A design thinking visualisation system and method for nodal behavior feedback calibration

Abstract

The application discloses a design thinking visualization system based on node behavior comparison and feedback calibration, which comprises a visualization editor interface, a node library and a simulation running module; the visualization editor interface is used for providing an intuitive node operation and network building environment for a user, supporting visualization modeling, configuration and debugging of design thinking; the node library comprises a plurality of atomic nodes integrated with AI agents, each atomic node adopts a three-layer architecture of input-processing-output, and at least comprises a context node C, a problem node P, a decision node D, an expected behavior node Be1, a scheme node S, an actual behavior node Ba2, a comparison node Compare and a value node V; each atomic node serves as an independent AI agent and is used for semantic understanding, data processing and logical operation of a corresponding link; and the simulation running module is used for driving node network operation, completing core data stream transmission and feedback calibration closed loop.

Description

Technical Field

[0001] This invention relates to the field of intelligent assisted design technology, specifically to a design thinking visualization system and method for node-based behavior feedback calibration. Background Technology

[0002] With the rapid development of digital design, design thinking has become a core methodology in product development. Its core lies in achieving efficient solution output through multi-stage iteration and a focus on user needs. However, current mainstream traditional design tools (such as Figma, Sketch, and Pixso) are still limited to basic functions like assisting with drawing and building static or dynamic prototypes. They cannot predict user behavior during the design implementation process, accurately compare design expectations with actual feedback, or automate calibration. This deficiency directly leads to a significant amount of time being spent on repeated verification during the design process, resulting in low design efficiency, high iteration costs, and a design value that heavily relies on the designer's personal experience, making standardization and large-scale reuse difficult.

[0003] While some node-based editors exist in the existing technology (such as UE for game design simulation and Touch Designer for interactive media), which have certain node-based logic building and simulation capabilities, these tools lack deep integration with AI agents for design thinking. They cannot effectively handle the difference between the expected and actual user behavior paths, and they do not form a closed-loop feedback mechanism of prediction-execution-comparison-calibration, thus failing to meet the core needs of non-linear iteration in design thinking.

[0004] Design thinking is not a linear process, but rather a non-linear, multi-dimensional process encompassing context awareness, problem definition, behavioral deduction, solution visualization, feedback verification, and value assessment. Existing technologies fail to explicitly transform implicit design experience into an operational and quantifiable networked structure, hindering the deep synergy between human designer intuition and AI computing power. Consequently, they struggle to address the core pain points of design logic being black-boxed, lack of feedback alignment, and low iteration efficiency. Summary of the Invention

[0005] This invention aims to address the technical problem of how to make implicit design logic explicit through artificial intelligence technology and how to use algorithms to automatically compare design expectations with actual feedback, thereby constructing an adaptive and quantifiable design iteration closed loop. The network architecture of this invention is a node-based design thinking system. Its core feature is that each atomic node in the system is an independent AI agent. The agent follows a unified three-layer input-processing-output architecture specification, uses strongly typed ports to achieve multimodal data transmission and collaborative computation, and ultimately achieves closed-loop feedback through comparison nodes, overcoming the aforementioned shortcomings of existing design tools.

[0006] The technical solution of the present invention is as follows: A design thinking visualization system for node-based behavioral feedback calibration includes a visualization editor interface, a node library, and a simulation operation module. The visualization editor interface provides users with an intuitive environment for node operation and network building, supporting the visualization modeling, configuration, and debugging of design thinking. The node library contains atomic nodes with various integrated AI agents. Each atomic node adopts a three-layer architecture of input-processing-output. Each atomic node includes at least a context node C, a problem node P, a decision node D, an expected behavior node Be1, a solution node S, an actual behavior node Ba2, a comparison node Compare, and a value node V. Each atomic node acts as an independent AI agent, used for semantic understanding, data processing, and logical operations in its corresponding stage. The simulation operation module drives the node network to complete the core data flow transmission and feedback calibration closed loop.

[0007] Furthermore, each atomic node supports role attribute declaration, which includes the human role and the machine role, with different role attributes corresponding to different algorithm preferences of the node processing layer.

[0008] Furthermore, the context node C is used to extract background semantics and quantize them. Its input layer supports the access of multimodal data such as text, images, and numerical data. The processing layer integrates a cross-modal coding agent, uses a feature fusion algorithm to extract semantic vectors and perform spatiotemporal background clustering, and the output layer outputs a structured context tensor. The problem node P is used to identify demand gaps and define design goals. The input layer receives the scenario tensor output by the scenario node C and preset industry benchmark data. The processing layer performs logical breakpoint scanning through the knowledge graph. The output layer outputs a structured demand tensor. The decision node D is used to anchor intent and set decision boundaries after demand identification. The input layer receives a multimodal input interface, including the context tensor of the context node C, the demand tensor of the problem node P, and the subjective intent input of the current role. The processing layer integrates a decision weight AI agent, which uses the role weighting operator Wr to set hard constraints and ethical boundaries of the design goals, and completes the weight fusion of multi-dimensional inputs and generation of decision benchmarks. The output layer outputs a structured decision constraint stream or a hierarchical guided decision data stream, which serves as the core criterion for subsequent solution generation.

[0009] Furthermore, the expected behavior node Be1 is used to deduce the user's operation path. The input layer receives the prototype parameters of the scheme node S and the target user profile. The processing layer uses the behavior dynamics simulation algorithm and the Monte Carlo algorithm to simulate the distribution of the user's operation path under different resistances. The output layer outputs the expected behavior feature vector sequence S_1. The actual behavior node Ba2 is used to collect and process actual interaction data. The input layer receives real-time embedded data, simulator operation log stream or interaction status capture data. The processing layer cleans, slices and transforms unstructured data into standard behavior vectors. The output layer outputs the actual behavior vector sequence S_2.

[0010] Furthermore, the comparison node Compare is used to compare the expected behavior with the actual behavior, and specifically performs the following steps: 1) Path feature extraction: the expected behavior feature vector sequence S_1 and the actual behavior vector sequence S_2 are encoded respectively; 2) Dynamic Time Warping (DTW) alignment: the shortest alignment path and minimum cumulative distance of the two variable-length sequences are calculated using the DTW algorithm to adapt to the differences in user operation rhythm; 3) Role alignment deviation calculation: the role weighting operator Wᵣ is introduced to calculate the semantic distance between the actual delivery trajectory and the design goal.

[0011] Furthermore, the solution node S is used to generate interaction strategies and digital prototypes. The input layer receives the pain point feature vector output by the problem node P, the structured decision constraint flow output by the decision node D, and the component library metadata. The processing layer uses a diffusion model or Transformer architecture to generate the topology or code framework. The output layer outputs digital prototype assets. The value node V is used to evaluate the effectiveness of the scheme and trigger calibration. The input layer receives the similarity score and ROI evaluation formula of the comparison node Compare. The processing layer performs the value function derivation. The output layer outputs a calibration signal to reset the parameters of the scheme node S or the expected behavior node Be1.

[0012] Furthermore, the node network forms two closed-loop links: one is the main logic flow, which is: Context node C → Problem node P → Decision node D → Expected behavior node Be1 → Solution node S → Actual behavior node Ba2; the other is the adaptive feedback flow, which is: Actual behavior node Ba2 → Compare node → Value node V → Solution node S / Expected behavior node Be1 / Decision node D; when the deviation value D calculated by the Compare node is greater than the preset threshold, the parameters are automatically reset through the adaptive feedback flow to complete the automatic fine-tuning of the solution.

[0013] Furthermore, the node library also includes container nodes, which are used to encapsulate the child node network, forming a hierarchical structure of composite nodes, aggregating the inputs / outputs of multiple child nodes, and supporting recursive processing; the container nodes provide external input ports and output ports to the outside world, and internally enable hierarchical viewing to view the complete micro-loop of child nodes and data flow path.

[0014] A design thinking visualization method based on node-based behavior comparison and feedback calibration, applied to the system described in any one of claims 1-8, includes the following steps: Step 1) Node Setup: Select atomic nodes from the node library through the visual editor interface, configure the role attributes and parameters of each node, connect each atomic node through a strongly typed port (input / output data format is consistent, data is transmitted without loss, and type-safe data flow) to form a node network, and define the link rules of the main logic flow and the adaptive feedback flow. Step 2) Data Input: Input multimodal context data into context node C, import preset industry benchmark data into problem node P, input the subjective intent data of the current role into decision node D, input the target user profile into expected behavior node Be1, and complete the initialization of all node data; Step 3) Simulation Operation: Drive the node network to run through the simulation operation module. The main logic flow completes the entire process of situation awareness, problem definition, decision constraint setting, expected behavior inference, solution generation and actual behavior collection, generating the expected behavior feature vector sequence S_1 and the actual behavior feature vector sequence S_2. Step 4) Behavior comparison: Compare the execution path feature extraction, DTW sequence alignment and role alignment deviation calculation of the comparison node, and generate the deviation value D and the deviation point marker; Step 5) Value assessment and closed-loop calibration: Value node V completes the scheme effectiveness assessment based on the deviation value D and the ROI assessment formula. If the deviation value D is greater than the preset threshold, a calibration signal is triggered. The parameters of the scheme node S, the expected behavior node Be1, or the decision node D are adjusted back through the adaptive feedback flow. Steps 3) to 5) are repeated until the deviation value D is less than or equal to the preset threshold, and the design iteration is completed.

[0015] Furthermore, in step 4), the behavior comparison process also includes generating a visual deviation heatmap and conflict warning lines, transforming the abstract algorithm calculation results into intuitive and perceptible visual information.

[0016] Compared with the prior art, the present invention has the following beneficial effects: 1) Achieve automated and precise iteration: By comparing nodes using the DTW algorithm, the failure points and behavioral deviations of the solution can be accurately located without the need for manual inspection of each step, significantly reducing the cost of manual inspection, improving the efficiency of design iteration, and reducing the iteration cycle and cost.

[0017] 2) Explicit and Role-Based Design Logic: For the first time, design thinking is upgraded from drawing to logical modeling. Implicit design experience is transformed into an operable and traceable network structure through a node-based architecture. At the same time, human-machine role attribute configuration is introduced to make design decisions traceable and auditable, solving the pain point of black box design logic.

[0018] 3) End-to-end semantic offset-free: The newly added problem node (D) and the strong-typed port mechanism of each node ensure that the semantic transmission of design intent from definition to execution is offset-free, avoids data mistransmission and intent deviation, and ensures the consistency between the design solution and the core requirements.

[0019] 4) Efficient Human-AI Collaboration: Through the deep integration of node-based AI agents, the intuitive judgment of human designers is combined with the efficient computing power of AI to achieve automated deduction, comparison and calibration of design thinking, reduce reliance on the personal experience of designers, and achieve the standardization and large-scale improvement of design value.

[0020] 5) Adapt to complex scenario requirements: Supports layered modeling of container nodes and multimodal data processing, which can adapt to design scenarios of different fields and different complexities. At the same time, the closed-loop calibration mechanism can dynamically adapt to changes in user behavior, improving the adaptability and stability of the design solution. Attached Figure Description

[0021] Figure 1 This is a diagram of the non-linear alignment algorithm model between behavior nodes Be1 and Ba2; Figure 2 This is a schematic diagram of the AI ​​agent atomic node architecture; Figure 3 This is a topology diagram of the main process flow of AI nodes and intelligent agent algorithms; Figure 4 This is a topology diagram of the adaptive feedback process of AI nodes and intelligent agents; Figure 5 This is a diagram of the ANC dynamic network evolution model in this embodiment. Detailed Implementation

[0022] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0023] like Figure 1-4 As shown, this invention provides a design thinking visualization system and method based on node-based behavior comparison and feedback calibration. Through a node-based AI agent architecture, core comparison algorithm, and closed-loop calibration logic, it achieves explicit and automated iterative optimization of design thinking. The specific technical solution is as follows: I. A Visualization System Based on Nodeated AI Agents The system includes a visual editor interface, a node library, and a simulation module. These modules work together to achieve visual modeling and automated calibration of design thinking. Visual editor interface: As the core interactive entry point of the system, it allows users to select atomic nodes from the node library, complete the node layout by dragging and placing, configure the role attributes and parameters of each node, and connect each atomic node through strongly typed ports to build a node network; at the same time, it displays the node network running status, data flow process and behavior comparison results (such as deviation heatmap, conflict warning line) in real time, transforming abstract algorithm logic into intuitive and perceptible visual information, supporting users to debug and iteratively optimize.

[0024] Node Library: Contains atomic nodes for various integrated AI agents. Each atomic node adopts a three-layer architecture of [input-processing-output] and realizes multimodal data transmission and collaborative computing through strongly typed ports. The strongly typed ports have a strict data type verification mechanism, which only allows ports with matching data types to establish connections, avoiding collaborative computing errors caused by mistransmission of different types of data, and ensuring the accuracy and stability of the data flow.

[0025] Simulation Module: Used to drive the node network, complete the core data flow transmission and feedback calibration closed loop, support the automated operation and cross-node collaboration of AI agents of each atomic node, and complete the simulation and iteration of the entire design thinking process without human intervention.

[0026] II. Configuration of Atomic Nodes and AI Agents Atomic nodes include at least a context node (C), a problem node (P), an expected behavior node (Be1), a solution node (S), an actual behavior node (Ba2), a comparison node (Compare), and a value node (V). Each atomic node acts as an independent micro-AI agent, responsible for semantic understanding, data processing, and logical operations in its corresponding stage. Each atomic node supports role attribute declaration, including Human and Machine roles. Different role attributes correspond to different algorithm preferences in the processing layer. Human roles focus on aesthetic and ethical calculations, while Machine roles focus on efficiency and compliance calculations. The specific configuration of each node is as follows: Context node (C): Utilizes LLM to extract contextual semantics and quantizes them. Input layer: Supports multimodal data access, including text, images, and numerical values ​​(market trend curves); The processing layer integrates a cross-modal coding agent, employs a feature fusion algorithm to extract semantic vectors and perform spatiotemporal context clustering; The output layer outputs a structured context tensor, providing basic data support for subsequent problem identification; Decision node (D) identifies pain points and sets intention anchors and decision boundaries.

[0027] Input layer: Receives input interfaces from the multimodal, including: the context tensor of the C node, the demand tensor of the P node, and the subjective intent input of the current role; Processing layer: Integrates decision weight intelligent agents, uses role-weighted operators W_r to set hard constraints or ethical boundaries of design goals, and generates design decision benchmarks; Output layer: Structured decision constraint stream or hierarchical wizard decision data stream, serving as criteria for subsequent scheme generation.

[0028] Problem node (P): Used to identify requirement gaps and define design goals; Input layer: Receives the context tensor output by the context node (C) and preset industry benchmark data. Processing layer: Performs logical breakpoint scanning through knowledge graph to automatically identify requirement gaps; Output layer: Outputs structured requirement tensors to clarify the core problems that the design needs to solve.

[0029] Expected Behavior Node (Be1): Used to deduce the user's operation path; Input layer: Receives prototype parameters and target user profiles from the scheme node (S); Processing layer: Employs behavioral dynamics simulation algorithms and Monte Carlo algorithms to simulate the distribution of user operation paths under different resistances; Output layer: Outputs the expected behavior feature vector sequence (S1) to form the expected behavior benchmark.

[0030] Solution Node (S): Used to generate interaction strategies and digital prototypes; Input layer: Receives the pain point vector and component library metadata output by the problem node (P); Processing layer: Uses diffusion model or Transformer architecture to generate topology or code framework; The output layer outputs digital prototype assets, thus concretizing the design scheme.

[0031] Actual Behavior Node (Ba2): Used to collect and process actual interaction data; Input layer: Integrates real-time event tracking data, simulator operation log streams, or interactive status capture data; Processing layer: Cleans, slices, and transforms unstructured data into standard behavior vectors; Output layer: Outputs the actual behavior vector sequence (S2), recording user behavior data in actual applications.

[0032] Value Node (V): Used to evaluate the effectiveness of the solution and trigger calibration; Input layer: Receives the similarity scores of the comparison nodes and the ROI evaluation formula; Processing layer: Performs value function derivation and determines the effectiveness of the solution; Output layer: Outputs calibration signals to reset the parameters of the scheme node (S) or expected behavior node (Be1) and drive scheme optimization.

[0033] Compare node: As the core of behavior comparison and role alignment, it is used to complete the accurate comparison between expected behavior and actual behavior. Specifically, it performs the following steps: (1) Path feature extraction: the expected behavior feature vector sequence (S1) and the actual behavior vector sequence (S2) are encoded respectively; (2) Dynamic time warping (DTW) alignment: the shortest alignment path and minimum cumulative distance of the two variable length sequences are calculated by the DTW algorithm to adapt to the differences in user operation rhythm; (3) Role alignment deviation calculation: the role weighting operator Wᵣ is introduced to calculate the semantic distance between the actual delivery trajectory and the design goal, and the serious deviation points are marked.

[0034] Input layer: The expected behavior tensor represents the ideal performance index of the design goal, and the actual behavior tensor represents the real performance of the system during operation or user testing. Optional context constraint tensor or value benchmark vector; Processing layer: data alignment and preprocessing, deviation calculation, feedback calibration generation; Output layer: Deviation vector, calibration signal vector, optional diagnostic output report; In addition, the system supports container nodes, which are used to encapsulate the child node network to form a hierarchical structure. The container nodes provide input and output ports to the outside world, and internally support hierarchical viewing of the complete child node micro-loop and data flow path, adapting to the hierarchical modeling needs of complex design scenarios.

[0035] Input layer: context input tensor, value input tensor, optional child node input; Processing layer: Input distribution and aggregation to internal child nodes, internal micro-loop management (supports drill-down viewing), feedback and calibration, supports arbitrary combination or disgrouping of child nodes to improve system scalability; Output layer: aggregated output tensor, value assessment data stream, and optional diagnostic output.

[0036] III. Closed-loop link and calibration logic The node network is connected via strongly typed ports, forming two closed-loop links to achieve fully automated iteration of the design process: (1) Main logic flow: The link is situation node (C) → problem node (P) → decision node (D) → expected behavior node (Be1) → decision node (D) → solution node (S) → actual behavior node (Ba2), which completes the whole process from situation perception, problem definition, expected deduction, solution generation to actual behavior collection, realizing the explicit modeling and execution of design thinking.

[0037] (2) Adaptive feedback flow: The link is actual behavior node (Ba2) → comparison node (Compare) → decision node (D) → value node (V) → decision node (D) → solution node (S) / expected behavior node (Be1); when the deviation value D calculated by the comparison node is greater than the preset threshold, the system automatically triggers parameter reset through adaptive feedback flow to complete the automatic fine adjustment of the solution; the calibration process can support two modes: implicit automatic detection trigger or explicit manual trigger, taking into account both automation efficiency and manual intervention flexibility.

[0038] IV. Visualization Methods for Design Thinking Based on the above system, this invention provides a design thinking visualization method, applied to the system, comprising the following steps: Step (1) Node building: Select atomic nodes from the node library through the visual editor interface, configure the role attributes and parameters of each node, connect each atomic node through strong typed ports to form a node network, and define the link rules of the main logic flow and the adaptive feedback flow. Step (2) Data input: Input multimodal situational data into situation node C, import preset industry benchmark data into problem node P, input subjective intent data of the current role into decision node D, input target user profile into expected behavior node Be1, and complete the initialization of all node data; Step (3) Simulation operation: Drive the node network to run through the simulation operation module. The main logic flow completes the entire process of situation perception, problem definition, decision constraint setting, expected behavior inference, solution generation and actual behavior collection, generating the expected behavior feature vector sequence S_1 and the actual behavior feature vector sequence S_2. Step (4) Behavior comparison: Compare the execution path feature extraction, DTW sequence alignment and role alignment deviation calculation of the comparison node, generate the deviation value D and the mark of the serious deviation point, and generate a visual deviation heat map and conflict warning line, transforming the abstract algorithm calculation results into intuitive and perceptible visual information; Step (5) Value assessment and closed-loop calibration: Value node V completes the scheme effectiveness assessment based on the deviation value D and ROI assessment formula. If the deviation value D is greater than the preset threshold, a calibration signal is triggered. The parameters of the scheme node S or the expected behavior node Be1 are adjusted back through the adaptive feedback flow. Step (3)-step (5) are repeated until the deviation value D is less than or equal to the preset threshold, and the design iteration is completed.

[0039] Example: Figure 5 As shown, the underlying system of this embodiment follows the ANC dynamic evolution network model W=(A, N, C), where A is the heterogeneous design intelligence library, N is the dynamic interaction topology between nodes, and C represents the driving situation space where different autonomous driving behavior designs are located.

[0040] At any time point t, after the system detects a long-tailed situation C(t), it will activate a specific agent from A (such as a perception agent A_{weather} that specializes in handling extreme weather) through the Global Controller to replace the general agent, thereby achieving dynamic adaptation to the problem space P(t).

[0041] The following is an example of setting up a long-tail scenario for autonomous driving: (1) Scenario: During a rainstorm at night, at an unprotected left-turn intersection, a food delivery vehicle is spotted driving against traffic 50 meters ahead; input multimodal data of images (intersection monitoring screen), numerical values ​​(distance, vehicle speed, rainfall) and text (intersection type, traffic rules) into the scenario node (C). The cross-modal coding agent of the scenario node performs feature fusion and spatiotemporal clustering, and outputs a structured scenario tensor (e.g., a tensor with the shape of [time step, feature dimension], containing quantitative features such as visibility and distance).

[0042] (2) Problem: Problem node (P) receives context tensor and obstacle avoidance benchmark data of autonomous driving industry, scans logical breakpoints through knowledge graph, identifies the core pain point that "the confidence of traditional perception model decreases under rain and fog obstruction, and there is a conflict between obstacle avoidance path and efficiency", and outputs structured demand vector.

[0043] (3) Main Logic Flow Decision (DN1: Between the Problem Node and the Expected Behavior Node): After P outputs the demand vector, the decision node DN1 is activated as a human-driven hierarchical guide. This node is driven by a rule base preset by human experts (such as a traffic law priority matrix and ethical guidelines), inputs the demand vector, and assesses potential risks (e.g., when the collision probability threshold is >0.1, mandatory safety priority is enforced). DN1 performs structured decision tree analysis, outputs an adjusted demand vector (e.g., with added safety weight factors), and passes it to the expected behavior node (Be1) to ensure that the decision is transparent and traceable.

[0044] (4) Expected behavior: Configure the expected behavior node (Be1) as a machine role, focusing on efficient and compliant operation; input the preset left turn obstacle avoidance prototype parameters (from the obstacle avoidance component library, sharing metadata with the solution node S), simulate the expected operation path of the delivery car decelerating and avoiding obstacles and the autonomous vehicle smoothly turning left through the Fogg behavior model (B=MAT) and Monte Carlo algorithm, and output the expected behavior feature vector sequence S_1.

[0045] (5) Main logic flow decision (DN2: between expected behavior node and solution node): After Be1 outputs S_1, the decision node DN2 is activated as another layer of human guidance. This node integrates the regulatory framework (such as the ISO 26262 safety standard), inputs S_1 and performs hierarchical review, and DN2 outputs a refined behavior sequence S_1' to ensure human intervention before solution generation and improve the robustness and legal compliance of the system.

[0046] (6) Scheme: The scheme node (S) receives the demand vector and metadata of the autonomous driving obstacle avoidance component library, uses the Transformer architecture (e.g., a variant based on the fine-tune of the autonomous driving dataset) to generate a smooth left turn obstacle avoidance topology and control code, and outputs a digital prototype asset (obstacle avoidance control scheme).

[0047] (7) Actual behavior: Real-time embedded data and simulator operation logs are collected by vehicle sensors to capture the actual behavior of the delivery vehicle that slips due to water accumulation on the road and speeds up. The actual behavior node (Ba2) cleans and slices the log data and transforms it into a standard behavior vector sequence S_2.

[0048] (8) Comparison: The Compare node performs path feature extraction, encodes S_1 and S_2, completes variable-length sequence alignment using the DTW algorithm, calculates the drastic deviation of the trajectory on the time axis (the delivery vehicle accelerates rather than decelerates), and introduces the role-weighted operator Wᵣ to mark the serious deviation point of "accelerating and rushing". The visual editor interface simultaneously generates a deviation heatmap and conflict warning line to intuitively display the behavioral deviation. Output deviation index, including a correction comparison for S_1' (if applicable).

[0049] (9) Adaptive Feedback Flow Decision (DN3: Between Comparison Node and Value Node): The comparison node outputs a deviation index to DN3, and calculates the deviation index (e.g., KL divergence measures path similarity). The decision node DN3 is activated as a human-made structured constraint. This node applies a predefined value framework (e.g., utility function: safety 0.6 + efficiency 0.3 + comfort * 0.1), inputs the deviation index, and executes the constraint decision; it outputs a constraint value vector, which is passed to the value node to ensure that the feedback does not deviate from human priorities.

[0050] (10) Value: The value node (V) receives the deviation score of the comparison node, calculates the safety risk score through the ROI evaluation formula, determines that the current solution is invalid (ROI is negative), outputs a calibration signal, triggers the feedback flow value node to apply the multi-objective optimization framework, calculates the utility delta, and outputs a calibration signal (e.g., a vector containing parameter adjustment instructions) if delta < threshold.

[0051] (11) Closed-loop calibration: (DN4: between the value node and the solution node) The system receives the constraint value vector of DN3 through the value node and quantifies the improvement potential (e.g., expected utility improvement of 15%). The decision node DN4 is activated as the final structured constraint. The final output of DN4 is an adaptive feedback stream that transmits the calibration signal to the solution node (S). The general obstacle avoidance agent is automatically suspended, and the high-order safety agent A with the wet road surface compensation operator is activated. The parameters of the solution node are adjusted, and the safety redundancy weight is increased from 0.3 to 0.95. The obstacle avoidance scheme of "emergency braking + steering compensation" is regenerated. The above steps are repeated until the deviation value D calculated by the comparison node is less than the preset threshold, and the iteration is completed.

[0052] In this embodiment (unprotected left turn at night in heavy rain), the system completes a closed loop from perception to decision calibration through the ANC architecture. When the autonomous vehicle enters this long-tail scenario, the Compare node on the visualization canvas immediately receives the standard compliant path vector S_1 generated by the Be1 expected behavior agent and the rain-induced slippage lag path vector S_2 obtained by the Ba2 actual observation agent. The system performs path feature extraction and nonlinear alignment through the DTW algorithm. This process is represented on the visualization terminal as a dynamic overlap analysis of the two behavioral trajectories in the time dimension: the algorithm accurately identifies phase lag and amplitude decay in the actual trajectory at a specific time step, indicating physical braking failure or a sharp decrease in road friction. At this time, the calculated values ​​are transformed into a visible deviation heatmap and conflict warning line on the canvas, transforming the abstract algorithm logic into a behavioral gap that designers can intuitively perceive.

[0053] Subsequently, value node V performs real-time value deduction based on the aforementioned visualized deviation, determining that the safety score of the current solution S has fallen below the steady-state threshold set by the design thinking model, and immediately generates a high-priority calibration signal. This signal, through the adaptive evolution mechanism of the ANC network, triggers a dynamic reorganization of node relationships on the canvas: the system automatically suspends the general driving agent, activates the higher-order safety agent A_{safety} with the slippery road surface compensation operator, and adjusts the edge attributes in the network topology in real time, increasing the safety redundancy weight from 0.3 to 0.95. This WYSIWYG node state switching not only gives the design thinking process extremely high transparency, but also enables the system to complete the solution reconstruction within 45ms after detecting the deviation. Ultimately, the vehicle's original expected action of "slow left turn" is calibrated in real time to the optimal obstacle avoidance strategy of "emergency braking and steering compensation".

[0054] This embodiment fully demonstrates that, through the ANC network and behavior alignment mechanism, the system can transform invisible perceptual deviations into a computable, observable, and automatically optimized design decision flow, significantly improving the safety resilience and evolution capability of the artificial intelligence-assisted system in extreme long-tail scenarios, thus verifying the effectiveness and practicality of the present invention.

Claims

1. A design thinking visualization system for node-based behavioral feedback calibration, characterized in that, It includes a visual editor interface, a node library, and a simulation operation module. The visual editor interface provides users with an intuitive environment for node operation and network building, supporting visual modeling, configuration, and debugging of design thinking. The node library contains atomic nodes of various integrated AI agents. Each atomic node adopts a three-layer architecture of input-processing-output. Each atomic node includes at least a context node C, a problem node P, a decision node D, an expected behavior node Be1, a solution node S, an actual behavior node Ba2, a comparison node Compare, and a value node V. Each atomic node acts as an independent AI agent, used for semantic understanding, data processing, and logical operations in the corresponding stage. The simulation operation module drives the node network to complete the core data flow transmission and feedback calibration closed loop.

2. The design thinking visualization system for node-based behavior feedback calibration according to claim 1, characterized in that, Each atomic node supports role attribute declaration, which includes the human role and the machine role. Different role attributes correspond to different algorithm preferences of the node processing layer.

3. The design thinking visualization system for node-based behavior feedback calibration according to claim 1, characterized in that, The context node C is used to extract background semantics and quantize them. Its input layer supports the access of multimodal data such as text, images, and numerical data. The processing layer integrates a cross-modal coding agent, uses a feature fusion algorithm to extract semantic vectors and perform spatiotemporal background clustering, and outputs a structured context tensor. The problem node P is used to identify demand gaps and define design goals. The input layer receives the scenario tensor output by the scenario node C and preset industry benchmark data. The processing layer performs logical breakpoint scanning through the knowledge graph. The output layer outputs a structured demand tensor. The decision node D is used to complete the intention anchoring and decision boundary setting after the requirement identification. The input layer receives a multimodal input interface, including the context tensor of the context node C, the requirement tensor of the problem node P, and the subjective intention input of the current role. The processing layer integrates the decision weight AI agent, which uses the role weighting operator Wr to set the hard constraints and ethical boundaries of the design goal, and completes the weight fusion of multi-dimensional inputs and the generation of decision benchmarks. The output layer outputs a structured decision constraint flow or a hierarchical wizard decision data flow, which serves as the core criterion for subsequent scheme generation.

4. The design thinking visualization system for node-based behavior feedback calibration according to claim 1, characterized in that, The expected behavior node Be1 is used to deduce the user's operation path. The input layer receives the prototype parameters of the scheme node S and the target user profile. The processing layer uses the behavior dynamics simulation algorithm and the Monte Carlo algorithm to simulate the distribution of the user's operation path under different resistances. The output layer outputs the expected behavior feature vector sequence S_1. The actual behavior node Ba2 is used to collect and process actual interaction data. The input layer receives real-time embedded data, simulator operation log stream or interaction status capture data. The processing layer cleans, slices and transforms unstructured data into standard behavior vectors. The output layer outputs the actual behavior vector sequence S_2.

5. The design thinking visualization system for node-based behavioral feedback calibration according to claim 4, characterized in that, The Compare node is used to compare the expected behavior with the actual behavior. Specifically, it performs the following steps: 1) Path feature extraction: Encode the expected behavior feature vector sequence S_1 and the actual behavior vector sequence S_2 respectively; 2) Dynamic Time Warping (DTW) alignment: Calculate the shortest alignment path and minimum cumulative distance between the two variable-length sequences using the DTW algorithm to adapt to the differences in user operation rhythm. 3) Role alignment deviation calculation: Introduce the role weighting operator Wr to calculate the semantic distance between the actual delivery trajectory and the design goal.

6. The design thinking visualization system for node-based behavioral feedback calibration according to claim 1, characterized in that, The solution node S is used to generate interaction strategies and digital prototypes. The input layer receives the pain point feature vector output by the problem node P, the structured decision constraint flow output by the decision node D, and the component library metadata. The processing layer uses a diffusion model or Transformer architecture to generate the topology or code framework. The output layer outputs digital prototype assets. The value node V is used to evaluate the effectiveness of the scheme and trigger calibration. The input layer receives the similarity score and ROI evaluation formula of the comparison node Compare. The processing layer performs the value function derivation. The output layer outputs a calibration signal to reset the parameters of the scheme node S or the expected behavior node Be1.

7. The design thinking visualization system for node-based behavior feedback calibration according to claim 1, characterized in that, The node network forms two closed-loop links: one is the main logic flow, which is: Context node C → Problem node P → Decision node D → Expected behavior node Be1 → Solution node S → Actual behavior node Ba2; the other is the adaptive feedback flow, which is: Actual behavior node Ba2 → Compare node → Value node V → Solution node S / Expected behavior node Be1 / Decision node D. When the deviation value D calculated by the Compare node is greater than the preset threshold, the parameters are automatically reset through the adaptive feedback flow to complete the automatic fine-tuning of the solution.

8. The design thinking visualization system for node-based behavior feedback calibration according to claim 1, characterized in that, The node library also includes container nodes, which are used to encapsulate the child node network, forming a hierarchical structure of composite nodes, aggregating the inputs / outputs of multiple child nodes, and supporting recursive processing. The container node provides external input and output ports, and internally, it enables hierarchical viewing to see the complete micro-loop of child nodes and the data flow path.

9. A design thinking visualization method for node-based behavioral feedback calibration, characterized in that, A design thinking visualization system for node-based behavior feedback calibration as described in any one of claims 1-8 includes the following steps: Step 1) Node building: Select atomic nodes from the node library through the visual editor interface, configure the role attributes and parameters of each node, connect each atomic node through type ports to form a node network, and define the link rules of the main logic flow and adaptive feedback flow. Step 2) Data Input: Input multimodal context data into context node C, import preset industry benchmark data into problem node P, input the subjective intent data of the current role into decision node D, input the target user profile into expected behavior node Be1, and complete the initialization of all node data; Step 3) Simulation Operation: Drive the node network to run through the simulation operation module. The main logic flow completes the entire process of situation awareness, problem definition, decision constraint setting, expected behavior inference, solution generation and actual behavior collection, generating the expected behavior feature vector sequence S_1 and the actual behavior feature vector sequence S_2. Step 4) Behavior comparison: Compare the execution path feature extraction, DTW sequence alignment and role alignment deviation calculation of the comparison node, and generate the deviation value D and the deviation point marker; Step 5) Value assessment and closed-loop calibration: Value node V completes the scheme effectiveness assessment based on the deviation value D and the ROI assessment formula. If the deviation value D is greater than the preset threshold, a calibration signal is triggered. The parameters of the scheme node S, the expected behavior node Be1, or the decision node D are adjusted back through the adaptive feedback flow. Steps 3) to 5) are repeated until the deviation value D is less than or equal to the preset threshold, and the design iteration is completed.

10. The method according to claim 9, characterized in that, In step 4), the behavior comparison process also includes generating a visual deviation heatmap and conflict warning lines, transforming the abstract algorithm calculation results into intuitive and perceptible visual information.

Images