Multi-disaster scene teenager science popularization education dynamic adaptation system based on artificial intelligence

Abstract

The invention relates to the field of artificial intelligence and disaster knowledge science popularization, and discloses a multi-disaster scene teenager science popularization education dynamic adaptation system based on artificial intelligence, and the system comprises a data collection module which is used for constructing a multi-dimensional disaster knowledge base and collecting student behavior data in real time; the ability evaluation module is used for enhancing the historical state memory ability through an index gating mechanism according to the student behavior data, and dynamically outputting an evaluation vector containing question type adaptation weights, content level scores and comprehensive ability values by combining normalized state variable stable gradient updating; the scene adaptation module is used for establishing a state transition model and an agency award function through dictionary function combination and generating an optimal adaptation strategy matched with the student state; and the feedback execution module is used for dynamically adjusting task parameters based on the evaluation result of the comprehensive capability value and the optimal adaptation strategy, and feeding back protection operation guidance in a multimedia form. According to the invention, the accuracy of the disaster education content and the intelligent degree of the teaching form are improved.

Description

Technical Field

[0001] The present invention relates to the fields of artificial intelligence and disaster knowledge popular science technology, and in particular to an artificial intelligence-based dynamic adaptation system for popular science education for teenagers in multiple disaster scenarios. Background Art

[0002] Traditional disaster education, often based on static courseware and offline lectures, lacks interactivity, real-time capabilities, and adaptability, making it difficult to meet the diverse and dynamic learning needs of young people. Especially when addressing individuals with varying cognitive abilities, learning motivations, and behavioral styles, traditional educational methods often suffer from problems such as "fixed content," "uniform pacing," and "delayed feedback."

[0003] At the same time, the integration of artificial intelligence and educational gaming technology offers a new path for disaster science education. By introducing an adaptive mechanism 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, and are unable to effectively capture the evolution of students' abilities, task selection preferences, and emotional feedback. Furthermore, in disaster-related game scenarios, due to practical constraints such as unobservable rewards, strong environmental uncertainty, and limited sample sizes, traditional reinforcement learning strategies often converge slowly and generalize poorly, making it difficult to develop efficient and deployable control strategies.

[0004] Therefore, there is an urgent need for an intelligent adaptation system that integrates high-dimensional time series modeling and reinforcement learning optimization. It can evaluate students' ability status in real time based on their behavioral data in a multi-disaster simulation environment, and automatically match disaster scenarios, content difficulty and task forms, so as to build a more immersive, interactive and educational youth disaster science popularization platform. Summary of the Invention

[0005] In response to the above problems, the present invention aims to provide an artificial intelligence-based dynamic adaptation system for youth science education in multiple disaster scenarios to improve the accuracy of disaster education content and the intelligence of teaching methods.

[0006] To achieve the above objectives, the following technical solutions are adopted:

[0007] An AI-based dynamic adaptation system for youth science education in multiple disaster scenarios, including:

[0008] A 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 capability through an exponential gating mechanism, combines the normalized state variable with stable gradient updates, and dynamically outputs an assessment vector containing question type adaptation weights, content level scores, and comprehensive ability values;

[0010] The scenario adaptation module integrates a sparse dynamic recognition reinforcement learning unit, establishes a state transition model and an agent reward function through a combination of dictionary functions, and generates the optimal adaptation strategy that matches the student's state;

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

[0012] Furthermore, the capability assessment module adopts a P-sLSTM model, including:

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

[0014] A block time series processing unit is used to divide the standardized time series data into a plurality of sub-block sequences of fixed lengths, independently extract local time series features from each sub-block, linearly fuse the global information, and output a sub-block feature embedding vector;

[0015] A channel-independent modeling unit is used to perform channel decoupling on the sub-block feature embedding vector, separate the M feature channels into independent data streams, process each channel in parallel through an independent LSTM module, and output a reconstructed channel-decoupled time series state matrix;

[0016] Exponential gating unit: This unit introduces an exponential gating mechanism into the memory unit. It expands the numerical range of the forget gate and input gate through the exponential function, enhances the ability to retain the remote historical state, and updates the cell state and outputs the updated cell state.

[0017] a state normalization unit configured to introduce a normalized state variable to balance the cell state update, calculate the normalized variable based on the updated cell state and the historical normalized state, and obtain a gradient-stabilized hidden state output;

[0018] The memory mixing unit is configured to fuse the capability features of different time 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 including a dynamic output including a question type adaptation weight, a content level score and a comprehensive ability value 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 to expand the output value range of the forget gate and input gate to the positive real number interval;

[0022] The retention strength of historical capability states is controlled by adjusting the forget gate weight matrix, and the update amplitude of capability evaluation based on current behavior data is controlled by the input gate weight matrix.

[0023] The output gate is used to adjust the normalized hidden state to form an encoding vector that represents the current ability.

[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 states and task actions through a combination of dictionary functions to construct a discrete-time dynamic model;

[0026] The sparse reward modeling unit uses a sparse regression algorithm to construct an interpretable proxy reward function based on the student's state vector and the task action vector recommended by the system;

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

[0028] The policy optimization unit uses the proximal policy optimization algorithm to maximize the expected cumulative agent reward. Through a two-stage iterative mechanism of simulated environment pre-training and real environment fine-tuning, 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 through a combination of nonlinear dictionary functions, and a discrete time modeling method is used to predict the student state evolution after task decision-making;

[0031] The dictionary function includes 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 the L2 regularization constraint is added to improve the generalization ability.

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

[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 transfer function and the agent reward function The sparse coefficient matrix of is jointly optimized through the block coordinate descent algorithm to improve model consistency.

[0037] Furthermore, the data acquisition module performs:

[0038] The knowledge base construction unit collaborates with medical institutions to supplement medical disaster protection knowledge and form an interdisciplinary knowledge map;

[0039] The behavioral analysis unit identifies students’ cognitive levels and motivation types through KAP questionnaires and qualitative interviews, and establishes a “disaster knowledge-cognitive ability” mapping database.

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

[0041] Triggering immediate corrective feedback based on the ability score output by the ability assessment module; wherein the immediate corrective feedback includes text guidance on incorrect operations and multimedia demonstrations of correct operations;

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

[0043] Generate a virtual-real scene linkage task based on the three-dimensional decision elements; wherein the virtual-real scene linkage task is associated with a geographic information system database to map the actual disaster shelter point into a virtual navigation target;

[0044] In addition, the feedback execution module integrates medical protection guidelines as interactive knowledge points and embeds clinical operation specifications in radiation leakage and heat stroke scenarios.

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

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

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

[0048] 3. The present invention designs a sparse reinforcement learning adaptation model for disaster tasks: combining sparse function dictionary modeling with reinforcement learning optimization algorithm, constructing a sparse dynamic recognition reinforcement learning system, learning state transition rules and preference feedback mechanism from students' behavior trajectories, and realizing accurate and personalized allocation of disaster scenario tasks.

[0049] 4. The present invention proposes an improved sparse recognition algorithm: it is used to establish a state transition model in the interactive process of disaster education, and to achieve efficient modeling of system behavior dynamics by combining dictionary functions under low sample conditions, thereby improving the model's adaptability to complex nonlinear disaster behaviors.

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

[0051] 6. This invention introduces sparse regression algorithms such as the sequential threshold least squares method: it automatically screens key influencing factors from the redundant feature dictionary, 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 strategy approximation module: the trained deep reinforcement learning strategy is compressed into an explicit control form through a sparse function expression, which reduces the model complexity and deployment cost, allowing the system to run on resource-constrained mobile terminals or embedded science popularization devices.

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

[0054] It should be understood that the contents described in the summary of the invention are not intended to limit the key or important features of the embodiments of the present invention, nor are they intended to limit the scope of the present invention. Other features of the present invention will become readily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] The above and other features, advantages and aspects of the embodiments of the present invention will become more apparent with reference to the following detailed description in conjunction with the accompanying drawings. The accompanying drawings are provided for a better understanding of the present invention and do not constitute a limitation of the present invention. In the accompanying drawings, the same or similar reference numerals represent the same or similar elements, among which:

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

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

[0058] Figure 3 It is a structural diagram of a scene adaptation module according to an embodiment of the present invention. DETAILED DESCRIPTION

[0059] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0060] In this document, the term "and / or" simply describes a relationship between related objects, indicating that three possible relationships exist. For example, "A and / or B" can represent: A exists alone, A and B exist simultaneously, or B exists alone. Furthermore, the character " / " in this document generally indicates that the related objects are in an "or" relationship.

[0061] Figure 1 The following is a module diagram of a dynamic adaptation system for youth science education in multiple disaster scenarios based on artificial intelligence. Figure 1 As shown, an artificial intelligence-based dynamic adaptation system 100 for youth science education in multiple disaster scenarios includes:

[0062] Data collection 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 a classification system of natural disasters, man-made disasters, and environmental disasters. Student behavior data includes answer records, operation paths, and interactive feedback;

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

[0064] The knowledge base construction unit 111 cooperates with medical institutions to supplement medical disaster protection knowledge and form an interdisciplinary knowledge map;

[0065] Based on a variety of common disaster types, such as natural disasters (such as earthquakes, floods, and typhoons), man-made disasters (such as fires and traffic accidents), and environmental disasters (such as pollution and radiation leaks), to enhance professionalism and practicality, we collaborated with expert teams from medical institutions such as Sichuan Cancer Hospital to supplement medical disaster prevention knowledge with modules on medical disasters, such as heat wave first aid and radiation protection measures. This has resulted in a comprehensive, multi-dimensional, and interdisciplinary disaster knowledge base. This knowledge base encompasses not only disaster principles but also emergency response operations, preventive measures, and rehabilitation guidance, ensuring that the content is scientific, authoritative, and relevant to reality.

[0066] Behavioral analysis unit 112, through KAP questionnaires and qualitative interviews, identifies students’ cognitive levels and motivation types, and establishes a “disaster knowledge-cognitive ability” mapping database.

[0067] Specifically, the behavioral analysis unit 112 is used to collect data on youth's disaster awareness and needs. By designing and implementing a KAP (Knowledge, Attitude, Practice) questionnaire, part of the supporting research project, this systematically collects information on youth's cognitive levels, interests, preferences, and weaknesses in response to different disaster types. Combined with qualitative interviews, this provides a deeper understanding of youth's acceptance habits, learning motivations, and emotional responses to disaster knowledge, ensuring the diversity and authenticity of the data. After cleaning and annotating the collected data, a "disaster knowledge-cognitive ability" database is constructed, providing foundational data support for subsequent intelligent adaptation algorithms and ensuring the scientific and accurate nature of the system's personalized recommendations.

[0068] Before designing the model, the design of dynamic adaptation dimensions is also included:

[0069] (1) Difficulty adaptation dimension design

[0070] Based on student behavioral data such as accuracy, response time, and number of repeated errors in interactive games, the system dynamically adjusts the difficulty of each level, enabling a progressive progression of knowledge. For example, if a student consistently masters the basics of "Identifying High Temperature Symptoms," the system automatically increases the difficulty and recommends more complex tasks, such as "Designing a Family Heatstroke Prevention Plan," to enhance the challenge and deepen the learning experience. This mechanism ensures that students learn within a difficulty range appropriate to their abilities, avoiding frustration caused by overly difficult tasks or boredom caused by overly easy tasks.

[0071] (2) Content adaptation dimension design

[0072] Combining the COM-B behavioral model (Capability, Motivation, Opportunity), we analyze students' motivational types and identify their different needs, such as achievement-driven, social-driven, or exploration-driven. Based on this analysis, we deliver personalized disaster scenarios and tasks tailored to students' interests and behavioral habits. For example, students who prefer teamwork are prioritized for multi-person disaster relief tasks to foster interaction and collaborative learning. For students who prefer individual challenges, we deliver strategy-based exercises or simulations to meet their needs for achievement and self-improvement.

[0073] The ability assessment module 120, based on the student's operation behavior, enhances the historical state memory capability through an exponential gating mechanism, combines the normalized state variable with stable gradient update, and dynamically outputs an assessment vector containing the question type adaptation weight, content level score, and comprehensive ability value;

[0074] The core task of building a student competency assessment model is to dynamically analyze students' behavioral data during interactions, thereby inferring comprehensive competency indicators such as their current knowledge mastery, strategic application ability, and cognitive load level. Based on this, the system can intelligently adjust the subsequent content structure and challenge difficulty to achieve "teaching students in accordance with their aptitude."

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

[0076] The model input data includes: (1) students’ historical answer record data: including behavioral characteristics such as answer accuracy rate, answer time, and distribution of incorrect question types; (2) content difficulty level classification in the disaster knowledge graph: structuring disaster knowledge points into multi-level difficulty levels (e.g., “heat wave protection” is divided into three levels: “basic cognition - emergency operation - scenario decision-making”).

[0077] Furthermore, the present invention adopts 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. It is used to enhance the ability to retain long-term dependent information in time series modeling and adapt to the ability change trajectory of teenagers during the game. Figure 2 Schematic diagram of the structure of the capability evaluation module of the embodiment of the present invention. Figure 2As shown, the capability assessment module 120 includes:

[0078] A preprocessing unit 121 is used to perform normalization and time series reconstruction on the original student behavior data and the disaster knowledge graph to obtain standardized time series data;

[0079] Before entering the model, the aforementioned student behavior history data (correctness rate / time consumption / error distribution) and disaster knowledge graph (structured difficulty level) undergo normalization (Z-score normalization) and time series reconstruction (original dimensions → B × L × M), resulting in standardized time series data (dimensions: B × L × M) to meet the requirements of time series modeling. Here, B is the number of samples in a batch, i.e., the number of samples processed in a single training run; L is the time step (e.g., 96 steps of historical data), i.e., the length of the input sequence; and M is the number of feature channels, i.e., the variable dimensions of the multivariate time series (e.g., question type, knowledge point, etc.).

[0080] A block time series processing unit 122 is used to divide the standardized time series data into a plurality of sub-block sequences of fixed lengths, independently extract local time series features from each sub-block, linearly fuse the global information, and output a sub-block feature embedding vector;

[0081] Building a Blocking Mechanism: This mechanism is combined with LSTM. Specifically, the student's historical behavior sequence is divided into several fixed-length sub-blocks along the time dimension. Local features are extracted from each sub-block independently (each sub-block independently learns short-term behavior patterns), and then global information aggregation is achieved through linear fusion. This mechanism, through the three steps of 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 transformed from the original dimension

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

[0084] In order to process multi-channel (M feature channels) time series data (L time steps) according to the block strategy (each block length P) and allow the subsequent independent modeling of the data of each channel, the original data X is reconstructed from the dimension [B, L, M] to [B×M, N, P]:

[0085]

[0086] in, Indicates that the original input data X (dimensions [B, L, M]) is reshaped 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 perform channel decoupling on the sub-block feature embedding vector ((B·M)×N×P), separate the M feature channels into independent data streams, process each channel in parallel through an independent LSTM module, and output a reconstructed channel-decoupled time series state matrix.

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

[0089] Build channel independence mechanism: For multivariate time series (such as answering behavior data of multiple question types and knowledge points), each channel is independently modeled to avoid information interference between channels and reduce the risk of parameter coupling and overfitting. During model training, the original data (multi-channel data independent input) is converted into Dimension, each channel is processed by independent blocks and LSTM modules, and the reconstructed time series state matrix is ​​output, that 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 Refactoring:

[0091] Among them, permute(0,3,1,2): dimension reordering [B,M,N,P]; reshape(B×M,N,P): channel merging into batch dimension.

[0092] Output reconstruction:

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

[0094] The sequence complexity is reduced through the block mechanism, the channel-independent modeling reduces interference, and the sparse behavioral data problem (such as low-frequency error patterns) is solved collaboratively.

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

[0096] Furthermore, the exponential gating mechanism is specifically manifested as: replacing the forget gate in the traditional LSTM AND input gate The activation function is replaced by an exponential function, so that the output range of the forget gate and input gate is extended to the positive real number interval , theoretically enhance the model's ability to retain key historical states, control the retention strength of historical capability states by adjusting the forget gate weight matrix, and control the update amplitude of current behavior data on capability evaluation by input gate weight matrix:

[0097]

[0098] in, : Forget gate output, used to control the degree of retention of historical student ability status (such as historical performance) at the current moment; the larger the value, the more historical ability status is retained, and the value range is ; The input gate output is used to control the update intensity of the current observation information (such as the current game performance) on the student's ability status; the larger the value, the stronger the new information input, and the value range is ;

[0099] The previous value of the forget gate activation, which represents the original input signal of the forget gate, combined with the current game input and the hidden state at the previous moment Characterizing the dynamic influence of students’ past abilities, the formula is:

[0100]

[0101] The forget gate input weight matrix is ​​used to map the current game input To the forgotten gate space; depict the impact of current game performance on the degree of retention of historical abilities. : Forget gate recurrent weight matrix, used to map the previous hidden state To the forget gate space; characterize the impact of historical ability status on the current forget decision. The forget gate bias vector is used to adjust the baseline value of the forget gate output; is the input vector (Input for the current time step).

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

[0103]

[0104] Input gate input weight matrix, used to map the current game input to the input gate space; characterize the impact of current game performance on ability update. Input gate recurrent weight matrix, used to map the previous hidden state to the input gate space; characterize the impact of historical capability states on current update decisions. Input gate bias vector, used to adjust the baseline value of the input gate output.

[0105] In this structure, the exponential gating mechanism expands the values ​​of the forget gate and the input gate to the positive real number interval, thereby enhancing the memory ability of remote historical states.

[0106] At the same time, the normalized state variables are introduced , improve hidden state Numerical stability during updates, preventing gradients from exploding or vanishing.

[0107] The state normalization unit 125 is configured to introduce a normalized state variable to balance the cell state update, and according to the updated cell state and historical normalized state Calculate the normalized variable to obtain the gradient-stabilized hidden state output

[0108] Introducing normalized state variables , improve hidden state Numerical stability during the update process prevents gradient explosion or vanishing. The output gate is used to explicitly adjust the normalized hidden state to form an encoding vector that contains the current capability representation.

[0109]

[0110] Normalized state, used to standardize the hidden state Calculation, to prevent the numerical explosion or disappearance of the student ability state gradient in the long time series, reflecting the comprehensive degree of historical and current information, the initial value 0 or 1. The normalized state at the previous moment represents the impact of history on the standardization of the current capability state. Hidden state, represents the student's ability performance at the current moment, combining the current input, historical state and output gate adjustment; ε is a smoothing term, The output gate uses the sigmoid activation function to determine the degree of explicit output of the normalized ability state at the current moment, which is similar to the degree of "display" of the current ability performance. The formula is:

[0111]

[0112] in, Output gate input weight matrix, used to map the current game input to the output gate space. The output gate recurrent weight matrix is ​​used to map the historical hidden state to the output gate space. Output gate bias vector, used to adjust the baseline response of the output gate. The cell state represents the student's ability state in long-term memory and is updated by combining the forget gate and input gate mechanisms. The formula is:

[0113]

[0114] The cell state at the previous moment, that is, the student's ability accumulated historically. Candidate cell state, which is determined by the current input and the historical state The activation function is generated to indicate the possible trend of ability change. 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 recurrent 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 effectively stabilizes the hidden state The gradient changes of the model can avoid unstable learning caused by ability fluctuations and improve the modeling effect of the medium and long-term behavioral trends.

[0118] The memory mixing unit 126 is configured to fuse capability features of different time 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 hybrid memory mechanism: Combining the multi-head mechanism with a block diagonal structured circulant weight matrix improves the model's ability to integrate memory features of different time granularities (short-term operations / long-term capabilities), thereby increasing its sensitivity to complex temporal behavior patterns or sudden changes in capabilities (such as sudden increases or fluctuations in students' capabilities).

[0120] (1) Multi-Head Attention

[0121] Input data: Channel independent output status (where B = batch size, T = time steps, and d = feature dimension);

[0122] Linear projection: 、 、

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

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

[0125] in, is a learnable weight matrix;

[0126] Multi-head split: reshape Q, K, V into or , and then transpose to or .

[0127] The projected Q, K, V tensors (dimensions ) is split along the head dimension, and the i-th head is (Dimension ), similarly

[0128] Single-headed attention , output dimension .

[0129] in, The ith head after segmentation is Q, K, and V, respectively, usually by linear projection to dimension, and then divided by the number of heads h.

[0130] Output Merge: Combine all (Dimension ) concatenate into Then project: .

[0131] is the output 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] Among them, X′ is the input of the memory mixing unit, that is, the output H (dimension [B, T, d]) of channel-independent modeling, and the self-attention parameters are: Q=K=V=X′.

[0135] (2) Block diagonal circulant weight matrix

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

[0137]

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

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

[0140]

[0141]

[0142] (3) Feature fusion mechanism

[0143] Feature fusion: , assuming and They are all [B, T, d], and after splicing they become [B, T, 2×d];

[0144]

[0145] in, It is multi-head attention output; Output for block diagonal loop; represents element-wise multiplication;

[0146] Gating weight calculation:

[0147] in,

[0148] is the gating weight matrix; Represents the concatenation operation, which concatenates two vectors into a long vector; α is a gated 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 including a dynamic output including a question type adaptation weight, a content level score and a comprehensive ability value according to the multi-scale fusion state vector.

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

[0151]

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

[0153] : Question type proportion vector, representing the recommendation weights of "multiple choice questions", "simulation operation questions" and "strategy questions", respectively, meeting the normalization conditions: .

[0154] : Content level recommendation vector, each element represents the student's current adaptability score for a certain level of content (such as "basic knowledge" to "advanced response solutions", the larger the value, the better the match), and the larger the value, the more suitable it is for that level.

[0155] : Student dynamic ability score, which indicates the comprehensive ability status of the student in the current period. The value range is usually Or it can be normalized through the softmax / sigmoid activation function to drive the game difficulty adjustment.

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

[0157]

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

[0159] The multi-head mechanism in the memory hybrid mechanism captures multi-scale features and the fully connected layer of the output layer is mapped into educational adaptation indicators to achieve end-to-end conversion from behavioral features to teaching decisions.

[0160] The ability assessment module 120 converts student behavior data into actionable dynamic teaching strategies through the progressive design of time-series block preprocessing, anti-interference channel modeling, memory-enhanced computing and educational indicator decoding, which core solves the key problem that traditional systems cannot capture the trajectory of ability evolution.

[0161] The scenario adaptation module 130 integrates a sparse dynamic recognition reinforcement learning unit, establishes a state transition model and an agent reward function through a combination of dictionary functions, and generates an optimal adaptation strategy that matches the student's state;

[0162] Traditional disaster education systems find it difficult to effectively account for individual differences in students' behavior patterns, social tendencies, and motivational preferences, which can easily lead to a mismatch between task push and student abilities, thus affecting immersion and educational effectiveness. In addition, existing reinforcement learning methods have problems such as low training efficiency, difficulty in generalization, and high deployment resource consumption when faced with scenarios such as small samples and high-dimensional inputs. To address the above problems, this step uses a disaster scenario adaptation model based on sparse dynamic recognition reinforcement learning. This model integrates sparse function dictionary modeling and policy optimization algorithms. By learning dynamic state transitions and preference feedback mechanisms from students' behavioral trajectories, it improves the scientificity and real-time performance of the disaster science education system in personalized response and task allocation. The implementation process of sparse reinforcement learning mainly includes: constructing a dictionary function, jointly optimizing state transitions and reward functions, PPO training, and compressing strategies into explicit expressions.

[0163] Further, Figure 3 Schematic diagram of the structure of the scene adaptation module of the embodiment of the present invention. Figure 3 As shown, the scene adaptation module 130 includes:

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

[0165] Dynamic environment modeling: The scene adaptation module 130 is based on an improved sparse recognition algorithm, in which the dynamic environment modeling unit 131 is based on historical student state-action pair data and establishes a state transition model in the disaster education interaction process through a combination of nonlinear dictionary functions. Specifically, the improved sparse recognition algorithm uses the student dataset (which contains sample pairs of student states and task actions in historical interactions) ) Fit the state-action transfer function, implement dynamic system modeling through sparse dictionary combination, and use discrete time modeling to predict the student state evolution after task decision:

[0166]

[0167] : The student state vector at the kth moment (such as task completion, psychological load, social behavior indicators, etc.); System decision / task action vector at time k (e.g., currently recommended disaster mission type, mission complexity level, collaboration method, etc.); : The predicted student state vector at the k+1th moment, which represents the expected transfer of the student state after the current task decision; : An approximate expression of the state transition function, which represents the dynamic relationship between the current state and task action and the next state;

[0168] The i-th nonlinear dictionary function reflects the feature combination between the state and the 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, which indicates the importance weight of this item in the modeling process and needs to be learned through the sparse regression algorithm; d: The total number of dictionary functions, which indicates the number of basis functions used for combined modeling.

[0169] The state environment modeling unit 131 outputs the state transfer function 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 uses a sparse regression algorithm to construct an interpretable proxy reward function (i.e., an indirect reward function) based on the student's state vector and the task action vector recommended by the system;

[0171] To address the problem that reward signals cannot be directly observed in some disaster scenarios, a sparse proxy reward function model is designed. That is, the proxy reward function is constructed by using characteristics such as student task completion, communication behavior, and feedback attitude as state input:

[0172]

[0173] in, The proxy reward value predicted by the model is used to measure the educational benefits or system feedback obtained by the student after performing a task action in the current state; x: The current student state vector, which includes behavioral characteristics 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 (such as fire escape, earthquake shelter), task difficulty level, and collaboration method; The jth nonlinear feature function reflects the interaction feature combination between state and action, for example ; : The sparse coefficient of the j-th feature function, which indicates the relative importance of the feature in constructing the reward function, and is usually learned through sparse regression methods.

[0174] The dynamic environment modeling unit 131 and the sparse reward modeling unit 132 share the same sparse dictionary function set; wherein the state transition function and the agent reward function The sparse coefficient matrix of is jointly optimized through the Block Coordinate Descent (BCD) algorithm to improve model consistency.

[0175] The sparse reward modeling unit 132 uses a sparse regression algorithm such as a sequential threshold least squares algorithm to select a sparse and effective feature subset from a redundant dictionary, that is, to screen key influencing factors. It also adds an L2 regularization constraint to improve generalization ability, identify the optimal reward function expression, and improve interpretability and model stability:

[0176]

[0177] Y: The reward vector (behavioral outcome feedback) actually observed by the student, 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 speeches, etc.); : A set of feature mapping functions, which represents mapping the state input X to a set of nonlinear feature spaces as explanatory variables of the reward function; : A sparse coefficient matrix, corresponding to the contribution strength of each nonlinear feature in reward modeling, which needs to be determined through learning (the goal is to select a small number of high-value features); : Apply the prediction error term of the Frobenius norm to measure the fitting error between the agent reward function and the actual reward. The smaller the error, 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 coefficient, which controls the degree of trade-off between the model's prediction error and model complexity (sparseness). The larger it is, the more concise and sparse the model tends to be; : Minimize the objective by optimizing the sparse coefficient matrix , making the predicted reward as close as possible to the true observation value while ensuring the simplicity and interpretability of the model.

[0178] This optimization objective is used to train a proxy reward function model, simulating the value of student behavior in disaster scenario education tasks and addressing the unobservable nature of rewards in educational settings. Because 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 converts it into an explicit function 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 approximately represented as a concise function form:

[0182]

[0183] u: The output task action decision value, which indicates the disaster task type or task adjustment control amount that the system will push to the student; x: The student's current state characteristic variable, such as the completion status of the previous stage task, psychological stress level, behavioral response speed, etc.; φ: The adjustable environmental parameters related to the task action, such as the probability of disaster occurrence, task urgency, or multi-student collaboration weight; : A sparse expression of task actions, which uses a combination of 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 generalization capability. By adding Gaussian noise to simulate environmental uncertainty, the policy robustness is enhanced, namely:

[0185]

[0186] : The student's next state is predicted by the simulated environment or agent model, which is used for strategy training and evaluation; : The state transition agent function is learned by the dynamic modeling module and describes the impact of the current state and task action on the student state evolution; : The mean is 0 and the variance is Gaussian noise is used to simulate environmental uncertainty and student behavior fluctuations in real interactions; : Noise variance parameter, which controls the disturbance amplitude during strategy training and is used to enhance model robustness and generalization ability

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

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

[0189] Use the Proximal Policy Optimization (PPO) algorithm (a policy gradient reinforcement learning algorithm) to train reinforcement learning strategies in the constructed model environment, train network strategies in a simulated environment, and continuously interact and iterate with the real environment to ultimately output the optimal strategy for adaptability:

[0190]

[0191] : The final output of the optimal adaptation strategy is the probability distribution of taking action u (such as which question type, scenario difficulty, interactive task, etc.) under the student state x; : Strategy function, parameter is φ, neural network model trained by PPO algorithm; : Student status at step t, such as the current disaster situation, task completion progress, interactive response data, etc. : The action decision recommended by the model in step t (such as scenario selection, question type configuration, prompt feedback mechanism, etc.); : The immediate reward predicted by the sparse agent reward model, which measures the contribution of the behavior to the student's learning effect or immersion; ƞ: Discount factor, which controls the impact of future rewards on the current strategy optimization, with a value range of , the larger the value, the more the model pays attention to long-term impact; TT: the maximum time step of the simulated training trajectory, representing the number of rounds or rounds of the task; : represents the expectation under the policy sampling trajectory, that is, simulating multiple rounds of task decision-making under the guidance of the current policy and estimating the expected cumulative reward; :The goal is to find the policy parameter φ that maximizes the cumulative reward expectation through training, that is, to output the optimal disaster scenario adaptation strategy

[0192] The strategy optimization unit 134 iterates through virtual and real environments: pre-training in simulated environments → fine-tuning in real environments, and outputs the optimal adaptation strategy , which realizes the output of task-action probability distribution 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 the progressive design of environment modeling → reward construction → strategy compression → closed-loop optimization, especially solving the implementation problem of 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 protection operation guidance in the form of multimedia.

[0195] Feedback execution module 140 aims to integrate the student capability 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 notification - effect feedback." By introducing a 3D interactive simulation engine, a behavior assessment mechanism, and dynamic task control, it achieves a precise match between disaster science content and student capabilities, enhancing the system's teaching effectiveness and interactive immersion.

[0196] Hardware carrier: VR helmet / tactile gloves (operation 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 degree.

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

[0199] The feedback execution module 140 is specifically used to achieve: triggering immediate corrective feedback based on the ability score output by the ability assessment module; wherein, the immediate corrective feedback includes text guidance for incorrect operations and multimedia demonstrations of correct operations; mapping the ability assessment results to the three-dimensional decision-making elements of the COM-B model: ability performance, environmental opportunities and motivation drive; generating virtual-reality scene linkage tasks based on the three-dimensional decision-making elements; wherein, the virtual-reality scene linkage tasks are associated with the geographic information system database, mapping the actual disaster shelter points to virtual navigation targets; and, the feedback execution module integrates medical protection guidelines as interactive knowledge points, and embeds clinical operation specifications in radiation leakage and heatstroke scenarios.

[0200] (1) Building an intelligent feedback mechanism for simulation scenarios

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

[0202] The feedback mechanism is deeply coupled with the P-sLSTM capability assessment model built by the capability assessment module 120. The system inputs the student’s behavior data into the model and dynamically updates its capability status representation. The system then matches the knowledge graph with the disaster knowledge graph to output the competency score and content suitability at that moment. For example, in the "heat wave protection" module, if the system detects that a student did not cover their mouth and nose with a wet towel during escape, the competency model will identify this behavior as low-level protection, deduct the corresponding points, and provide immediate feedback on the correct approach through text guidance and video replay.

[0203] This mechanism not only constitutes an instant evaluation system for students' operations, but also provides real interaction samples for ability status modeling, feeding back to the continuous optimization of the ability assessment network.

[0204] (2) Constructing the 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 converts the student's behavior sequence (ability status evaluated by the capability model) into , motivational dimensions, task response, etc.) are mapped to the three elements of the COM-B model:

[0205] C (Capability): This is represented by the multidimensional capability score output by the P-sLSTM model, including knowledge mastery, operational proficiency, and strategic 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 shelters on a map) to generate real-world disaster shelter navigation tasks;

[0207] M (Motivation): Task preferences extracted from system behavior trajectories, such as challenge tendency, collaboration frequency, and post-failure reactions, are used to allocate collaborative / single-player challenge modes.

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

[0209] For example, for students who show slow reaction in the "Heat Wave First Aid" level but have strong learning motivation, the system will prioritize recommending the "Heat Stroke Symptom Identification" and "Summer Resort Navigation" tasks, and increase 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 connections between virtual and reality, and enhance their ability to immerse themselves in scenes and transform actions.

[0210] In addition, by combining cooperative resources with medical institutions, the system can embed real clinical protection guidelines as interactive knowledge points in high-risk disaster scenarios (such as radiation leaks and heat stroke), achieving an integrated improvement in medical knowledge popularization and disaster prevention education.

[0211] This integrated architecture connects the closed loop of capability assessment - scenario recommendation - interactive feedback - capability improvement to form an artificial intelligence disaster education platform with adaptive regulation capabilities, providing young people with a continuous and advanced disaster response training path, while ensuring the scientific nature of the content, personalized push, and instant 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. Reference can be made to the common and similar parts between the various embodiments. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the description of the methods.

[0213] It should also be noted that, in the embodiments of the present application, relational terms such as first and second, etc. are merely used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply the existence of any such actual relationship or order between these entities or operations. Moreover, the terms "comprise", "include" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the statement "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device comprising the elements.

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

Claims

1. A dynamic adaptation system for youth science education in multiple disaster scenarios based on artificial intelligence, characterized by: include: A 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 capability through an exponential gating mechanism, combines the normalized state variable with stable gradient updates, and dynamically outputs an assessment vector containing question type adaptation weights, content level scores, and comprehensive ability values; The scenario adaptation module integrates a sparse dynamic recognition reinforcement learning unit, establishes a state transition model and an agent reward function through a combination of dictionary functions, and generates the optimal adaptation strategy that matches the student's state; The feedback execution module dynamically adjusts the task parameters based on the evaluation results of the comprehensive capability value and the optimal adaptation strategy, and provides protection operation guidance in the form of multimedia.

2. The artificial intelligence-based dynamic adaptation system for multi-disaster scenario science education for young people according to claim 1 is characterized by: in, The capability assessment module adopts the P-sLSTM model, which includes: The preprocessing unit is used to standardize and reconstruct the time series of the original student behavior data and disaster knowledge graph to obtain standardized time series data; A block time series processing unit is used to divide the standardized time series data into a plurality of sub-block sequences of fixed lengths, independently extract local time series features from each sub-block, linearly fuse the global information, and output a sub-block feature embedding vector; A channel-independent modeling unit is used to perform channel decoupling on the sub-block feature embedding vector, separate the M feature channels into independent data streams, process each channel in parallel through an independent LSTM module, and output a reconstructed channel-decoupled time series state matrix; Exponential gating unit: This unit introduces an exponential gating mechanism into the memory unit. It expands the numerical range of the forget gate and input gate through the exponential function, enhances the ability to retain the remote historical state, and updates the cell state and outputs the updated cell state. a state normalization unit configured to introduce a normalized state variable to balance the cell state update, calculate the normalized variable based on the updated cell state and the historical normalized state, and obtain a gradient-stabilized hidden state output; The memory mixing unit is configured to fuse the capability features of different time 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 including a dynamic output including a question type adaptation weight, a content level score and a comprehensive ability value based on the multi-scale fusion state vector.

3. The artificial intelligence-based dynamic adaptation system for multi-disaster scenario science education for young people according to claim 2 is characterized by: The exponential gating mechanism is specifically manifested as follows: The activation function uses an exponential function to expand the output value range of the forget gate and input gate to the positive real number interval; The retention strength of historical capability states is controlled by adjusting the forget gate weight matrix, and the update amplitude of capability evaluation based on current behavior data is controlled by the input gate weight matrix. The output gate is used to explicitly adjust the normalized hidden state to form an encoding vector that represents the current ability.

4. The artificial intelligence-based dynamic adaptation system for multi-disaster scenario science education for young people according to claim 1 is characterized by: in, The scene adaptation module includes: The dynamic environment modeling unit, based on an improved sparse recognition algorithm, fits the transition relationship between student states and task actions through a combination of dictionary functions to construct a discrete-time dynamic model; The sparse reward modeling unit uses a sparse regression algorithm to construct an interpretable proxy reward function based on the student's state vector and the task action vector recommended by the system; The policy compression unit approximates the deep reinforcement learning policy in a sparse form and converts it into an explicit function expression, outputting a lightweight policy expression; The policy optimization unit uses the proximal policy optimization algorithm to maximize the expected cumulative agent reward. Through a two-stage iterative mechanism of simulated environment pre-training and real environment fine-tuning, it outputs a probability distribution mapping from student state to task action.

5. The artificial intelligence-based dynamic adaptation system for multi-disaster scenario science education for young people according to claim 4 is characterized by: The dynamic environment modeling unit includes: Based on historical student state-action pair data, a state transition model is established through a combination of nonlinear dictionary functions, and a discrete time modeling method is used to predict the student state evolution after task decision-making; The dictionary function includes a combination of polynomial and trigonometric function features.

6. The artificial intelligence-based dynamic adaptation system for multi-disaster scenario science education for young people according to claim 5 is characterized by: The sparse reward modeling unit includes: The sequential threshold least squares algorithm is used to filter key influencing factors from the redundant feature dictionary, and the L2 regularization constraint is added to improve the generalization ability.

7. The artificial intelligence-based dynamic adaptation system for multi-disaster scenario science education for young people according to claim 4 is characterized by: The policy compression unit introduces a noisy trajectory sampling mechanism to simulate environmental uncertainty by adding Gaussian noise to enhance the robustness of the policy.

8. The artificial intelligence-based dynamic adaptation system for multi-disaster scenario science education for young people according to claim 4 is characterized by: in, The dynamic environment modeling unit and the sparse reward modeling unit share the same sparse dictionary function set; Among them, the sparse coefficient matrices of the state transfer function and the agent reward function are jointly optimized through the block coordinate descent algorithm to improve model consistency.

9. The artificial intelligence-based dynamic adaptation system for multi-disaster scenario science education for young people according to claim 1 is characterized by: The data acquisition module includes: The knowledge base construction unit collaborates with medical institutions to supplement medical disaster protection knowledge and form an interdisciplinary knowledge map; The behavioral analysis unit uses KAP questionnaires and qualitative interviews to identify students' cognitive levels and motivation types, and establishes a "disaster knowledge-cognitive ability" mapping database.

10. The artificial intelligence-based dynamic adaptation system for multi-disaster scenario science education for young people according to claim 1 is characterized by: The feedback execution module is used to implement: Triggering immediate corrective feedback based on the ability score output by the ability assessment module; wherein the immediate corrective feedback includes text guidance on incorrect operations and multimedia demonstrations of correct operations; Mapping the capability assessment results to the three-dimensional decision-making elements of the COM-B model: capability performance, environmental opportunities, and motivational drivers; Generate a virtual-real scene linkage task based on the three-dimensional decision elements; wherein the virtual-real scene linkage task is associated with a geographic information system database to map the actual disaster shelter point into a virtual navigation target; In addition, the feedback execution module integrates medical protection guidelines as interactive knowledge points and embeds clinical operation specifications in radiation leakage and heat stroke scenarios.

Images