasks, and adjust the teaching logic sequence in real time according to the Pareto optimal solution set, interpretable data and ability evaluation report. The intelligent controller for the teaching process is electrically connected to the sustainability quantification and multi-source feature fusion module, the structured display and decision-making process simulation module, the interpretable multi-objective collaborative optimization module, and the decision-making process tracking and dynamic cognitive evaluation module.

2. The engineering education system integrating interpretable multi-objective intelligent optimization according to claim 1, characterized in that, The respective design variable sets Physical simulation analysis and logical analysis based on design features and process knowledge are performed to obtain a standardized multi-dimensional index vector of the initial engineering design scheme. The specific content is as follows: By calling the physics simulation engine, the set of design variables is... Physical simulation analysis is performed to obtain the technical performance index vector of the initial engineering design scheme. ; Based on the design feature parsing rule base and material-process knowledge graph, the design variable set... Logical analysis based on design features and process knowledge is performed to obtain the sustainability index vector of the initial engineering design scheme. ; according to and The standardized multi-dimensional index vector of the initial engineering design scheme is obtained. .

3. The engineering education system integrating interpretable multi-objective intelligent optimization according to claim 2, characterized in that, The design variable set is based on a design feature parsing rule base and a material-process knowledge graph. Logical analysis based on design features and process knowledge is performed to obtain the sustainability index vector of the initial engineering design scheme. The specific content is as follows: Utilize predefined feature recognition rules to construct a design feature parsing rule base; Using the aforementioned design feature parsing rule base, the design variable set is... Perform process semantic recognition to obtain the manufacturing features and assembly relationships of the initial engineering design scheme, and map the obtained manufacturing features and assembly relationships into a list of manufacturing activities of the initial engineering design scheme; Construct a material-process knowledge graph to store material entities, processing technology entities, and environmental load factors associated with each entity; Based on the list of manufacturing activities, the environmental load factor corresponding to each manufacturing activity is obtained by querying the material-process knowledge graph; Based on the set of design variables Calculate the geometric parameters of the initial engineering design scheme, and establish a mapping model between the geometric parameters and the environmental load factors based on the preset environmental load factor mapping relationship library; Based on the mapping model, the sustainability index of the initial engineering design scheme is calculated. .

4. The engineering education system integrating interpretable multi-objective intelligent optimization according to claim 3, characterized in that, The specific content of the structured display and decision-making process simulation module is as follows: For standardized multidimensional indicator vectors For any indicator, assign a weight to that indicator; Based on the pre-set structured reason knowledge base, a structured reason is associated with the weighting behavior of the indicator, generating triple data containing the indicator, weight, and structured reason; Based on the triplet data, calculate the weighted comprehensive utility value of the current engineering design scheme; Based on the weighted comprehensive utility value, sensitivity analysis is used to identify the critical weight value that leads to a decision reversal in the current engineering design scheme. Student users, based on the aforementioned critical weight values, select from a standardized multi-dimensional indicator vector. Several indicators are selected as optimization objectives, and all unselected indicators are used as constraints.

5. The engineering education system integrating interpretable multi-objective intelligent optimization according to claim 4, characterized in that, The specific content of the explainable multi-objective collaborative optimization module is as follows: Under the condition that the constraints are met, and using the optimization objective as the standard, an interpretable multi-objective optimizer is employed to optimize the set of design variables. Perform multi-objective optimization search to generate a Pareto optimal solution set, and simultaneously output the interpretability data corresponding to each solution in the Pareto optimal solution set; The interpretable multi-objective optimizer mentioned above is implemented using an interpretable optimizer based on the Transformer neural network architecture; The generated Pareto optimal solution set is provided to student users, who can then select to view interpretable data and standardized multi-dimensional index vectors corresponding to any solution in the Pareto optimal solution set. A solution is selected from the Pareto optimal solution set as the final engineering design scheme, and a structured demonstration report is generated. During the process of student users making selections and viewings, a sequence of behavioral interactions recorded in chronological order is generated based on the student users' selections.

6. The engineering education system integrating interpretable multi-objective intelligent optimization according to claim 5, characterized in that, The method for constructing the interpretable optimizer based on the Transformer neural network architecture is as follows: Construct a Transformer model, including a decoder and an encoder; wherein the input to the encoder includes: constraints, optimization objective, and a set of design variables. The encoder is used to map the encoder input to a high-dimensional semantic feature space using a multi-head attention mechanism to obtain the encoder output, and to extract the attention weight matrix generated by the multi-head attention mechanism. The decoder is used to decode the output of the encoder and generate a sequence of candidate engineering design schemes. A multi-task loss function with an interpretability regularization term is used to train the Transformer model, resulting in a well-trained Transformer model, which is then used as an interpretable optimizer based on the Transformer neural network architecture. The multi-task loss function is a weighted sum of a performance loss term, a diversity loss term, and an interpretability regularization term. The performance loss term drives the Pareto solution set generated by the interpretable optimizer to converge to the true Pareto front. The diversity loss term maximizes the coverage of the Pareto solution set generated by the interpretable optimizer in the target space. The interpretability regularization term applies structured constraints to the attention weight matrix, including at least one of sparsity constraints, block structure constraints, or alignment constraints with the engineering prior knowledge matrix.

7. The engineering education system integrating interpretable multi-objective intelligent optimization according to claim 6, characterized in that, The interpretable data includes: attention weight matrix Variable-objective relationship heatmap, list of key decision factors, and text describing trade-offs; The attention weight matrix , used to characterize the relationship between design variables and optimization objectives; For the attention weight matrix Normalization and visualization mapping are performed to generate a variable-target correlation heatmap; the variable-target correlation heatmap is used to display the correlation between the design variables and the optimization target corresponding to the solution currently selected by the student user; The list of key decision factors is as follows: Based on the variable-objective correlation heatmap, the influence of each design variable on the current Pareto optimal solution set or a specific solution is obtained, and the design variables are sorted in descending order of influence to obtain the sorting result; The trade-off relationship description text is generated by using a predefined text template or a lightweight natural language generation model, based on the variable-objective correlation heatmap and the objective function value of the current solution, to describe the trade-off relationship between design variables and optimization objectives in the current engineering design scheme.

8. The engineering education system integrating interpretable multi-objective intelligent optimization according to claim 7, characterized in that, The specific content of the decision-making process tracking and dynamic cognitive evaluation module is as follows: Define a Markov Decision Process (MDP) as a quintuple. ;in For state space; For action space; The state transition probability; For the reward function; As a discount factor, and ; The state space Used for storage State vector at time step The state vector mentioned above This includes: the current standardized multi-dimensional indicator vector Weights of each indicator The coordinates of the Pareto solution that the current student user is viewing or selecting. and historical behavior summary information; The action space Used for storage Moment of action The actions described therein For student users The interactive behavior at any given moment is obtained from the sequence of behavioral interactions; The state transition probability , used to describe the state Next action Afterwards, transition to the new state. The probability of; The reward function Used to monitor student users' status The following actions Conduct immediate value assessment; The reward logic defined in the reward function includes at least one of the following behaviors: Value clarification rewards are used to positively reward actions that increase the weight of sustainability indicators and provide reasonable justifications. The depth exploration reward is used to give a positive reward when a student user continuously selects or views a Pareto solution and the Euclidean distance in the target space is greater than a preset threshold. The causal inquiry reward is used to give positive rewards to student users who actively trigger the viewing of the interpretable data; Logical consistency reward is used to give positive rewards to behaviors that are consistent with the value preferences and exploration trajectory stated in the early stage when choosing the final engineering design scheme. Based on the quintuple The behavioral interaction sequence is modeled as the trajectory of a Markov Decision Process (MDP), and the cumulative discount reward for student users is calculated. The cumulative discount rewards are decomposed into sources, and multi-dimensional ability indicators of student users are calculated based on the decomposition results; wherein the multi-dimensional ability indicators include at least one of the following: decision consistency index, exploration completeness, sustainability value sensitivity, and causal analysis depth. Based on a pre-set competency evaluation report template, a competency evaluation report for student users is generated according to the multi-dimensional competency indicators of student users.

9. The engineering education system integrating interpretable multi-objective intelligent optimization according to claim 1, characterized in that, The engineering teaching system integrating interpretable multi-objective intelligent optimization further includes: a user interface; the user interface is electrically connected to the teaching process intelligent controller, and is used to display the output data of each module in the sustainability quantification and multi-source feature fusion module, the interpretable multi-objective collaborative optimization module, and the decision process tracking and dynamic cognitive evaluation module pushed by the teaching process intelligent controller to student users, and to receive interactive operation instructions from student users.

10. An engineering education method integrating interpretable multi-objective intelligent optimization, implemented using an engineering education system integrating interpretable multi-objective intelligent optimization as described in any one of claims 1-9, characterized in that, This method The process includes the following: Based on the teaching logic sequence preset by the teacher user, deploy and start the engineering education system that integrates interpretability and multi-objective optimization; Obtain the initial engineering design scheme provided by student users and extract the set of design variables for that scheme; Physical simulation analysis and logical analysis based on design features and process knowledge are performed on the set of design variables respectively to obtain a standardized multi-dimensional index vector of the initial engineering design scheme. Assign weights with structured rationale to each indicator in the standardized multidimensional indicator vector, and calculate the weighted comprehensive utility value; Based on the weighted comprehensive utility value, sensitivity analysis is used to identify the critical weight values ​​that lead to decision reversal, thereby assisting student users in determining optimization objectives and constraints. Based on the optimization objective and constraints, an interpretable multi-objective optimizer is used to perform multi-objective optimization search on the set of design variables, generate a Pareto optimal solution set, and simultaneously output interpretable data. Student users can select to view interpretability data corresponding to any solution in the Pareto optimal solution set, choose the final engineering design scheme from the Pareto optimal solution set, and generate a sequence of behavioral interactions based on the student user's selection. Based on the behavioral interaction sequence, a reinforcement learning-based evaluation model is used to calculate the multi-dimensional ability indicators of student users and generate an ability evaluation report for student users. Based on student users' competency evaluation reports, the pre-set teaching logic sequence for teachers is adjusted, and then the engineering education system is redeployed to form a closed-loop teaching process.

Images