A cognitive evolution reconstruction method and system based on cognitive digital twinning

Abstract

The application relates to a cognitive evolution reconstruction method and system based on cognitive digital twinning, and belongs to the technical field of cognitive data processing. The method comprises the following steps: acquiring multi-modal cognitive data of a target object, performing layered cognitive imaging processing on the multi-modal cognitive data, locking a deviation bottom mechanism label and generating a cognitive evidence atom; performing state machine determination on a cognitive state, and outputting a ladder intervention task; generating a logic isomorphism and surface heteromorphism verification task for the cognitive state after intervention, and returning the non-migrated state to the intervention stage according to the cognitive structure migration determination result; performing steady-state training on the migrated cognitive structure and generating a corresponding training task; and writing the cognitive evidence atom, the verification determination result and the steady-state training result into a cognitive digital twinning carrier through a state streaming writing architecture, and outputting a layered decision signal. The method realizes accurate positioning of cognitive bias layer by layer and steady-state reconstruction of a cognitive structure, and improves the accuracy and effectiveness of cognitive intervention.

Description

Technical Field

[0001] This invention belongs to the field of cognitive data processing technology, specifically relating to a cognitive evolution reconstruction method and system based on cognitive digital twins. Background Technology

[0002] With the continuous development of intelligent education technology, diagnosing learners' cognitive deficiencies and implementing personalized interventions using information technology has become a research hotspot. Most existing cognitive diagnostic solutions only conduct superficial analysis based on the final results of answers, merely determining whether the answer is right or wrong, failing to reconstruct the complete thought process or pinpoint the deep structural roots of cognitive biases. Some solutions introduce knowledge graphs to model cognitive structures, but their diagnostic granularity remains at the level of knowledge point mastery, without deconstructing the deep mechanisms of the thought process.

[0003] Current systems generally lack a progressive, step-by-step analytical mechanism for thinking, failing to break down learners' problem-solving processes, thought processes, and behavioral habits layer by layer, and making it difficult to distinguish between short-term rote memorization and genuine cognitive mastery. Traditional intervention models suffer from process fragmentation: after intervention is initiated, there is a lack of standardized verification procedures, making it difficult to determine whether learners have truly completed cognitive structure transfer; and when some pre-interventions fail to achieve the desired results, there is a lack of controllable retry mechanisms, easily leading to multiple inefficient guidance sessions.

[0004] Meanwhile, existing solutions have vague intervention boundaries, with drawbacks such as directly outputting complete answers and replacing learners' independent reasoning; they lack tiered guidance strategies and cannot match graded prompts according to the severity of deviations, easily leading to over-intervention or under-guidance. Learner cognitive data is generally stored in a static snapshot manner, recording only scattered information at a single point in time, making it difficult to continuously track the dynamic evolution of cognitive abilities with training, and failing to construct a complete and traceable cognitive timeline trajectory.

[0005] In addition to process and technical deficiencies, the existing system also has shortcomings in terms of trust management: intervention decisions are issued entirely autonomously by the algorithm, lacking a channel for manual intervention by teachers; various cognitive processing operations lack unified auditing and record-keeping methods, which is not conducive to the division of teaching responsibilities and data compliance management.

[0006] The aforementioned problems collectively lead to a high degree of blindness in cognitive interventions, a lack of interpretable structured evidence to support diagnostic results, and difficulty in achieving long-term stable cognitive structure reconstruction. This invention aims to provide a method and system for cognitive evolution reconstruction based on cognitive digital twins. Summary of the Invention

[0007] To address the aforementioned problems in the existing technology, this invention provides a cognitive evolution reconstruction method and system based on cognitive digital twins. The objective of this invention can be achieved through the following technical solutions: Acquire multimodal cognitive data of the target object’s explicit response information, implicit behavioral information, and subjective expression information; perform hierarchical cognitive imaging processing on the multimodal cognitive data; use the analysis results of the previous level as the input of the next level to gradually narrow the scope of cognitive bias localization, lock the underlying mechanism label of the bias and generate cognitive evidence atoms; Using the cognitive evidence atom as the trigger event input, the cognitive state machine is subjected to state transition determination, and a step-by-step intervention task is output. Based on the cognitive state after intervention, logically isomorphic but superficially heterogeneous verification tasks are generated and cognitive structure transfer is judged. Cognitive states whose corresponding cognitive structures have not been transferred are reverted to the intervention stage and a step-by-step intervention task is re-outputted. Steady-state training is performed on the transferred cognitive structures and corresponding training tasks are generated. The cognitive evidence atoms, verification judgment results, and steady-state training results are written into the cognitive digital twin carrier through a state-stream writing architecture, and the hierarchical decision signal is output after being updated by a time-series cognitive snapshot.

[0008] Specifically, the hierarchical cognitive imaging process includes five levels: The result layer determines the correctness of the answer data based on the cognitive task, generates a result layer feature vector, and marks abnormal answers. The path layer uses the abnormal answer item as the retrieval anchor point to locate the corresponding deviation point in the problem-solving step sequence and generate a path layer deviation vector. The expression layer uses the steps corresponding to the deviation points as the focus area, performs semantic analysis on the text and voice data of the target object within the focus area, identifies abnormal patterns in thought expression, and generates abnormal features of the expression layer. The behavior layer uses the steps corresponding to the deviation points as time windows to perform behavior pattern analysis on the interaction sequence and handwriting trajectory data of the target object within the window, and determines the type of cognitive behavior inertia. The evolution layer performs joint temporal modeling based on the outputs of the result layer, path layer, expression layer, and behavior layer, and outputs the underlying mechanism label of the deviation.

[0009] Specifically, the generation process of the cognitive evidence atoms includes: The underlying mechanism label of the deviation is used as the deviation type field, the output results of each level and the deviation point sequence are used as the structured evidence chain in the evidence set field, and the deviation type field is mapped to the predefined intervention protocol identifier; the deviation type field, the structured evidence chain and the intervention protocol identifier are assembled into the cognitive evidence atom.

[0010] Specifically, during the execution of each level of the hierarchical cognitive imaging process, the confidence level of the feature vector output by each level is monitored. When the confidence level is lower than a preset threshold, a forced interaction interruption signal is triggered, and three types of structured prompt texts are extracted from a preset prompt template library. The first type of prompt text is used to induce the target object to state the basis of the deduction, the second type of prompt text is used to induce the target object to state the dependency relationship between knowledge points, and the third type of prompt text is used to induce the target object to re-execute the deduction under the assumption of reversing the original task conditions. The response text of the target object to the three types of prompts is obtained, semantic feature extraction is performed on the response text, the extracted semantic features are written back to the evidence set field as supplementary evidence, and the locking operation of the bias underlying mechanism label is re-executed based on the updated evidence set to generate updated cognitive evidence atoms.

[0011] Specifically, the path layer includes: Obtain the sequence of steps and nodes of the standard cognitive path, and perform an alignment operation between the sequence of problem-solving steps and the sequence of steps and nodes; mark the steps and nodes that do not match in the alignment operation as deviation points, and assign a deviation type label to each deviation point; output the sequence of deviation points containing the deviation type label as the path layer deviation vector.

[0012] Specifically, the state transition determination of the cognitive state machine includes: Based on a preset set of transition conditions between cognitive stages, the feature vector of the current cognitive state is compared with the trigger threshold of each condition in the set of transition conditions. When any trigger threshold is met, the current cognitive state is transferred to the next cognitive stage corresponding to the trigger threshold. The set of transition conditions includes at least three types of trigger conditions: the first condition is when cognitive evidence atoms are generated, the transition is from the deviation identification stage to the intervention execution stage; the second condition is when the verification is passed, the transition is from the variant verification stage to the steady-state training stage; and the third condition is when the verification fails, the transition is from the variant verification stage back to the intervention execution stage.

[0013] Specifically, the state transition determination also includes: In response to a verification failure event triggering a rollback to the intervention execution stage, the number of rollbacks is recorded. After the intervention task is re-output, the verification task generation is executed again. When the cognitive structure migration determination result is again determined to be non-migrated and the number of rollbacks reaches the preset upper limit, the current cognitive state is marked as pending manual intervention, the automatic flow is paused, an alarm signal is sent to the administrator terminal, and the automatic intervention distribution process is terminated.

[0014] Specifically, the generation process of the verification task includes: Extract the inference nodes and concept dependency edges from the original cognitive task corresponding to the cognitive state after intervention, and use the inference nodes and concept dependency edges to form a cognitive logic skeleton; When the cognitive evidence atom contains supplementary semantic features obtained through additional interactive prompts, the cognitive logic skeleton composed of the inference node and the concept dependency edge is updated based on the supplementary semantic features. Select a target template that matches the cognitive logic skeleton from the predefined variant template library, and perform replacement on the context elements and numerical representations according to the replacement rules defined in the target template to generate a verification task.

[0015] Specifically, the generation process of the tiered intervention task includes: Based on the severity level of the bias in the cognitive evidence atoms, a corresponding guidance level is matched. The guidance level includes a directional prompt level, a step-by-step guidance level, and a demonstration deduction level. Each level is associated with a directional prompt template, a step-by-step guidance template, and a demonstration deduction template, respectively. An intervention task is generated according to the template corresponding to the matched guidance level. In the second interaction mode, the generated items associated with the complete answer in the intervention task are marked as masked items, and the masked items are filtered out at the output.

[0016] Specifically, the steady-state training includes: A sequence of inference tasks is generated based on the inference nodes and concept dependency edges corresponding to the transferred cognitive structure. Each inference task in the sequence adds a concept dependency edge to the previous inference task while keeping the inference nodes unchanged. The inference tasks are pushed to the target object in sequence, and the completion accuracy and response latency of each inference task are collected. The cognitive stability index is updated based on the completion accuracy and response latency. When the cognitive stability index reaches a preset solidification threshold, the transferred cognitive structure is marked as a steady-state cognitive structure.

[0017] Specifically, when the state-stream writing architecture performs a write operation, it encapsulates the cognitive evidence atom, the verification judgment result, and the steady-state training result into event records. Each event record contains a timestamp and a version number, and is stored in the cognitive digital twin carrier in chronological order by appending the write operation. In response to the time-series cognitive snapshot update request, it reads all event records within the time window corresponding to the current cognitive state and merges all event records in chronological order to generate a time-series cognitive snapshot.

[0018] Specifically, the method further includes: Each cognitive processing operation generates an immutable audit record with a timestamp; the step-by-step intervention task must receive a manual confirmation instruction before it is issued, and the intervention issuance process is terminated if no manual confirmation instruction is received.

[0019] The beneficial effects of this invention are as follows: (1) This invention constructs a hierarchical cognitive imaging processing mechanism, which sequentially performs result-level judgment, path-level comparison, expression-level analysis, behavior-level identification and evolution-level tracking on learners' multimodal cognitive data, realizing the deconstruction of the thinking process from the surface result to the deep structure, accurately locating the underlying mechanism of cognitive bias, effectively distinguishing between occasional operational errors and stable cognitive structure biases, overcoming the shortcomings of existing technologies that only rely on the answer results to carry out shallow analysis and cannot restore the complete thinking path, and providing traceable structured evidence support for precise cognitive intervention.

[0020] (2) This invention establishes a closed-loop cognitive governance system that integrates state machine determination, step-by-step intervention, logically isomorphic and superficially heterogeneous verification, and steady-state training. By accurately determining the cognitive structure transfer effect through logically isomorphic and superficially heterogeneous verification tasks, the untransferred cognitive state is returned to the intervention stage to re-output the intervention task. Steady-state training is then carried out on the transferred cognitive structure to achieve root cause repair of cognitive bias and long-term stable solidification of cognitive structure. This overcomes the shortcomings of existing technologies, such as lack of standardized verification after intervention and difficulty in distinguishing between short-term rote memory and the degree of real cognitive mastery.

[0021] (3) This invention uses a state-stream writing architecture to continuously write cognitive evidence atoms, verification judgment results, and steady-state training results into the cognitive digital twin carrier. It relies on time-series cognitive snapshots to dynamically update and output hierarchical decision signals, thereby realizing the assetization of cognitive data and dynamic tracking of the entire life cycle of cognitive state. This overcomes the shortcomings of existing technologies that use static snapshot storage, can only record single-point scattered information, and cannot construct a complete cognitive time-series evolution trajectory.

[0022] (4) This invention generates tamper-proof audit records with timestamps for each cognitive processing operation, and sets up a manual confirmation control link before the step-by-step intervention task is issued. If no manual confirmation instruction is received, the intervention issuance process is terminated, which improves the credible control mechanism of the algorithm's autonomous decision-making, makes up for the shortcomings of the existing solution that lacks a teacher manual intervention channel and the operation process has no audit trace, facilitates the division of teaching responsibilities, and effectively meets the data compliance management needs in the intelligent teaching scenario. Attached Figure Description

[0023] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0024] Figure 1 This is a flowchart illustrating a cognitive evolution reconstruction method based on cognitive digital twins according to the present invention. Figure 2 This is a schematic diagram of the architecture of a cognitive evolution reconstruction system based on cognitive digital twins in this invention; Figure 3 This is a schematic diagram of the collaborative architecture of system components in this invention. Detailed Implementation

[0025] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided.

[0026] The overall technical architecture of this invention adopts a top-down five-layer progressive layered architecture design, namely the experience layer, control layer, engine layer, data layer, and trust layer. The overall system architecture can be found in the appendix. Figure 2 The system comprises the following layers: the experience layer, responsible for multimodal cognitive data acquisition and standardized processing, and hierarchical decision signal output; the control layer, responsible for cognitive state machine scheduling, process orchestration and control, and human-machine instruction reception; the engine layer, which centrally carries all the core algorithm operations of this method, including five-layer progressive cognitive imaging and analysis, heterogeneous verification task generation, cognitive transfer quantification and judgment, branch backflow control, and steady-state training operations; the data layer, relying on the cognitive digital twin carrier, realizes streaming data storage and time-series snapshot aggregation; and the trust layer realizes full-link auditing and manual control, ensuring that the entire process operation is traceable and verifiable.

[0027] Based on the five-layer macro-architecture, the core functional logic of this invention is implemented collaboratively by three core functional components. The component collaboration architecture can be found in the appendix. Figure 3 Specifically, it includes a cognitive scheduling and control unit, an MRT cognitive imaging engine, and a CDT cognitive digital twin carrier. The cognitive scheduling and control unit corresponds to the system control layer, responsible for global state scheduling and process management; the MRT cognitive imaging engine corresponds to the system engine layer, undertaking the core work of cognitive data parsing and algorithm inference; and the CDT cognitive digital twin carrier corresponds to the system data layer, responsible for the persistent storage and dynamic evolution output of cognitive time-series data. These three components form a closed-loop collaborative flow through scheduling command streams, structured cognitive data streams, and historical data callback streams, providing hardware and software support for the entire process of cognitive diagnosis, intelligent intervention, effect verification, and steady-state solidification of this invention.

[0028] Based on the aforementioned hierarchical architecture and component collaboration system, the method of this invention sequentially executes a complete closed-loop process: multimodal cognitive data hierarchical imaging and diagnosis, intelligent intervention of cognitive state machines, heterogeneous verification and transfer judgment, cognitive steady-state solidification training, and time-series data storage and hierarchical decision output. This process gradually converges cognitive biases, solidifies high-quality cognitive structures, and ultimately achieves long-term reconstruction of learners' cognitive abilities. The following detailed, detailed, and reproducible description of the implementation methods for each step, using specific embodiments, provides a complete explanation.

[0029] like Figures 1-3 A cognitive evolution reconstruction method based on cognitive digital twins: Acquire multimodal cognitive data of the target object, perform hierarchical cognitive imaging processing on the multimodal cognitive data, use the analysis results of the previous level as the input of the next level to gradually narrow the range of cognitive bias localization, lock the underlying mechanism label of the bias and generate cognitive evidence atoms; The cognitive state is determined by a state machine, and the cognitive state that meets the transition conditions is used as an intervention trigger signal to output a tiered intervention task. The system generates logically isomorphic but superficially heterogeneous verification tasks for the cognitive state after intervention. Based on the verification results, it determines the transfer of cognitive structure and reverts the cognitive state for which the corresponding cognitive structure has not been transferred to the intervention stage to output a step-by-step intervention task. The system performs steady-state training on the transferred cognitive structure and generates corresponding training tasks. The cognitive evidence atoms, verification judgment results, and steady-state training results are written into the cognitive digital twin carrier through a state-stream writing architecture, and the hierarchical decision signal is output after being updated by a time-series cognitive snapshot.

[0030] Step 1: Obtain multimodal cognitive data of the target object, perform hierarchical cognitive imaging processing on the multimodal cognitive data, use the analysis results of the previous level as the input of the next level to gradually narrow the range of cognitive bias localization, lock the underlying mechanism label of the bias and generate cognitive evidence atoms.

[0031] In this embodiment, the system first collects multimodal cognitive data. The raw data is then aggregated and preliminarily cleaned by the multimodal data stream receiver of the MRT cognitive imaging engine, and finally, a five-layer recursive parsing is performed by the hierarchical cognitive imaging processor. This embodiment corresponds to the core algorithm capabilities of the system engine layer and the core functions of the MRT cognitive imaging engine, executing operations such as multimodal cognitive data acquisition, standardized preprocessing, structured vectorized encoding, five-layer hierarchical cognitive imaging parsing, metacognitive information completion, and cognitive evidence atom encapsulation. The finally generated cognitive evidence atoms are transmitted to the cognitive scheduling and control unit through the cognitive data output interface of the MRT cognitive imaging engine for subsequent state machine determination. The specific steps are as follows: 1. Multimodal cognitive data collection and refined screening The system collects data from three types of heterogeneous data sources. Each type of data is equipped with clear capture rules, cleaning standards, and effective information filtering logic, covering learners' explicit response information, implicit behavioral information, and subjective cognitive information, comprehensively covering the data dimensions required for cognitive diagnosis.

[0032] Explicit answer text data collection: This includes the text of solution steps, the text of the answer conclusion, and the answer content corresponding to the question stem. The system monitors user input behavior in real time through the front-end DOM, capturing the user's input character stream line by line, and automatically cutting independent step units according to carriage returns, paragraph breaks, and question numbers. Simultaneously, data cleaning operations are performed to filter invalid spaces, redundant line breaks, meaningless punctuation, and duplicate input content, retaining only complete, coherent, and valid step-by-step solution logic text, ultimately forming a structured and ordered sequence of solution steps.

[0033] Interactive temporal behavior data collection: This includes total answering time for a single question, text modification points, and pause intervals. The system monitors front-end text editing events with millisecond-level precision, recording the offset position and corresponding timestamp of every content addition, deletion, and replacement operation. A 2-second threshold is set for valid pauses; intervals with no operation lasting longer than 2 seconds are considered valid pauses. The system automatically segments pause intervals, statistically analyzing the number of pauses, the duration of single pauses, and the overall pause distribution characteristics to accurately characterize learners' implicit behavioral features during problem-solving, such as hesitation, logical blockage, and interruption of thought.

[0034] Learner subjective feedback text data collection includes reflections on incorrect answers, self-descriptions of problem-solving approaches, and descriptions of difficult questions. This type of data is used to compensate for implicit cognitive biases that cannot be covered by objective answer data and behavioral data, supplement learners' subjective cognitive perspectives, improve the information dimensions of cognitive diagnosis, and avoid biased and one-sided positioning caused by single data.

[0035] 2. Multimodal data fusion vectorized encoding processing Since the original text data is unstructured and cannot be directly used in multi-level quantization and parsing operations, the system adopts a differentiated fusion encoding strategy to generate a unified 300-dimensional fixed-dimensional feature vector based on the differences between long and short texts, thereby achieving the standardization and homogenization of multi-source heterogeneous data.

[0036] Short text encoding processing: For short text content such as single-sentence problem-solving steps, short-sentence answers, and brief concept descriptions, a pre-trained Word2Vec word embedding model for the education domain is used for encoding. Based on a dedicated educational vocabulary, each word is mapped to a 300-dimensional dense vector. Then, average pooling is used to integrate all word vectors within a single text to generate the overall feature vector corresponding to a single short text, fully preserving the text's semantics and knowledge point association features.

[0037] Weighted Correction Processing for Long Texts: For long texts such as reflections on incorrect answers, complete thought processes, and explanations of difficult points, a TF-IDF statistical model is used to weight and correct the Word2Vec base vectors. By statistically analyzing word frequencies and inverse document frequencies, core knowledge points and bias characteristic keywords are selected, increasing the vector weight of core feature words, weakening the interference of invalid modifiers and redundant sentences, avoiding the semantic dilution problem of long texts, and strengthening the representation ability of cognitive bias characteristics.

[0038] Formal definition of a standardized dataset: After encoding, a multimodal standardized dataset with completely unified dimensions and units is formed, formally expressed as: D={d1,d2,...,d n}, Where D represents the set of all standardized samples in a single cognitive diagnostic task; n is the total number of valid samples in the current diagnostic scenario; This is the 300-dimensional standardized feature vector of the i-th element after cleaning, word segmentation, and fusion encoding, providing a unified standardized input for subsequent hierarchical progressive parsing.

[0039] 3. Five-layer progressive hierarchical cognitive imaging analysis This hierarchical analysis strictly follows the core design principles of layer-by-layer convergence, layer-by-layer dependency, from surface to depth, and progressive in-depth exploration. Each layer utilizes the analytical markers (abnormal answers, deviation sequence, and labels) of the previous layer as focusing constraints, extracting corresponding subsets from the original multimodal data for targeted analysis, gradually compressing invalid features, and narrowing the convergence deviation location range. The specific layered operation is as follows: First layer: Result layer judgment (surface deviation locking) Using a standardized multimodal dataset D as input, the core principle is to identify surface-level answer biases. The system globally compares all sample feature vectors with a pre-built standard answer feature library, statistically analyzing three core indicators: correct / incorrect binary labels, score percentage, and answer feature bias distance, and then fuses these to generate result-level feature vectors. This quantitatively characterizes the overall deviation of the final answer results; simultaneously, it marks abnormal answers (i.e., a list of question IDs that were answered incorrectly or not at all), providing retrieval anchors for the next level. This completes the initial positioning of the surface-level answer questions.

[0040] Second layer: Path layer comparison (step logic deviation decomposition) This layer uses the abnormal answers marked in the previous layer (result layer) as retrieval anchors, extracting the subsequences of steps corresponding to the abnormal questions from the original problem-solving step sequence as comparison objects. The system pre-stores standard problem-solving step sequences and the user's actual problem-solving step sequences, formally defined as: S std ={s1,s2,...,s m},Susr ={u1,u2,...,u k}, Among them, S std For the standard cognitive step sequence, s i S is the feature vector of the i-th standard step; usr This is a sequence of actual problem-solving steps for the user, u j Let be the feature vector of the user at step j; m and k are the total number of standard steps and the total number of steps actually performed by the user, respectively.

[0041] The step matching error is calculated using a sequence Euclidean alignment algorithm to quantify the logical differences between steps. The formula is as follows: , Among them, ||s i -u j || represents the L2 norm distance, used to characterize the degree of difference in single-step problem-solving logic; L align Match the total loss to the global steps. L align The higher the value, the lower the match between the user's problem-solving steps and standard cognitive logic, and the more significant the cognitive bias. The system automatically marks mismatched steps based on loss distribution characteristics, designating unmatched step nodes in the alignment operation as deviation points, and assigning deviation type labels to each deviation point (e.g., conceptual confusion, logical omission, calculation error). The sequence of deviation points containing deviation type labels is then used as the path-layer deviation vector. The output provides the positioning basis for the focus area and time window for the next level.

[0042] The third layer: Expression layer analysis (syntactic and logical defect detection) The steps corresponding to the deviation points located in the previous layer (path layer) are defined as the focus area. The text input stream and corresponding speech-to-text within this focus area are extracted from the original multimodal data as the analysis objects. A syntactic dependency parsing algorithm is used to semantically decompose the text within the focus area, extracting four types of abnormal expression features: sentence confusion, logical jumps, conceptual misuse, and incomplete expression. These features are then quantified and integrated to generate an expression layer feature vector. It accurately characterizes the implicit cognitive deficiencies in users' thinking and expression.

[0043] Fourth layer: Behavioral layer recognition (distinguishing and judging deviation attributes) The steps corresponding to the deviation points located in the previous layer (path layer) are used as time windows. Handwritten trajectory data, UI operation records, and pause intervals within this time window are extracted from the original interaction time sequence logs as the analysis objects. Behavioral features such as handwritten trajectory pressure variance, UI operation backtracking times, and effective thinking pause intervals within this window are extracted. A hidden semi-Markov model is used to determine the type of cognitive behavioral inertia, and a behavioral layer feature vector is output. This completes the accurate classification of deviation attributes.

[0044] Fifth layer: Evolutionary layer tracing (root cause analysis of deviations) Output feature vectors from the behavior layer For and the output of the first four layers , , As input, through joint temporal modeling (improved LSTM network, with a forgetting gate introducing a cognitive decay factor), the feature vectors output from the result layer, path layer, expression layer, and behavior layer are integrated to output a unique bias underlying mechanism label. bias It accurately pinpoints the root causes of cognitive problems, completing the full-dimensional deviation localization from surface phenomena to underlying mechanisms.

[0045] 4. Metacognitive Information Completion Mechanism To address the issues of missing data and insufficient feature confidence during hierarchical parsing, a real-time confidence monitoring mechanism is configured throughout the process, with a preset fixed confidence threshold Th. conf =0.6. The system calculates the confidence score of the output features level by level to determine the validity and completeness of the parsing results at the current level.

[0046] Interruption Trigger Operation: If the confidence level of any output feature is <0.6, immediately pause all subsequent lower-level parsing processes and trigger a forced interactive interruption signal. The system extracts three types of structured prompt text from a pre-set prompt template library: the first type of prompt text is used to induce the target object to state the basis of deduction (i.e., evidence question); the second type of prompt text is used to induce the target object to state the dependency relationship between knowledge points (i.e., structure question); the third type of prompt text is used to induce the target object to re-execute the deduction under the assumption of reversing the original task conditions (i.e., reverse question).

[0047] Semantic completion operation: For the collected user subjective feedback text, a pre-trained BERT model is used to complete deep semantic encoding. Through bidirectional context modeling, semantic similarity calculation, and implicit cognitive feature mining, supplementary semantic feature vectors are generated. The semantic feature is written back as supplementary evidence to the evidence set field of the cognitive evidence atom, and the locking operation of the bias underlying mechanism label is re-executed based on the updated evidence set to generate an updated cognitive evidence atom. The updated cognitive evidence atom is further used to update the cognitive logic skeleton on which subsequent verification tasks are generated, so that the verification tasks can specifically verify the bias underlying mechanism label corrected by the supplementary evidence.

[0048] Atomic structured encapsulation of cognitive evidence After completing the analysis of all levels and the completion of metacognitive information, the system integrates all dimensional feature information, structures and encapsulates it to generate standardized cognitive evidence atoms, which serve as the sole standardized output of this cognitive diagnosis. Formalized as follows: Atom={Label bias ,Evd seq ID inter}, Among them, Label bias As a root cause type label for the bias, it uniquely identifies the underlying cause of cognitive bias; Evd seq To form a multi-level feature fusion evidence chain, by , , , , Weighted concatenation generates a complete and traceable system of evidence of cognitive bias; ID inter The intervention strategy is encoded to correspond to the deviation, so as to achieve a precise binding between the deviation type and the personalized intervention plan.

[0049] Step 2: Perform state machine judgment on the cognitive state, and use the cognitive state that meets the transition conditions as the intervention trigger signal to output a step-by-step intervention task.

[0050] In this embodiment, cognitive state quantitative control and automated flow are performed based on a preset cognitive state space, a cyclical intervention flow restriction mechanism is configured, and a tiered intervention task is generated according to the cognitive deviation level and intelligent answer shielding is performed to achieve refined and controllable cognitive intervention guidance.

[0051] In this embodiment, when the cognitive state machine determines that the intervention triggering condition is met (see...), Figure 1 The cognitive scheduling control unit, through Figure 2 The global scheduling command stream shown sends scheduling commands containing cognitive evidence atoms and state information to the MRT cognitive imaging engine. This embodiment corresponds to the system control layer's scheduling and management capabilities and the core functions of the cognitive scheduling control unit (such as...). Figure 2 , Figure 3 The architecture specifically implements operations such as cognitive state quantitative modeling, automated rigid state transition, cyclical intervention limit, three-tiered intervention matching, and intelligent intervention task masking. After the intervention task is generated, it is processed by the answer masking strategy and pushed to the learner's terminal through the experience layer interface. The specific steps are as follows: 1. Cognitive state space quantification modeling The system constructs a four-state closed-loop, mutually exclusive and controllable cognitive state space for the entire cognitive repair lifecycle, fully covering the entire process of cognitive diagnosis, intervention, verification, and consolidation, formally expressed as: St={St bias St interSt verify St train}, Symbol Explanation: St bias In the state of deviation identification, the characterization system has completed full-dimensional cognitive diagnosis and generated complete and valid cognitive evidence atoms; St inter To indicate the intervention execution status, the system has completed deviation grading and is awaiting the issuance of personalized intervention guidance tasks; St verify To verify the judgment status, indicating that the user has completed the intervention learning and is waiting for the system to verify the intervention effect; St train This represents a steady-state training state, indicating that cognitive biases have been corrected, and the program has entered a phase of long-term cognitive consolidation and reinforcement. The four states are independent of each other, form a closed loop, and are interconnected without any process breaks.

[0052] 2. Rigid Quantization State Flow Control The system is configured with three sets of quantifiable, rigid state transition rules that eliminate subjective judgment. All state transitions rely on explicit pre-quantifiable conditions, enabling automated and precise process flow. Diagnosis to Intervention Transition: When a complete cognitive evidence atom is successfully generated and the confidence levels of all five analytical levels of features meet the criteria, the cognitive state transitions from St... bias →St inter The intervention process will be initiated automatically.

[0053] Verification of steady-state transition: When the subsequent cognitive transfer score meets the standard, the cognitive bias is deemed to have been effectively corrected, and the cognitive state transitions from St to steady state. verify →St train The steady-state solidification training process is automatically initiated.

[0054] Verification of regression intervention jump: When the cognitive transfer score fails to meet the standard, the intervention is deemed ineffective and the bias is not corrected, and the cognitive state changes from St. verify →St inter This will automatically trigger the intervention and retry process.

[0055] 3. Circular intervention and flow restriction control operation To avoid ineffective loop interventions, conserve system resources, and avoid overlearning issues, the system is configured with quantized rollback counting and rate limiting termination mechanisms: Counter initialization: When a single cognitive bias correction process is initiated, the global rollback counter is initialized to 0. Formal expression: Cnt roll =0.

[0056] Counter update rule: For each verification failure or state rollback to the intervention state, the counter increments by Cnt. roll =Cnt roll +1.

[0057] Rate limiting termination rule: The system presets the maximum retry threshold Throll =3. When Cnt roll ≥Th roll When the current cognitive state is determined to be in a state requiring manual intervention, the automatic process is paused and an alarm signal is sent to the administrator's terminal. The automatic intervention issuance process is terminated, and the process is resumed or terminated after waiting for manual confirmation from the administrator.

[0058] 4. Three-tiered, step-by-step intervention and matching procedures The system uses the bias level Lv determined by the atomic scale of cognitive evidence. bias Following the hierarchical principle of "the more severe the deviation, the finer the guidance granularity, the higher the prompt level, and the stronger the intervention targeting," corresponding intervention strategies are matched to achieve differentiated and precise intervention: Level 1 Mild Deviation Intervention: For minor deviations such as occasional oversights and weak memorization of superficial knowledge points, only directional and inspirational hints are provided, along with problem-solving entry points, the scope of core knowledge points, and directions for thinking. Users are left to review and correct their own mistakes.

[0059] Level 2 Moderate Deviation Intervention: For moderate deviations such as missing logic, incomplete understanding of concepts, and gaps in step-by-step deduction, step-by-step guidance content is provided to break down the standard reasoning chain segment by segment, fill in the logical gaps in the user's logic in a targeted manner, and gradually guide the user to improve their own cognitive logic.

[0060] Level 3 Severe Deviation Intervention: For severe deviations such as confusion of core concepts, persistent thinking errors, and lack of underlying cognitive framework, it outputs standard deduction framework demonstration content, fully demonstrating the standardized reasoning logic and thinking architecture, and helps users completely reconstruct their underlying cognitive system.

[0061] 5. Detailed content of intervention tasks and answer masking operation. In the second interactive mode, to prevent the system from directly outputting standard answers and replacing users' independent thinking and deduction, the system performs a refined answer masking process throughout the entire process. The specific steps are as follows: Resource pre-construction: Build a subject-specific standard answer resource library in advance, including four types of sensitive feature data: conclusion keywords, final solution of formulas, closed deduction steps, and result statements.

[0062] Text semantic annotation: The BiLSTM sequence annotation model is used to perform character-by-character semantic annotation on the generated intervention text to accurately distinguish between open-ended heuristic statements and closed-ended conclusion statements.

[0063] Sensitive content processing: Accurately match sensitive feature segments in the text, and perform removal and masking replacement operations on complete problem-solving conclusions, final answers, and closed-loop deduction content.

[0064] Content output control: Only open-ended thinking guidance, logical questions, and segmented thought process prompts are retained to ensure that intervention only serves to inspire and guide, and to fully preserve the learning space for users to make their own deductions and think independently.

[0065] After the intervention task content is generated and masked, the system packages it into a standardized intervention task and pushes it to the front end, awaiting user interaction and response. Once the user completes the learning process, the system automatically switches to the verification state. This embodiment fully discloses all the underlying implementation details of state modeling, flow logic, counting and rate limiting, hierarchical matching, and content masking, with clearly quantified rules and clear and complete steps.

[0066] Step 3: Generate a verification task that is logically isomorphic but superficially heterogeneous for the cognitive state after intervention. Based on the verification results, determine the transfer of cognitive structure. Revert the cognitive state for which the corresponding cognitive structure has not been transferred to the intervention stage and re-output the step-by-step intervention task. Perform steady-state training on the transferred cognitive structure and generate the corresponding training task.

[0067] In this embodiment, a fixed cognitive logic framework is constructed based on the cognitive content that has been learned through intervention. A verification task with logical isomorphism and surface heterogeneity is generated. User response data is collected and cognitive transfer scores are quantitatively calculated. Based on the score results, process branch judgment and cognitive state regression are executed to achieve accurate verification and closed-loop correction of cognitive transfer effects.

[0068] In this embodiment, as Figures 1-3 After users complete the verification task, the system collects accuracy and response time data. The MRT cognitive imaging engine converts the verification task data and transfer judgment results into structured cognitive data packets, which are then processed by the system. Figure 2 The structured cognitive data stream shown is written into the cognitive digital twin carrier to achieve persistent storage of the verification results. This embodiment integrates the process control of the linkage system control layer with the verification computation capabilities of the engine layer, relying on the cognitive scheduling and control unit and the MRT cognitive imaging engine for collaborative implementation. Specifically, this embodiment presents the operations of constructing the cognitive logic skeleton, generating heterogeneous variant tasks, normalizing and preprocessing the response data, quantifying and scoring cognitive transfer, determining process branches, and implementing state rollback. The specific steps are as follows: 1. Construction of the core cognitive logic framework After the user completes the tiered intervention learning, the system first solidifies the core cognitive framework corresponding to this intervention, which serves as the sole benchmark for generating verification tasks and judging cognitive effects. This framework remains unchanged throughout the process to ensure that the core of the verification assessment does not deviate, and is formally defined as follows: Frame={Node,Edge} Among them, Node is the core reasoning node set, representing the key problem-solving steps, core operation units and reasoning links that must be mastered for this knowledge point; Edge is the knowledge point dependency edge set, representing the logical deduction relationship, causal relationship, operation constraint and knowledge point association rules between each reasoning node.

[0069] Conditional Linkage Update of Cognitive Logic Framework The system detects whether the evidence set fields of the current cognitive evidence atom contain supplementary semantic features obtained through additional interactive prompts (i.e., the response extraction results of the aforementioned three types of structured prompt texts). If included, the cognitive logic skeleton (Frame) composed of the inference node and concept dependency edge is updated based on the supplementary semantic feature, so that the subsequent variant template matching and substitution generation operations performed based on the skeleton can specifically verify the bias underlying mechanism label after the supplementary evidence correction.

[0070] 2. Generation of Logically Isomorphic and Surface Heterogeneous Verification Tasks Based on a fixed cognitive logic framework, the system performs dual heterogeneous transformation operations to generate verification questions with entirely new appearances but consistent core logic. This completely eliminates the interference of users' memorization through rote memorization and accurately detects their true cognitive mastery. Specifically, the transformation operations are divided into two categories: Contextual substitution transformation: Completely replace the background scene, text description sentence structure, question posing method and external expression form of the question, thoroughly change the appearance of the question, but do not add, delete or modify any core reasoning nodes and logical dependencies of knowledge points, ensuring that the core of the assessment remains unchanged.

[0071] Numerical substitution transformation: Under the premise of strictly preserving the original calculation formulas, reasoning logic, variable constraints, and knowledge point relationships, the original numerical values, variable value ranges, and parameter distribution ranges of the question are replaced to achieve heterogeneous updates of the numerical dimension.

[0072] 3. Normalization of response data and quantitative calculation of cognitive transfer After users complete heterogeneous verification tasks, the system simultaneously collects two core data points: step accuracy and response time. Through normalized preprocessing and a weighted fusion model, the system quantifies and calculates the cognitive transfer effect. The overall formula for cognitive transfer score: Score trans =w1·Acc+w2·Speed, Among them, Acc is the accuracy rate of the verification task steps, with a value range of [0,1]. It is calculated by comparing the system step by step with the standard reasoning logic and statistically analyzing the ratio of correct steps to total steps. Speed ​​is the normalized response speed feature. The weights w1=0.7 and w2=0.3 are the core weights that prioritize the accuracy of logical mastery while taking into account the proficiency of thinking.

[0073] Answering speed normalization: Since the original answering time is affected by the difficulty of the question type, the number of questions, and the complexity of the test points, it is not possible to directly compare them horizontally. The system uses the max-min normalization algorithm to achieve dimensionless standardization, as shown in the following formula: Speed ​​= (T) max -T raw )⁄(T max -T min ), Among them, T raw The actual time spent by the user answering a single question; T max T min This represents the historical maximum and minimum time extremes for questions of the same difficulty and type. After normalization, Speed∈[0,1], enabling unified calculation and comparison of speed indicators for different question types.

[0074] 4. Rigid branch determination and state rollback operation The system presets a fixed migration judgment threshold Th trans =0.7, based on the migration score, rigid and unambiguous process branch determination is performed: Migration failed branch: When Score trans If the score is less than 0.7, the user is deemed to have not achieved effective cognitive transfer, only grasping the surface-level answers to the questions and failing to understand the underlying reasoning logic. The system automatically triggers a state rollback mechanism, reverting the cognitive state from the verification state to the intervention execution state. Combining the characteristics of the current answer deviation with historical intervention records, a new round of adapted tiered intervention tasks is generated, while the rollback counter is simultaneously incremented.

[0075] Successful migration branch: When Score trans If the score is ≥0.7, it is determined that the user has a good grasp of the underlying reasoning logic, the cognitive bias has been effectively corrected, the cognitive structure has been stably transferred, and the system will automatically jump to the steady-state solidification training process.

[0076] This embodiment fully discloses the principle of variant construction, data standardization implementation, quantitative scoring model, and branch control logic. Each technical operation is detailed to the point of being directly implementable execution steps. Formal symbols are fully covered, parameter explanations are complete, and judgment logic is unambiguous, enabling precise quantitative verification and closed-loop correction of intervention effects.

[0077] Step 4: Write the cognitive evidence atoms, verification judgment results, and steady-state training results into the cognitive digital twin carrier through a state-stream writing architecture, and output hierarchical decision signals after updating with time-series cognitive snapshots.

[0078] In this embodiment, a progressive steady-state training sequence is constructed based on the transferred cognitive structure, the cognitive stability index is quantitatively calculated to determine the degree of cognitive solidification, full-process data retention is achieved through streaming time-series writing, time-series cognitive snapshots are dynamically updated and hierarchical teaching decisions are output, and the long-term solidification and closed-loop management of cognitive structure are achieved in conjunction with full-process control.

[0079] In this embodiment, after the steady-state training process is completed, the system relies on Figure 2 The cognitive digital twin carrier shown completes the full-volume cognitive data streaming time-series writing and time-series snapshot aggregation and updating, forming a closed-loop backflow of historical cognitive data through the data callback interface, while relying on... Figure 3 The trust layer enables end-to-end operational auditing, traceability, and control. This embodiment details the operations of progressive steady-state training sequence construction, cognitive stability quantitative assessment, full-data streaming time-series storage, dynamic snapshot aggregation and updating, hierarchical teaching decision output, and end-to-end controllable management. The specific step-by-step operations are as follows: 1. Construction of progressive steady-state training sequences For cognitive structures that have undergone effective transfer, the system follows the training principle of "fixed core logic and progressively increasing reasoning complexity," constructing a steady-state training sequence with gentle gradients to avoid cognitive regression and knowledge forgetting. Fixed core logic: The core reasoning nodes of the already mastered cognitive structure are retained throughout the process to ensure that the training content is consistent with the user's existing cognitive foundation and avoid learning gaps caused by abrupt changes.

[0080] Gradually increase complexity: Slightly increase the number of knowledge point dependencies in each question, slightly extend the length of the reasoning chain, and slowly add related knowledge points to gradually increase the difficulty of training, allowing users to consolidate and deepen their understanding on the basis of mastery.

[0081] Training data collection: During the training process, the accuracy rate (Acc) of each question is collected. t The original response time was calculated, and the original time was normalized to obtain the standardized time delay feature T. norm,t This provides data support for stability assessment.

[0082] 2. Quantitative assessment calculation of cognitive stability The system constructs a proprietary quantitative assessment formula for cognitive stability to accurately determine the degree of solidification of a user's cognitive structure and distinguish between short-term memory and long-term stable cognitive abilities. The formula is as follows: , Where: N is the total number of questions in this round of steady-state training; Acc t The correct answer rate for question t; To improve the average accuracy during training, positively represent the stability of cognitive mastery; T norm,tThe normalized response time for a single question; λ=0.2 is the balance coefficient, used to weaken the random interference of lag and momentary errors in answering, and to strengthen the weight of accuracy in judging steady-state ability; To train the average latency feature, we can inversely represent the fluency of thought and cognitive proficiency.

[0083] Steady-state curing determination rule: The system presets a steady-state curing threshold Th. stable =0.75. When Index stable If the value is ≥0.75, it is determined that the user's cognitive structure has moved beyond short-term memory attributes and formed a long-term stable reasoning ability and knowledge system, and the current round of steady-state training will automatically terminate; if the indicator does not meet the standard, the system will continue to iterate and train until the steady-state solidification condition is met.

[0084] 3. Full-scale cognitive data streaming time-series storage: After completing the entire cognitive processing process, the system initiates time-series data encapsulation and streaming write operations to achieve complete retention of the cognitive evolution trajectory: Data encapsulation: Cognitive evidence atoms, tiered intervention records, cognitive transfer scores, steady-state training indicators, variant verification data, and full-process state transition records are uniformly encapsulated into time-series event records carrying unique timestamps, version numbers, and user identifiers.

[0085] Incremental writing: The backend incremental append writing mode is adopted. All newly generated data is appended to the backend for storage. No historical data is overwritten, modified, or deleted. The entire life cycle evolution trajectory of cognition from the generation of deviation, diagnosis, intervention, repair to solidification is fully preserved.

[0086] 4. Dynamic time-series snapshot aggregation and hierarchical decision output: The system supports snapshot aggregation updates within custom time windows, enabling dynamic review of cognitive states and precise teaching decisions. Snapshot aggregation operation: Extract all time-series event records within a user-specified start and end time window, perform feature fusion, indicator aggregation, and trajectory fitting in chronological order, and dynamically generate the latest time-series cognitive snapshot.

[0087] Tiered decision output: Based on the cognitive state, deviation correction progress, and training consolidation effect integrated by snapshots, the system adaptively outputs tiered teaching decisions, including differentiated control strategies such as continuous consolidation training, intervention gradient downsampling, cognitive topic switching, and specialized enrichment training, to achieve data-driven precision teaching.

[0088] 5. Full-process controllable management and control mechanism: The system implements standardized control over the entire cognitive processing operation, ensuring that the process is standardized, controllable, and traceable. Full-process traceability: For all operations such as hierarchical cognitive analysis, cognitive state transition, intervention task issuance, verification task execution, steady-state training iteration, and data streaming writing, log records with unique timestamps are generated to achieve full-process traceability of behavior.

[0089] Manual intervention and control: Before all tiered intervention tasks are officially issued, a manual confirmation instruction must be received. If no valid confirmation is received, the issuance of the task will be automatically terminated. This achieves an organic combination of intelligent algorithm scheduling and manual teaching control, avoiding unreasonable algorithmic intervention.

[0090] This embodiment fully completes the steady-state training progression rules, the stability index quantification principle, the streaming storage writing mechanism, and the time-series snapshot aggregation logic. All formulas, parameters, operations, and hierarchical principles are fully elaborated, and the technical implementation details are complete and sufficient, ultimately achieving a closed-loop implementation of the entire chain from cognitive diagnosis to intervention, verification, consolidation, storage, and decision-making.

[0091] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A cognitive evolution reconstruction method based on cognitive digital twins, characterized in that, include: Acquire multimodal cognitive data of the target object and perform hierarchical cognitive imaging processing. Use the analysis results of the previous level as the input of the next level to gradually narrow down the scope of cognitive bias localization, lock the underlying mechanism label of the bias and generate cognitive evidence atoms. Using the cognitive evidence atom as the trigger event input, the cognitive state machine is subjected to state transition determination, and a step-by-step intervention task is output. Based on the cognitive state after intervention, a logically isomorphic but superficially heterogeneous verification task is generated and a cognitive structure transfer judgment is made. Cognitive states that have not been transferred are regressed and a step-by-step intervention task is re-output. Stabilization training is performed on the transferred cognitive structures, and corresponding training tasks are generated; The cognitive evidence atoms, verification and judgment results, and steady-state training results are written into the cognitive digital twin carrier through a state-stream writing architecture, and hierarchical decision signals are output.

2. The method according to claim 1, characterized in that, The hierarchical cognitive imaging process includes five levels: The result layer determines the correctness of answers based on the answer data of the cognitive task, generates a result layer feature vector, and marks abnormal answers. The path layer uses abnormal answer items as retrieval anchors to locate the corresponding deviation points in the problem-solving step sequence and generate a path layer deviation vector. The expression layer focuses on the steps corresponding to the deviation points, performs semantic analysis on the text and speech data of the target object within the focus area, and generates expression layer abnormal features. The behavioral layer uses the steps corresponding to the deviation point as a time window to perform behavioral pattern analysis on the interaction sequence and handwriting trajectory data of the target object within the window, and determines the type of cognitive behavioral inertia. The evolution layer performs joint temporal modeling based on the outputs of the result layer, path layer, expression layer, and behavior layer, and outputs the underlying mechanism label of the deviation.

3. The method according to claim 1, characterized in that, The generation process of the cognitive evidence atoms includes: The underlying mechanism label of the deviation is used as the deviation type field, and the output results and deviation point sequence of each level are used as the structured evidence chain of the evidence set field. The deviation type field is mapped to the predefined intervention protocol identifier. The deviation type field, structured evidence chain and intervention protocol identifier are encapsulated into cognitive evidence atoms.

4. The method according to claim 3, characterized in that, During the execution of each level of the hierarchical cognitive imaging process, the confidence level of the output results of each level is monitored. When the confidence level is lower than a preset threshold, a forced interaction interruption signal is triggered. The structured prompt text is extracted from the preset prompt template library, and the corresponding semantic features are obtained as supplementary evidence and written back to the evidence set field. The locking operation of the bias underlying mechanism label is performed to generate updated cognitive evidence atoms.

5. The method according to claim 2, characterized in that, The path layer includes: Obtain the sequence of steps and nodes of the standard cognitive path, and perform an alignment operation between the sequence of problem-solving steps and the sequence of steps and nodes; mark the steps and nodes that do not match in the alignment operation as deviation points, and assign a deviation type label to each deviation point; output the sequence of deviation points containing the deviation type label as the path layer deviation vector.

6. The method according to claim 1, characterized in that, The state transition determination of the cognitive state machine includes: The feature vector of the current cognitive state is compared with the trigger threshold of each condition in the preset transition condition set; when any trigger threshold is met, the current cognitive state is transferred to the next cognitive stage corresponding to the trigger threshold.

7. The method according to claim 6, characterized in that, The state transition determination also includes: In response to a verification failure event triggering a rollback to the intervention execution stage, the number of rollbacks is recorded. After the intervention task is re-output, the verification task generation is executed again. When the cognitive structure migration determination result is determined to be non-migrated again and the number of rollbacks reaches the preset upper limit, the current cognitive state is marked as pending manual intervention, the automatic flow is paused, and an alarm signal is sent to the administrator terminal.

8. The method according to claim 1, characterized in that, The process of generating the verification task includes: Extract inference nodes and concept dependency edges from the original cognitive task corresponding to the cognitive state after intervention, and use inference nodes and concept dependency edges to form a cognitive logic skeleton; when the cognitive evidence atom contains supplementary semantic features obtained through additional interactive prompts, update the cognitive logic skeleton formed by inference nodes and concept dependency edges; select a target template that matches the cognitive logic skeleton from a predefined variant template library, perform replacement on contextual elements and numerical representations, and generate a verification task.

9. The method according to claim 1, characterized in that, The generation process of the tiered intervention task includes: Based on the severity level of the bias in the cognitive evidence atoms, a corresponding guidance level is matched. The guidance level includes a directional prompt level, a step-by-step guidance level, and a demonstration and deduction level. Each level is associated with a directional prompt template, a step-by-step guidance template, and a demonstration and deduction template, respectively. Tiered intervention content is generated according to the template corresponding to the matched guidance level.

10. The method according to claim 1, characterized in that, The steady-state training includes: A sequence of inference tasks is generated based on the inference nodes and concept dependency edges corresponding to the transferred cognitive structure. The inference tasks are pushed to the target object in sequence, and the completion accuracy and response latency of each inference task are collected. The cognitive stability index is updated based on the completion accuracy and response latency. When the cognitive stability index reaches the preset solidification threshold, the transferred cognitive structure is marked as a steady-state cognitive structure.

11. The method according to claim 1, characterized in that, When the state-stream writing architecture performs a write operation, it encapsulates the cognitive evidence atom, the verification judgment result and the steady-state training result into an event record, and stores them in the cognitive digital twin carrier in chronological order by appending the write operation. In response to the temporal cognitive snapshot update request, all event records within the time window corresponding to the current cognitive state are read and merged in chronological order to generate a temporal cognitive snapshot.

12. A cognitive evolution reconstruction system based on cognitive digital twins, comprising a processor for performing the method as described in any one of claims 1 to 11.

Images