AI-based dynamic adaptation system for science education for teenagers in multiple disaster scenarios

Abstract

The present invention relates to the fields of artificial intelligence and disaster knowledge popularization, and discloses an artificial intelligence-based dynamic adaptation system for youth science popularization education in multiple disaster scenarios. The system comprises: a data acquisition module for constructing a multi-dimensional disaster knowledge base and collecting student behavior data in real time; an ability assessment module for enhancing historical state memory capabilities through an exponential gating mechanism based on the student behavior data, combining stable gradient updates of normalized state variables, and dynamically outputting an evaluation vector containing question type adaptation weights, content level scores, and comprehensive ability values; a scenario adaptation module for establishing a state transition model and an agent reward function through a combination of dictionary functions to generate an optimal adaptation strategy that matches the student state; and a feedback execution module for dynamically adjusting task parameters based on the evaluation results of the comprehensive ability value and the optimal adaptation strategy, and providing feedback on protective operation guidance in multimedia form. The present invention improves the accuracy of disaster education content and the intelligence of teaching methods.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and disaster knowledge popularization technology, and in particular to a dynamic adaptation system for popular science education for teenagers in multiple disaster scenarios based on artificial intelligence. Background Technology

[0002] Traditional disaster education primarily relies on static courseware and offline lectures, lacking interactivity, real-time relevance, and adaptability, making it difficult to meet the diverse and dynamic learning needs of teenagers. Especially when dealing with individuals with different cognitive abilities, learning motivations, and behavioral styles, traditional educational methods often exhibit problems such as "fixed content," "monotonous pace," and "delayed feedback."

[0003] Meanwhile, the integration of artificial intelligence and educational game technology offers a new path for disaster science education. By introducing adaptive mechanisms driven by behavioral data, personalized delivery of disaster knowledge content and dynamic adjustment of teaching tasks can be achieved. However, current mainstream personalized recommendation systems generally lack in-depth modeling of the complexity of disaster scenarios and the temporal nature of learning behaviors, failing to effectively capture changes in students' ability evolution, task selection preferences, and emotional feedback. Furthermore, in disaster-themed game scenarios, due to practical constraints such as unobservable rewards, strong environmental uncertainty, and limited sample size, traditional reinforcement learning strategies often exhibit slow convergence and poor generalization, making it difficult to form deployable and efficient control strategies.

[0004] Therefore, there is an urgent need for an intelligent adaptation system that integrates high-dimensional temporal modeling and reinforcement learning optimization. This system can assess students' abilities in real time based on their behavioral data in multi-hazard simulation environments and automatically match disaster scenarios, content difficulty, and task formats, thereby building a more immersive, interactive, and educational disaster science popularization platform for teenagers. Summary of the Invention

[0005] To address the aforementioned issues, this invention aims to provide a dynamic adaptation system for science education for teenagers in multiple disaster scenarios based on artificial intelligence, thereby improving the accuracy of disaster education content and the intelligence of teaching methods.

[0006] To achieve the above objectives, the following technical solution is adopted:

[0007] A dynamic adaptation system for science education for teenagers in multiple disaster scenarios based on artificial intelligence includes:

[0008] The data acquisition module is used to build a multi-dimensional disaster knowledge base and collect student behavior data in real time through a 3D disaster simulation engine. The disaster knowledge base covers a classification system of natural disasters, man-made disasters and environmental disasters. The student behavior data includes answer records, operation paths and interactive feedback.

[0009] The ability assessment module, based on the student behavior data, enhances the historical state memory ability through an index gating mechanism, combines normalized state variable stable gradient updates, and dynamically outputs an assessment vector containing question type adaptation weights, content level scores, and comprehensive ability values.

[0010] The scene adaptation module integrates a sparse dynamic recognition reinforcement learning unit, which establishes a state transition model and a proxy reward function through dictionary function combination to generate the optimal adaptation strategy that matches the student's state.

[0011] The feedback execution module dynamically adjusts task parameters based on the comprehensive capability value evaluation results and the optimal adaptation strategy, and provides protection operation guidance through multimedia.

[0012] Furthermore, the capability assessment module employs a P-sLSTM model and includes:

[0013] The preprocessing unit is used to standardize and reconstruct the original student behavior data and disaster knowledge graph to obtain standardized time-series data;

[0014] The block-based temporal processing unit is used to divide the standardized temporal data into several fixed-length sub-block sequences. Each sub-block independently extracts local temporal features and then linearly fuses global information to output a sub-block feature embedding vector.

[0015] The channel-independent modeling unit is used to decouple the sub-block feature embedding vectors, separate the M feature channels into independent data streams, and process each channel in parallel through independent LSTM modules to output the reconstructed channel-decoupled temporal state matrix.

[0016] The exponential gating unit introduces an exponential gating mechanism into the memory unit. By expanding the numerical range of the forget gate and the input gate through an exponential function, it enhances the ability to retain remote historical states, updates the cell state, and outputs the updated cell state.

[0017] The state normalization unit is configured to introduce normalized state variables to balance cell state updates. It calculates normalized variables based on the updated cell state and the historical normalized state to obtain gradient-stable hidden state output.

[0018] The memory fusion unit is configured to fuse capability features at different temporal granularities through a multi-head attention mechanism and a block diagonal cyclic weight matrix to generate a multi-scale fused state vector.

[0019] The output decoding unit is configured to generate an evaluation vector that includes dynamic output, question type adaptation weights, content level scores, and comprehensive ability values, based on the multi-scale fusion state vector.

[0020] Furthermore, the exponential gating mechanism is specifically manifested as follows:

[0021] The activation function uses an exponential function, which expands the output range of the forget gate and the input gate to the positive real number range;

[0022] The retention strength of historical ability states is controlled by adjusting the forget gate weight matrix, and the update magnitude of the ability assessment by the current behavior data is controlled by the input gate weight matrix.

[0023] The output gate is used to explicitly adjust the output of the normalized hidden state, forming an encoding vector that contains the representation of the current capability.

[0024] Furthermore, the scene adaptation module includes:

[0025] The dynamic environment modeling unit, based on an improved sparse recognition algorithm, fits the transition relationship between student state and task action through dictionary function combination to construct a discrete-time dynamic model;

[0026] The sparse reward modeling unit constructs an interpretable surrogate reward function based on the student's state vector and the task action vector recommended by the system using a sparse regression algorithm.

[0027] The policy compression unit approximates the deep reinforcement learning policy in a sparse form and transforms it into an explicit functional expression, outputting a lightweight policy expression.

[0028] The strategy optimization unit uses a proximal policy optimization algorithm to maximize the expected cumulative agent reward. Through a two-stage iterative mechanism of pre-training in a simulated environment and fine-tuning in a real environment, it outputs a probability distribution mapping from student state to task action.

[0029] Furthermore, the dynamic environment modeling unit includes:

[0030] Based on historical student state-action pair data, a state transition model is established by combining nonlinear dictionary functions, and a discrete-time modeling method is used to predict the evolution of student state after task decision.

[0031] The dictionary function contains a combination of polynomial and trigonometric function features.

[0032] Furthermore, the sparse reward modeling unit includes:

[0033] The sequential threshold least squares algorithm is used to screen key influencing factors from the redundant feature dictionary, and L2 regularization constraints are added to improve generalization ability.

[0034] Furthermore, the policy compression unit introduces a noisy trajectory sampling mechanism, which enhances policy robustness by adding Gaussian noise to simulate environmental uncertainties.

[0035] Furthermore, the dynamic environment modeling unit and the sparse reward modeling unit share the same sparse dictionary function set;

[0036] Among them, the state transition function With agent reward function The sparse coefficient matrix is ​​jointly optimized using a block coordinate descent algorithm to improve model consistency.

[0037] Furthermore, the data acquisition module performs the following:

[0038] Knowledge base construction units, in collaboration with medical institutions, supplement medical disaster prevention knowledge to form an interdisciplinary knowledge graph;

[0039] The behavioral analysis unit identifies students' cognitive levels and motivation types through the KAP questionnaire and qualitative interviews, and establishes a "disaster knowledge-cognitive ability" mapping database.

[0040] Furthermore, the feedback execution module is used to implement:

[0041] The ability score output by the ability assessment module triggers immediate corrective feedback; wherein, the immediate corrective feedback includes textual guidance for incorrect operations and multimedia demonstrations of correct operations;

[0042] Mapping the competency assessment results to the three-dimensional decision elements of the COM-B model: competency performance, environmental opportunities, and motivational drivers;

[0043] A virtual-real scene linkage task is generated based on the three-dimensional decision elements; wherein, the virtual-real scene linkage task is associated with a geographic information system database, and real disaster avoidance points are mapped as virtual navigation targets;

[0044] Furthermore, the feedback execution module integrates medical protection guidelines as interactive knowledge points, embedding clinical operation standards in scenarios such as radiation leakage and heatstroke.

[0045] Compared with the prior art, the present invention achieves the following beneficial effects:

[0046] 1. This invention proposes a block-enhanced structured long short-term memory network (P-sLSTM) model: by introducing an exponential gating mechanism and normalized state variables, it solves the memory decay and gradient instability problems of traditional LSTM in long sequence modeling, realizes high-fidelity modeling of the changing trends of key abilities in the behavioral sequences of adolescent students, and enhances the recognition accuracy of individual ability states.

[0047] 2. This invention constructs a block structure and structured memory fusion mechanism: it proposes to divide the student behavior history sequence into fixed-length sub-blocks, and integrate global features through linear fusion based on local modeling, thereby alleviating the impact of behavior sequence sparsity on prediction ability and improving the generalizability and stability of the model.

[0048] 3. This invention designs a sparse reinforcement learning adaptation model for disaster tasks: combining sparse function dictionary modeling and reinforcement learning optimization algorithms, a sparse dynamic recognition reinforcement learning system is constructed to learn state transition rules and preference feedback mechanisms from student behavior trajectories, thereby achieving accurate and personalized allocation of disaster scenario tasks.

[0049] 4. This invention proposes an improved sparse identification algorithm: used to establish a state transition model in the disaster education interaction process. Under low sample conditions, it achieves efficient modeling of system behavior dynamics by combining dictionary functions, thereby improving the model's adaptability to complex nonlinear disaster behavior.

[0050] 5. This invention establishes a sparse agent reward function model: by analyzing students' task completion, collaborative behavior and feedback attitude, an agent reward function is constructed to solve the problem of "reward signals being difficult to observe directly" in disaster education, thereby enhancing the applicability and interpretability of reinforcement learning in educational scenarios.

[0051] 6. This invention introduces sparse regression algorithms such as sequential threshold least squares method: it automatically selects key influencing factors from redundant feature dictionaries, establishes a concise and stable reward function expression structure, and significantly improves the model's ability to be generalized to new students and new scenarios.

[0052] 7. The present invention constructs a sparse policy approximation module: the trained deep reinforcement learning policy is compressed into an explicit control form through sparse function expressions, reducing model complexity and deployment costs, so that the system can run on resource-constrained mobile devices or embedded science popularization devices.

[0053] In summary, this invention combines advanced deep learning, reinforcement learning, and sparse modeling methods to form a complete intelligent disaster education technology framework, from student ability identification and dynamic content allocation to strategy generation and interpretable output, effectively improving the disaster response capabilities and scientific literacy of young people.

[0054] It should be understood that the description in the Summary of the Invention is not intended to limit the key or essential features of the embodiments of the present invention, nor is it intended to restrict the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0055] The above and other features, advantages, and aspects of the various embodiments of the present invention will become more apparent from the accompanying drawings and the following detailed description. The drawings are provided for a better understanding of the invention and are not intended to limit the invention. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein:

[0056] Figure 1 This is a schematic diagram of a module of a dynamic adaptation system for science education for teenagers in multiple disaster scenarios based on artificial intelligence, according to an embodiment of the present invention.

[0057] Figure 2 This is a schematic diagram of the capability assessment module according to an embodiment of the present invention;

[0058] Figure 3 This is a schematic diagram of the scene adaptation module in an embodiment of the present invention. Detailed Implementation

[0059] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0060] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0061] Figure 1 A schematic diagram of a module for a dynamic adaptation system for science education for teenagers in multiple disaster scenarios based on artificial intelligence is shown. Figure 1 As shown, a dynamic adaptation system 100 for science education for teenagers in multiple disaster scenarios based on artificial intelligence includes:

[0062] The data acquisition module 110 is used to build a multi-dimensional disaster knowledge base and collect student behavior data in real time through a 3D disaster simulation engine. The disaster knowledge base covers the classification system of natural disasters, man-made disasters and environmental disasters. The student behavior data includes answer records, operation paths and interactive feedback.

[0063] Furthermore, the data acquisition module 110 includes:

[0064] Knowledge base construction unit 111, in conjunction with medical institutions, supplements medical disaster prevention knowledge and forms an interdisciplinary knowledge graph;

[0065] Based on various common disaster types encountered in daily life, such as natural disasters (e.g., earthquakes, floods, typhoons), man-made disasters (e.g., fires, traffic accidents), and environmental disasters (e.g., pollution, radiation leaks), and to enhance professionalism and practicality, this project collaborates with expert teams from medical institutions such as Sichuan Cancer Hospital to supplement medical-related disaster prevention knowledge, including modules on emergency response to heat waves and radiation protection measures. This forms a comprehensive disaster knowledge base covering multiple dimensions and interdisciplinary fields. The knowledge base not only includes disaster principles but also covers emergency procedures, preventative measures, and rehabilitation guidance, ensuring that the content is scientific, authoritative, and closely related to reality.

[0066] Behavioral analysis unit 112 identifies students' cognitive levels and motivation types through the KAP questionnaire and qualitative interviews, and establishes a "disaster knowledge-cognitive ability" mapping database.

[0067] Specifically, the behavioral analysis unit 112 is used to collect data on adolescents' disaster cognition and needs. Through the design and implementation of the KAP (Knowledge, Attitude, Practice) questionnaire in the accompanying research project, the system collects adolescents' cognitive levels, interests, and weaknesses under different disaster types. Combined with qualitative interviews, it delves into adolescents' habits of accepting disaster knowledge, their learning motivations, and emotional responses, ensuring the diversity and authenticity of the data. After cleaning and labeling, the collected data is used to construct a "Disaster Knowledge-Cognitive Ability" database, providing foundational data support for subsequent intelligent adaptation algorithms and ensuring the scientific validity and accuracy of the system's personalized recommendations.

[0068] Before designing the model, the design of dynamically adaptable dimensions is also included:

[0069] (1) Difficulty adaptation dimension design

[0070] Based on students' behavioral data in interactive games, such as accuracy, response time, and number of repeated errors, the system dynamically adjusts the difficulty of each level, progressively increasing the level's complexity. For example, if students consistently grasp the basic knowledge of "high-temperature symptom identification," the system automatically increases the difficulty, recommending more complex tasks such as "designing a family heatstroke prevention plan," thus enhancing the challenge and depth of learning. This mechanism ensures that students learn within a difficulty range suitable for their abilities, avoiding frustration from overly difficult tasks or boredom from overly easy ones.

[0071] (2) Content Adaptation Dimension Design

[0072] By combining the COM-B behavioral model (Capability, Motivation, Opportunity), we analyze students' motivation types to identify different needs such as achievement-driven, social-driven, and exploration-driven motivation. Based on the analysis results, we personalize the delivery of disaster scenarios and tasks that match students' interests and behavioral habits. For example, for students who prefer teamwork, we prioritize assigning multi-person team disaster relief tasks to promote interaction and collaborative learning; for students who prefer individual challenges, we provide strategy questions or simulated operation questions to satisfy their needs for achievement and self-improvement.

[0073] The ability assessment module 120, based on the student's operational behavior, enhances the ability to remember historical states through an exponential gating mechanism, combines normalized state variables to stabilize gradient updates, and dynamically outputs an assessment vector that includes question type adaptation weights, content level scores, and comprehensive ability values.

[0074] The core task of building a student ability assessment model is to dynamically analyze students' behavioral data during interactions, thereby inferring their current knowledge mastery, strategy application ability, cognitive load level, and other comprehensive ability indicators. Based on this, the system can intelligently adjust the structure and difficulty of subsequent content to achieve "personalized instruction."

[0075] The ability assessment module 120 aims to dynamically assess and predict students' ability levels in science popularization education in multiple disaster scenarios by constructing an artificial intelligence model based on a patch-enhanced structured long short-term memory network (P-sLSTM). By introducing exponential gating mechanisms and normalized states, it enhances the capture of key ability cues in behavioral sequences and the memory of historical ability accumulation information.

[0076] The model input data includes: (1) Student historical answer record data: including answer accuracy rate, answer time, distribution of incorrect question types and other behavioral characteristics; (2) Difficulty level classification of content in the disaster knowledge graph: the disaster knowledge points are structured into multi-level difficulty levels (such as "heat wave protection" being divided into three levels: "basic cognition - emergency operation - scenario decision-making").

[0077] Furthermore, this invention employs an improved LSTM model—P-sLSTM. The P-sLSTM model mainly includes key steps such as block processing, channel-independent LSTM / exponential gating, state normalization, and multi-head attention fusion, which are used to enhance the ability to retain long-term dependent information in temporal modeling and adapt to the ability change trajectory of adolescents during the game process. Figure 2 This is a structural schematic diagram of the capability assessment module according to an embodiment of the present invention, as shown below. Figure 2As shown, the competency assessment module 120 includes:

[0078] The preprocessing unit 121 is used to standardize and reconstruct the original student behavior data and disaster knowledge graph to obtain standardized time-series data.

[0079] The aforementioned student historical behavior sequences (answer accuracy / time consumption / error distribution) and disaster knowledge graphs (structured difficulty levels) are standardized (Z-score normalization) and reconstructed into time series data (original dimension → B×L×M) before being input into the model. This results in standardized time series data (dimension: B×L×M) that meets the requirements of time series modeling. Here, B represents the batch size, i.e., the number of samples processed in a single training iteration; L represents the time step (e.g., 96 steps of historical data), i.e., the length of the input sequence; and M represents the number of feature channels, i.e., the variable dimension of the multivariate time series (e.g., question type, knowledge point, etc.).

[0080] The block-based temporal processing unit 122 is used to divide the standardized temporal data into several fixed-length sub-block sequences. Each sub-block independently extracts local temporal features and then linearly fuses global information to output a sub-block feature embedding vector.

[0081] A block-based mechanism is constructed: combining block-based mechanisms with LSTM, specifically dividing the student's historical behavior sequence into several fixed-length sub-blocks along the time dimension. Each sub-block independently extracts local features (each sub-block independently learns short-term behavior patterns), and then global information is aggregated through linear fusion. This mechanism, through three steps—sub-block segmentation, local feature extraction, and global fusion—effectively alleviates the information dilution problem that occurs in long-term series modeling.

[0082] The long sequence is divided into N fixed-length sub-blocks (block length P), and the input data is from the original dimension.

[0083] Reconstructed After feature extraction from sub-blocks, they are mapped to the embedding space and combined with LSTM for sequence modeling. N: Number of blocks, i.e., the number of sub-blocks after segmenting a long sequence. Each "sample" (a channel of the original sample) is divided into N consecutive subsequences (blocks). P: block length, time step of each subsequence (e.g., ...). This indicates that each sub-block contains 16 time steps. When L=96 and P=16, then N=6 blocks are calculated.

[0084] To process multi-channel (M feature channels) time series data (L time steps) using a block-based strategy (each block length P), and to allow for independent modeling of each channel's data, the original data X is reconstructed from dimensions [B, L, M] to [B×M, N, P]:

[0085]

[0086] in, This indicates that the original input data X (dimension [B, L, M]) is reconstructed into a three-dimensional tensor [B×M, N, P]. The dimension B×M means that the batch dimension B and the feature channel dimension M are merged into a new batch dimension, that is, B×M independent processing units are formed, which is equivalent to treating the data of M channels of B samples as B×M independent "samples".

[0087] The channel-independent modeling unit 123 is used to decouple the sub-block feature embedding vector ((B·M)×N×P) into channels, separating the M feature channels into independent data streams. Each channel is processed in parallel by an independent LSTM module, and the reconstructed channel-decoupled temporal state matrix is ​​output.

[0088] Assuming batch size B=32, number of channels M=5 (e.g., answer accuracy / time consumption / error rate / number of retries / knowledge point relevance), and block size N=6; independent channels: the 5 channels are separated into 160 (32×5) independent processing units.

[0089] Constructing a channel-independent mechanism: For multivariate time series (such as answer behavior data with multiple question types and knowledge points), each channel is modeled independently to avoid information interference between channels and reduce the risk of parameter coupling and overfitting. During model training, the original data (independent input of multi-channel data) is transformed... In terms of dimensions, each channel is processed independently through a block and an LSTM module, outputting a reconstructed time-series state matrix, which is the reconstructed channel-decoupled time-series state matrix. T: Prediction length, such as the predicted value for the next 24 time steps.

[0090] Input Restructuring:

[0091] Among them, permute(0,3,1,2): dimensional reordering [B,M,N,P]; reshape(B×M,N,P): channels are merged into batch dimensions.

[0092] Output reconstruction:

[0093] After channel-independent LSTM processing ( After feature compression, data with dimensions [B×M, T] or [B×M, T, H] (where T is the prediction time step) is obtained. First, the independent dimensions of batch B and channel M are restored, resulting in [B, M, T] (or [B, M, T, H]). Then, the data is adjusted to the standard time-series input format [B, T, M] (or [B, T, M, H]), i.e., (batch, time step, channel), through a dimension permute operation, making the time step T the primary sequence dimension for subsequent processing (such as memory mixing units).

[0094] By using a block-based mechanism to reduce sequence complexity, and by modeling channels independently to reduce interference, we can collaboratively address the problem of sparse behavioral data (such as low-frequency error patterns).

[0095] The exponential gating unit 124 introduces an exponential gating mechanism into the memory unit. By expanding the numerical range of the forget gate and input gate through an exponential function, it enhances the ability to retain remote historical states, updates the cell state, and outputs the updated cell state. .

[0096] Furthermore, the exponential gating mechanism is specifically manifested as follows: the forget gate in traditional LSTM is used... Input gate The activation function is replaced with an exponential function, which expands the output range of the forget gate and input gate to the positive real number range. Theoretically, this enhances the model's ability to retain key historical states. The retention strength of historical ability states is controlled by adjusting the forget gate weight matrix, while the update magnitude of the ability assessment based on current behavioral data is controlled by the input gate weight matrix.

[0097]

[0098] in, The forget gate output controls the degree to which historical student abilities (such as past performance) are retained at the current moment; a larger value indicates that more historical abilities are retained, and the value range is [value missing]. ; The input gate output controls the strength of how well the current observation information (such as current competition performance) updates the student's ability status; a larger value indicates stronger new information input, and the value range is [value range missing]. ;

[0099] The pre-activation value of the forget gate represents the original input signal of the forget gate, combined with the current game input. and the hidden state of the previous moment The formula representing the dynamic impact of students' past abilities is:

[0100]

[0101] The forget gate input weight matrix is ​​used to map the current game input. To the Forgotten Gate space; depicting the impact of current match performance on the degree to which historical abilities are retained. Forget gate loop weight matrix, used to map the previous hidden state. Enter the forgetting gate space; depict the impact of historical ability states on current forgetting decisions. The forget gate bias vector is used to adjust the baseline value of the forget gate output. For input vectors (Input of the current time step).

[0102] The input gate activation value represents the original input signal of the input gate, used to guide the impact of the current game input on the update of ability status. The formula is:

[0103]

[0104] The input gate input weight matrix is ​​used to map the current game input. Enter the input gate space; depict the impact of current game performance on ability updates. The input gate loop weight matrix is ​​used to map the previous hidden state. Enter the input gate space; characterize the impact of historical capability states on current update decisions. The input gate bias vector is used to adjust the baseline value of the input gate output.

[0105] In this structure, the exponential gating mechanism extends the values ​​of the forget gate and the input gate to the range of positive real numbers, thereby enhancing the ability to remember distant historical states.

[0106] Simultaneously, normalized state variables are introduced. Enhance hidden state Numerical stability during the update process to prevent gradient explosion or vanishing.

[0107] State normalization unit 125 is configured to introduce normalized state variables to balance cell state updates, based on the updated cell state. and historical normalization state Calculate the normalized variables to obtain the gradient-stable hidden state output.

[0108] Introducing normalized state variables Enhance hidden state Numerical stability during the update process is ensured to prevent gradient explosion or vanishing. An output gate is used to explicitly adjust the output of the normalized hidden state, forming an encoding vector that represents the current capability.

[0109]

[0110] Normalized states are used to standardize hidden states. The calculation prevents numerical explosion or vanishing of student ability state gradients in long-term series, reflects the integration of historical and current information, and uses initial values. It can be 0 or 1. The normalized state at the previous moment represents the influence of history on the standardization of the current capability state. The hidden state represents the student's current performance level, combining the current input, historical state, and output gate adjustment; ε is a smoothing term. The output gate, through the sigmoid activation function, determines the degree to which the normalized ability state is explicitly displayed at the current moment, similar to the degree of "display" of the current ability performance. The formula is:

[0111]

[0112] in, The output gate input weight matrix is ​​used to map the current game input to the output gate space. The output gate loop weight matrix is ​​used to map the historical hidden states to the output gate space. The output gate bias vector is used to adjust the baseline response of the output gate. Cellular state, representing the student's ability state in long-term memory, is updated using a combination of forgetting and input gate mechanisms. The formula is:

[0113]

[0114] The cell state at the previous moment, that is, the student's accumulated ability over time. Candidate cell state is determined by the current input and historical state. The activation function is generated to represent the possible trends in ability changes; the formula is:

[0115]

[0116] in, : Candidate state input weight matrix, used to map the current game input to the candidate state space. : Candidate state cyclic weight matrix, used to map historical states to candidate state space. : Candidate state bias vector, used to adjust the baseline of the candidate state.

[0117] Normalized state The introduction of this effectively stabilized the hidden state. The gradient changes avoid unstable learning phenomena caused by fluctuations in ability and improve the model's modeling effect on medium- and long-term behavioral trends.

[0118] The memory mixing unit 126 is configured to fuse capability features at different temporal granularities through a multi-head attention mechanism and a block diagonal cyclic weight matrix to generate a multi-scale fused state vector.

[0119] Constructing a memory fusion mechanism: Combining a multi-head mechanism with a block diagonal structure of cyclic weight matrix, the model's ability to fuse memory features (short-term operations / long-term abilities) at different temporal granularities is improved, thereby enhancing the sensitivity to complex temporal behavioral patterns or ability mutations (such as sudden increases or fluctuations in students' abilities).

[0120] (1) Multi-Head Attention Mechanism

[0121] Input data: Independent output status of channels (Where B = batch size, T = time step, d = feature dimension);

[0122] Linear projection: , ,

[0123] Among them, the projection dimensions of Q, K, and V ( The relationship between () and the number of heads h ,generally

[0124] ), It is the dimension of each head.

[0125] in, It is a learnable weight matrix;

[0126] Multi-head partitioning: Reshape Q, K, and V respectively as or Then transpose to or .

[0127] The projected Q, K, V tensors (dimensions) Segment along the head dimension to obtain the i-th head. (dimension) Similarly,

[0128] Single-head attention Output dimension .

[0129] in, These are the i-th heads after Q, K, and V segmentation, typically projected linearly to... Dimensions, then segmented by the number of heads h.

[0130] Output merging: Merge all (dimension) Concatenate Then project: .

[0131] This is for outputting the projection matrix.

[0132] Multi-head attention: Split the hidden state into D groups (e.g., 4 heads); calculate the attention weights at different time granularities.

[0133]

[0134] Where X′ is the input of the memory mixing unit, i.e. the output H (dimension [B, T, d]) of the channel-independent modeling, and the self-attention parameters are: Q=K=V=X′.

[0135] (2) Block diagonal cyclic weight matrix

[0136] Matrix structure: N is the number of blocks, where each submatrix N is the number of blocks, and let the hidden state dimension be... , It is the hidden sub-state dimension (block size) corresponding to each block sub-matrix; State update formula:

[0137]

[0138] in, It is the input vector at time t (input dimension) dim ); For the input weight matrix (dimension d × input) dim ); Let b be the tanh activation function and b be the bias vector (dimension d).

[0139] It is the hidden state vector (dimension d) at time t-1; It is the hidden state vector (dimension d) at time t.

[0140]

[0141]

[0142] (3) Feature fusion mechanism

[0143] Feature fusion: Assuming and Both are [B, T, d], which concatenate to form [B, T, 2×d].

[0144]

[0145] in, For multi-head attention output; Output the block diagonally in a loop; This represents element-wise multiplication;

[0146] Gating weight calculation:

[0147] in,

[0148] This is the gated weight matrix; This indicates a concatenation operation, which concatenates two vectors into a single long vector; α is a gating scalar (range 0~1) that controls the fusion ratio of the two vectors, and the output dimension is [B, T, 1].

[0149] The output decoding unit 127 is configured to generate an evaluation vector containing dynamic output, including question type adaptation weights, content level scores, and comprehensive ability values, based on the multi-scale fusion state vector.

[0150] Prediction output: The P-sLSTM model outputs a vector consisting of three parts:

[0151]

[0152] The final model prediction output includes suggestions on question type distribution, content level suggestions, and dynamic ability scores, which are used for personalized push notifications.

[0153] : Question type weight vector, representing the recommendation weights of "multiple choice questions", "simulation questions", and "strategy questions", satisfying the normalization condition: .

[0154] : Content level suggestion vector, each element represents the student's current adaptation score to a certain level of content (such as "basic knowledge" to "advanced coping strategies", the larger the value, the better the match). The larger the value, the more suitable the student is for that level.

[0155] Student dynamic ability score: Represents a student's overall ability status at the current time point; the value range is typically within... Alternatively, it can be normalized using softmax / sigmoid activation functions to drive game difficulty adjustment.

[0156] This output vector can be obtained through a fully connected regression layer using the following regression prediction function:

[0157]

[0158] Output the weight matrix, which represents the hidden state. The mapping is to an output vector, the dimension of which is determined by the task output structure. Output bias vector to adjust the baseline of the predicted values. : Hidden state vector, generated by the P-sLSTM model at time step Output is a combination of current input and historical capability status.

[0159] By capturing multi-scale features through the multi-head mechanism in the memory mixing mechanism and mapping the fully connected layer of the output layer to educational adaptation indicators, an end-to-end transformation from behavioral features to teaching decisions is achieved.

[0160] The Ability Assessment Module 120 transforms student behavior data into actionable dynamic teaching strategies through a progressive design that includes time-series block preprocessing, anti-interference channel modeling, memory enhancement computation, and educational indicator decoding. This addresses the key problem that traditional systems cannot capture the trajectory of ability evolution.

[0161] The scene adaptation module 130 integrates a sparse dynamic recognition reinforcement learning unit, which establishes a state transition model and a proxy reward function through dictionary function combination to generate the optimal adaptation strategy that matches the student's state.

[0162] Traditional disaster education systems struggle to effectively consider individual differences among students in behavioral patterns, social tendencies, and motivational preferences, easily leading to mismatches between task assignments and student abilities, thus impacting immersion and educational effectiveness. Furthermore, existing reinforcement learning methods suffer from low training efficiency, poor generalization, and high deployment resource consumption when facing scenarios with small samples and high-dimensional inputs. To address these issues, this paper proposes a disaster scenario adaptation model based on sparse dynamic recognition reinforcement learning. This model integrates sparse function dictionary modeling and policy optimization algorithms, learning dynamic state transitions and preference feedback mechanisms from students' behavioral trajectories to improve the scientific rigor and real-time performance of disaster science education systems in personalized responses and task allocation. The implementation process of sparse reinforcement learning mainly includes: constructing dictionary functions, jointly optimizing state transition and reward functions, PPO training, and compressing policies into explicit expressions.

[0163] Furthermore, Figure 3 This is a schematic diagram of the scene adaptation module according to an embodiment of the present invention, as shown below. Figure 3 As shown, the scene adaptation module 130 includes:

[0164] The dynamic environment modeling unit 131, based on an improved sparse recognition algorithm, fits the transition relationship between student state and task action through dictionary function combination to construct a discrete-time dynamic model.

[0165] Dynamic environment modeling: The scenario adaptation module 130 is based on an improved sparse recognition algorithm. The dynamic environment modeling unit 131 within it establishes a state transition model for the disaster education interaction process based on historical student state-action pair data through a combination of nonlinear dictionary functions. Specifically, the improved sparse recognition algorithm utilizes a student dataset (containing sample pairs of student states and task actions from historical interactions). By fitting a state-action transition function and modeling the dynamic system through sparse dictionary combinations, a discrete-time modeling approach is used to predict the evolution of student states after task decisions.

[0166]

[0167] : The student's state vector at time k (such as task completion status, psychological load, social behavior indicators, etc.); The system decision / task action vector at time k (such as the currently recommended disaster task type, task complexity level, collaboration method, etc.); : The predicted student state vector at time k+1, representing the expected transition of the student's state after the current task decision; : An approximate expression of the state transition function, representing the dynamic relationship between the current state and task actions and the next state;

[0168] The i-th nonlinear dictionary function reflects the feature combination between state and task behavior (such as...). (etc.), which are predefined basis functions or characteristic functions. The basis function library contains polynomial terms (such as x², x·u) and periodic terms (such as sin(u), cos(x)). : The sparse coefficient corresponding to the i-th dictionary function, representing the importance weight of this item in the modeling process, which needs to be learned through the sparse regression algorithm; d: The total number of dictionary functions, representing the number of basis functions used for combined modeling.

[0169] Dynamic environment modeling unit 131 outputs the state transition function It can predict the evolution of the student's state after performing task u. The dynamic environment modeling unit 131 adopts a discrete-time modeling method to avoid continuous-time integration errors, and is suitable for high-frequency response modeling of state changes in complex disaster tasks.

[0170] The sparse reward modeling unit 132 constructs an interpretable surrogate reward function (i.e., indirect reward function) based on the student's state vector and the task action vector recommended by the system using a sparse regression algorithm.

[0171] To address the issue that reward signals are not directly observable in some disaster scenarios, a sparse surrogate reward function model is designed. This model uses features such as student task completion, communication behavior, and feedback attitude as state inputs to construct the surrogate reward function.

[0172]

[0173] in, The agent reward value predicted by the model is used to measure the educational benefits or system feedback obtained by the student after performing a certain task action in the current state; x: the current student's state vector, which includes behavioral features such as task completion, communication frequency, emotional feedback, and operation accuracy; u: the task action vector recommended by the current system, such as the recommended disaster scenario type (e.g., fire escape, earthquake refuge), task difficulty level, collaboration method, etc. The j-th nonlinear characteristic function reflects the combination of interactive features between state and action, for example... ; : The sparse coefficients of the j-th feature function, representing the relative importance of the feature in constructing the reward function, are usually learned through sparse regression methods.

[0174] Dynamic environment modeling unit 131 and sparse reward modeling unit 132 share the same sparse dictionary function set; among which, the state transition function With agent reward function The sparse coefficient matrix is ​​jointly optimized using the Block Coordinate Descent (BCD) algorithm to improve model consistency.

[0175] The sparse reward modeling unit 132 employs sparse regression algorithms such as the sequential threshold least squares method to select a sparse and effective subset of features from a redundant dictionary, i.e., to screen key influencing factors. It also adds L2 regularization constraints to improve generalization ability, identify the optimal reward function expression, and enhance interpretability and model stability.

[0176]

[0177] Y: The reward vector (behavioral outcome feedback) actually observed by students, including a set of objective and subjective indicators such as task success rate, situational response score, and interaction enthusiasm; X: The student status input dataset, including historical task performance, current disaster scenario, and behavioral response characteristics (such as time consumption, error rate, number of times speaking, etc.). : A set of feature mapping functions, representing the mapping of the state input X to a set of nonlinear feature spaces, serving as explanatory variables for the reward function; The sparse coefficient matrix corresponds to the contribution strength of each nonlinear feature in the reward modeling, and needs to be determined through learning (the goal is to select a small number of high-value features). The prediction error term, derived from the Frobenius norm, measures the fitting error between the surrogate reward function and the actual reward; the smaller the term, the more accurate the modeling. The L2 regularization term limits the magnitude of the coefficient matrix, prevents overfitting, and enhances the model's generalization ability in new scenarios. Regularization weight coefficients control the degree to which the model balances prediction error with model complexity (sparseness). The larger the value, the more concise and sparse the model becomes; Minimize the objective by optimizing the sparse coefficient matrix. This ensures that the predicted reward is as close as possible to the actual observed value, while maintaining the simplicity and interpretability of the model.

[0178] This optimization objective is used to train a proxy reward function model to simulate the value of student behavior in educational tasks within disaster scenarios, addressing the problem of unobservable rewards in educational settings. Since student rewards cannot be directly observed in some scenarios, the system trains a sparse function combination using finite state-behavioral data to automatically identify which state-action features have the greatest impact on reward outcomes, thereby driving personalized task recommendation strategies.

[0179] The sparse reward modeling unit 132 has the advantages of strong interpretability and high learning efficiency, and can effectively support the strategy optimization process even when student behavior data is scarce.

[0180] The policy compression unit 133 approximates the deep reinforcement learning policy in a sparse form and transforms it into an explicit functional expression, outputting a lightweight policy expression.

[0181] To improve the deployment feasibility and inference speed of the model, the trained deep reinforcement learning (neural network) policy is approximated as an explicit control expression in a sparse form. That is, the policy mapping is compressed into a sparse function combination through dictionary fitting. For example, the learned policy can be approximated as a concise function form:

[0182]

[0183] u: Output task action decision value, indicating the type of disaster task or task adjustment control amount that the system will push to students; x: Student's current state characteristic variable, such as the completion status of the previous stage task, psychological stress level, behavioral response speed, etc.; φ: Adjustable environmental parameters related to task actions, such as the probability of disaster occurrence, the urgency of the task, or the weight of multi-student collaboration, etc. The sparse representation of task actions uses predefined functions (such as polynomials and trigonometric functions) to achieve policy compression representation, thereby replacing the original deep network decision results. This formula is an exemplary expression, and the actual coefficients can be obtained through sparse regression learning.

[0184] Furthermore, the policy compression unit 133 introduces a noisy trajectory sampling mechanism to enhance the policy's generalization ability. By adding Gaussian noise to simulate environmental uncertainties, the policy's robustness is enhanced.

[0185]

[0186] The student's next state, predicted through a simulated environment or agent model, is used for strategy training and evaluation. The state transition proxy function, learned by the dynamic modeling module, describes the impact of the current state and task actions on the evolution of the student's state. The mean is 0 and the variance is 0. Gaussian noise was used to simulate environmental uncertainties and fluctuations in student behavior during real-world interactions. Noise variance parameter controls the magnitude of perturbations during strategy training, used to enhance model robustness and generalization ability.

[0187] This compression mechanism makes the model more suitable for deployment on resource-constrained platforms (such as mobile devices or embedded educational terminals).

[0188] The strategy optimization unit 134 uses a proximal policy optimization algorithm to maximize the expected cumulative agent reward. Through a two-stage iterative mechanism of pre-training in a simulated environment and fine-tuning in a real environment, it outputs the probability distribution mapping from student state to task action.

[0189] The reinforcement learning policy is trained in the constructed model environment using the Proximal Policy Optimization (PPO) algorithm (a policy gradient reinforcement learning algorithm), and the network policy is trained in a simulated environment. It continuously interacts and iterates with the real environment, ultimately outputting the optimal policy for the best fit.

[0190]

[0191] The final output is the optimal adaptation strategy, which represents the probability distribution of taking action u (e.g., recommending question type, scene difficulty, interactive task, etc.) under student state x. : Policy function, with parameter φ, is the neural network model trained using the PPO algorithm; The student's status at step t, such as the current disaster situation, task completion progress, and interaction response data; The action decision recommended by the model at step t (such as scenario selection, question type configuration, prompt feedback mechanism, etc.). : Immediate reward predicted by the sparse agent reward model, measuring the contribution of this behavior to student learning effectiveness or engagement; ƞ: Discount factor, controlling the impact of future rewards on current policy optimization, with a value range of... A larger value indicates that the model emphasizes long-term impact; TT: the maximum time step of the simulated training trajectory, representing the number of rounds or cycles the task lasts; : Represents the expectation under the policy sampling trajectory, that is, to simulate multiple rounds of task decision-making and estimate the expected cumulative reward under the guidance of the current policy; The goal is to find the policy parameter φ that maximizes the expected cumulative reward through training, i.e., to output the optimal disaster scenario adaptation policy.

[0192] The policy optimization unit 134 iterates through virtual and real environments: pre-training in a simulated environment → fine-tuning in a real environment, outputting the optimal adapted policy. This implements the output of the probability distribution of task actions for a given student state x.

[0193] The scenario adaptation module 130 achieves the dual goals of scientific decision-making and efficient deployment of personalized disaster education through a progressive design of environment modeling, reward construction, strategy compression, and closed-loop optimization. In particular, it solves the problem of implementing reinforcement learning in resource-constrained scenarios.

[0194] The feedback execution module 140 dynamically adjusts the task parameters based on the evaluation results of the comprehensive capability value and the optimal adaptation strategy, and provides feedback on the protection operation guidance through multimedia.

[0195] The feedback execution module 140 aims to integrate the student ability assessment model and the disaster scenario adaptation model into the overall system framework, forming a closed-loop process of "real-time perception—intelligent judgment—personalized push—effect feedback." By introducing a 3D interactive simulation engine, behavior assessment mechanism, and dynamic task control, it achieves precise matching between disaster science popularization content and students' ability status, thereby improving the system's teaching effectiveness and interactive immersion.

[0196] Hardware carriers: VR headset / haptic gloves (operation data acquisition), physiological sensors (heart rate monitoring and stress level).

[0197] Data flow: Behavioral trajectory: path selection, operation sequence, response time; Environmental parameters: disaster type (fire / earthquake), scene complexity; Physiological signals: eye-tracking attention focus, hand tremor level.

[0198] For example, if a student does not cover their mouth and nose with a wet towel in a "fire escape" scenario, the system will mark it as a lack of protective measures.

[0199] The feedback execution module 140 is specifically used to: trigger immediate corrective feedback based on the ability score output by the ability assessment module; wherein, the immediate corrective feedback includes textual guidance for incorrect operations and multimedia demonstrations of correct operations; map the ability assessment results to the three-dimensional decision elements of the COM-B model: ability performance, environmental opportunities, and motivational drive; generate virtual-real scene linkage tasks based on the three-dimensional decision elements; wherein, the virtual-real scene linkage tasks are associated with a geographic information system database, mapping real disaster avoidance points as virtual navigation targets; and, the feedback execution module integrates medical protection guidelines as interactive knowledge points, embedding clinical operation standards in radiation leakage and heatstroke scenarios.

[0200] (1) Constructing an intelligent feedback mechanism for simulated scenarios

[0201] The system integrates an immersive 3D disaster simulation engine (such as typical scenarios like fire escape, flood avoidance, and earthquake response), and combines virtual reality and interactive behavior acquisition technology to track students' dynamic performance in real time during the game, including path selection, protective measure selection, interaction reaction time, and operation sequence.

[0202] The feedback mechanism is deeply coupled with the P-sLSTM ability assessment model constructed by the ability assessment module 120. The system inputs students' behavioral data into the model and dynamically updates their ability status representation. The system matches the data with a disaster knowledge graph and outputs the ability score and content fit at that moment. Taking the "heat wave protection" module as an example, if the system detects that a student did not cover their mouth and nose with a wet towel during the escape, the ability model will identify this behavior as a low-level protection performance, deduct the corresponding score, and provide immediate feedback on the correct approach through text guidance, video replay, and other means.

[0203] This mechanism not only constitutes an instant evaluation system for student operations, but also provides real interactive samples for competency status modeling, which in turn feeds back into the competency assessment network for continuous optimization.

[0204] (2) Constructing a COM-B model (Capability, Opportunity, Motivation) mapping and personalized ability improvement path. Based on the disaster scenario adaptation model constructed by the scenario adaptation module 130, the system will map the student's behavioral sequence (ability status assessed by the ability model) to the disaster scenario adaptation model. (Motivation dimension, task response, etc.) are mapped to the three elements of the COM-B model:

[0205] C (Capability) Performance: A multi-dimensional capability score represented by the output of the P-sLSTM model, including knowledge mastery, operational proficiency, and policy response level, which determines the task difficulty threshold.

[0206] O (Opportunity) Environmental opportunities: such as allocable disaster scenarios, collaborative task resource pools, and real-world scenario linkage modules (such as summer resorts on the map), used to generate navigation tasks for real-world disaster relief points;

[0207] M (Motivation): Task preferences extracted from the system's behavioral trajectory, such as challenge inclination, collaboration frequency, and reaction after failure, are used to assign collaborative / solo challenge modes.

[0208] Based on the above information (COM-B model output), the disaster scenario adaptation model combines a state transition mechanism with the PPO reinforcement learning strategy to output the optimal task recommendation decision. To achieve the optimal match between students and disaster tasks.

[0209] For example, for students who show slow reaction but strong learning motivation in the "Heat Wave First Aid" level, the system will prioritize recommending the "Heatstroke Symptom Recognition" and "Summer Resort Navigation" tasks, and improve their success rate and sense of participation by appropriately reducing the time pressure of the tasks; at the same time, it will link with the "Public Summer Resort Center" database to generate the "Virtual Navigation to Nearby Summer Resorts" task in the game, so that students can establish a connection between virtual and reality and enhance their ability to immerse themselves in the scene and transform their actions.

[0210] In addition, by combining resources from medical institutions, the system can embed real clinical protection guidelines as interactive knowledge points in high-risk disaster scenarios (such as radiation leakage and heatstroke), thereby achieving a synergistic improvement in medical knowledge popularization and disaster prevention education.

[0211] This integrated architecture connects capability assessment, scenario recommendation, interactive feedback, and capability enhancement in a closed loop, forming an AI-powered disaster education platform with adaptive adjustment capabilities. It provides teenagers with a continuous and progressive disaster response training path, while ensuring the scientific nature of the content, personalized delivery, and real-time feedback, significantly improving the effectiveness and practical value of disaster prevention education.

[0212] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0213] It should also be noted that, in the embodiments of this application, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0214] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined in the embodiments of this application may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown in this application, but is to be accorded the widest scope consistent with the principles and novel features disclosed in the embodiments of this application.

Claims

1. A dynamic adaptation system for science education for teenagers in multiple disaster scenarios based on artificial intelligence, characterized in that, include: The data acquisition module is used to build a multi-dimensional disaster knowledge base and collect student behavior data in real time through a 3D disaster simulation engine. The disaster knowledge base covers a classification system of natural disasters, man-made disasters and environmental disasters. The student behavior data includes answer records, operation paths and interactive feedback. The ability assessment module, based on the student behavior data, enhances the historical state memory ability through an index gating mechanism, combines normalized state variable stable gradient updates, and dynamically outputs an assessment vector containing question type adaptation weights, content level scores, and comprehensive ability values. The scene adaptation module integrates a sparse dynamic recognition reinforcement learning unit, which establishes a state transition model and a proxy reward function through dictionary function combination to generate the optimal adaptation strategy that matches the student's state. The feedback execution module dynamically adjusts task parameters based on the evaluation results of the comprehensive capability value and the optimal adaptation strategy, and provides feedback on protection operation guidance through multimedia. The capability assessment module employs a P-sLSTM model and includes: The preprocessing unit is used to standardize and reconstruct the original student behavior data and disaster knowledge graph to obtain standardized time-series data; The block-based temporal processing unit is used to divide the standardized temporal data into several fixed-length sub-block sequences. Each sub-block independently extracts local temporal features and then linearly fuses global information to output a sub-block feature embedding vector. The channel-independent modeling unit is used to decouple the sub-block feature embedding vectors, separate the M feature channels into independent data streams, and process each channel in parallel through independent LSTM modules to output the reconstructed channel-decoupled temporal state matrix. The exponential gating unit introduces an exponential gating mechanism into the memory unit. By expanding the numerical range of the forget gate and the input gate through an exponential function, it enhances the ability to retain remote historical states, updates the cell state, and outputs the updated cell state. The state normalization unit is configured to introduce normalized state variables to balance cell state updates. It calculates normalized variables based on the updated cell state and the historical normalized state to obtain gradient-stable hidden state output. The memory fusion unit is configured to fuse capability features at different temporal granularities through a multi-head attention mechanism and a block diagonal cyclic weight matrix to generate a multi-scale fused state vector. The output decoding unit is configured to generate an evaluation vector that includes dynamic output, question type adaptation weights, content level scores, and comprehensive ability values, based on the multi-scale fusion state vector. The index gating mechanism is specifically manifested as follows: The activation function uses an exponential function, which expands the output range of the forget gate and the input gate to the positive real number range; The retention strength of historical ability states is controlled by adjusting the forget gate weight matrix, and the update magnitude of the ability assessment by the current behavior data is controlled by the input gate weight matrix. The output gate is used to explicitly adjust the output of the normalized hidden state, forming an encoding vector that contains the representation of the current capability.

2. The dynamic adaptation system for science education for teenagers in multiple disaster scenarios based on artificial intelligence as described in claim 1, characterized in that, in, The scenario adaptation module includes: The dynamic environment modeling unit, based on an improved sparse recognition algorithm, fits the transition relationship between student state and task action through dictionary function combination to construct a discrete-time dynamic model; The sparse reward modeling unit constructs an interpretable surrogate reward function based on the student's state vector and the task action vector recommended by the system using a sparse regression algorithm. The policy compression unit approximates the deep reinforcement learning policy in a sparse form and transforms it into an explicit functional expression, outputting a lightweight policy expression. The strategy optimization unit uses a proximal policy optimization algorithm to maximize the expected cumulative agent reward. Through a two-stage iterative mechanism of pre-training in a simulated environment and fine-tuning in a real environment, it outputs a probability distribution mapping from student state to task action.

3. The dynamic adaptation system for science education for teenagers in multiple disaster scenarios based on artificial intelligence according to claim 2, characterized in that, The dynamic environment modeling unit includes: Based on historical student state-action pair data, a state transition model is established by combining nonlinear dictionary functions, and a discrete-time modeling method is used to predict the evolution of student state after task decision. The dictionary function contains a combination of polynomial and trigonometric function features.

4. The dynamic adaptation system for science popularization education for teenagers in multiple disaster scenarios based on artificial intelligence according to claim 3, characterized in that, The sparse reward modeling unit includes: The sequential threshold least squares algorithm is used to screen key influencing factors from the redundant feature dictionary, and L2 regularization constraints are added to improve generalization ability.

5. The dynamic adaptation system for science education for teenagers in multiple disaster scenarios based on artificial intelligence according to claim 2, characterized in that, The strategy compression unit introduces a noisy trajectory sampling mechanism, which enhances the robustness of the strategy by adding Gaussian noise to simulate environmental uncertainties.

6. The dynamic adaptation system for science education for teenagers in multiple disaster scenarios based on artificial intelligence according to claim 2, characterized in that, in, The dynamic environment modeling unit and the sparse reward modeling unit share the same sparse dictionary function set; The sparse coefficient matrices of the state transition function and the agent reward function are jointly optimized using a block coordinate descent algorithm to improve model consistency.

7. The dynamic adaptation system for science education for teenagers in multiple disaster scenarios based on artificial intelligence according to claim 1, characterized in that, The data acquisition module includes: Knowledge base construction units, in collaboration with medical institutions, supplement medical disaster prevention knowledge to form an interdisciplinary knowledge graph; The behavioral analysis unit identifies students' cognitive levels and motivation types through the KAP questionnaire and qualitative interviews, and establishes a "disaster knowledge-cognitive ability" mapping database.

8. The dynamic adaptation system for science education for teenagers in multiple disaster scenarios based on artificial intelligence according to claim 1, characterized in that, The feedback execution module is used to implement: The ability score output by the ability assessment module triggers immediate corrective feedback; wherein, the immediate corrective feedback includes textual guidance for incorrect operations and multimedia demonstrations of correct operations; Mapping the competency assessment results to the three-dimensional decision elements of the COM-B model: competency performance, environmental opportunities, and motivational drivers; A virtual-real scene linkage task is generated based on the three-dimensional decision elements; wherein, the virtual-real scene linkage task is associated with a geographic information system database, and real disaster avoidance points are mapped as virtual navigation targets; Furthermore, the feedback execution module integrates medical protection guidelines as interactive knowledge points, embedding clinical operation standards in scenarios such as radiation leakage and heatstroke.

Images