Multi-scenario Intelligent Decision-Making System and Method Based on Mental Modeling

Abstract

This invention relates to the field of artificial intelligence technology, specifically to a multi-scenario intelligent decision-making system and method based on mental modeling. The method includes: acquiring multimodal environmental perception data; performing data alignment and preprocessing; generating perception input vectors and mapping them to a high-dimensional orthogonal feature space; extracting the high-dimensional orthogonal feature mappings into continuous cognitive waveform signals; dynamically inputting the cognitive waveforms into a mental evolution graph network for message passing and node state updates; dynamically adjusting the structure of the mental evolution graph based on node state similarity, information entropy, and decision error feedback; obtaining mental state vectors from the evolution of the mental evolution graph; inputting these vectors into a decision decoder to obtain decision branches and generate candidate decision schemes; mapping decision outputs according to task type; evaluating the decision execution effect and using the evaluation results as feedback signals; and iteratively executing the preceding processes based on the feedback signals. This invention reduces scene switching costs and supports continuous learning and knowledge accumulation.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a multi-scenario intelligent decision-making system and method based on mental modeling. Background Technology

[0002] Artificial intelligence (AI) is the theory, methods, technologies, and application systems that utilize digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In recent years, with breakthroughs in deep learning technology, generative AI, represented by large language models, has made significant progress, driving a paradigm shift in artificial intelligence from analysis and judgment to content creation.

[0003] However, while existing AI systems perform well with specific training data distributions, their performance drops sharply when faced with scenarios outside of these distributions. They suffer from data distribution dependence and a lack of generalization ability, resulting in significant domain limitations. Existing AI systems rely on a single large language model as the decision-making hub. Although they can call external tools through the MCP protocol, their reasoning process remains a "black box" end-to-end generation process. The vulnerability of the single-agent architecture and the lack of collaborative verification lead to serious issues of decision-making uniformity. Existing AI systems enter a "frozen state" after training, unable to continuously learn from new experiences. Furthermore, the entire model needs to be retrained every time the environment changes, resulting in high costs and low efficiency. Existing intelligent decision-making systems suffer from insufficient decision security, lacking effective internal decision review mechanisms. Traditional systems often use input filtering methods, which cannot achieve effective security control over the system's autonomously generated decisions. Summary of the Invention

[0004] The present invention aims to provide a multi-scenario intelligent decision-making method based on mental modeling to solve the problem of insufficient security of autonomously generated decisions in existing technologies.

[0005] According to one aspect of the present invention, a multi-scenario intelligent decision-making method based on mental modeling includes the following steps: Step 1: Acquire multimodal environment perception data, perform data alignment and preprocessing on the multimodal environment perception data, and generate perception input vectors; Also includes: Step 2: Map the perceptual input vector to a high-dimensional orthogonal feature space, extract the high-dimensional orthogonal features, and map the high-dimensional orthogonal features into a continuous cognitive waveform signal. The cognitive waveform is encoded in a multi-channel time-frequency domain, with each channel corresponding to a class of orthogonal feature components. Step 3: Perform multi-dimensional operations based on the pre-constructed mental evolution graph network to generate a decision candidate set. Dynamically input the cognitive waveform into the pre-constructed mental evolution graph network to perform message passing and node state updates. Based on the similarity of node states, changes in information entropy, and decision error feedback, dynamically adjust the structure of the mental evolution graph. The mental state vector is obtained from the evolution of the mental evolution graph. Step 4: Input the mental state vector into the decision decoder to generate candidate decision schemes. Candidate decision schemes are forcibly intercepted and input into a serial three-layer security decision-making mechanism for verification: First, a red line test of the safety circuit breaker is performed. If it fails, the candidate decision-making scheme is discarded directly. If it passes, it will proceed to the rule verification layer for compliance testing; if it fails, the proposal will be rejected. If it passes, it will proceed to the utility optimization layer for scoring; Only after a candidate decision solution passes the security circuit breaker layer and the rule verification layer is it allowed to map the utility-optimized decision to a specific output. Step 5: Evaluate the effectiveness of the decision implementation after the decision output, and use the evaluation results as feedback signals; Step 6: Iterate through steps 1 to 5 based on the feedback signal.

[0006] The beneficial effects of this plan are: For input multimodal environmental perception data, a three-layer security decision-making mechanism is first used to avoid decisions that could lead to significant security problems and prevent security issues arising under the guidance of artificial intelligence. The mental evolution mechanism, which digitizes the cognitive evolution, concept formation, associative learning, and knowledge reconstruction mechanisms of artificial intelligence, is used to construct an intelligent decision-making method that integrates high-dimensional orthogonal feature extraction, dynamic evolution of mental evolution graph network, decision generation and risk assessment, and closed-loop iterative optimization. This method achieves adaptability, robustness, interpretability, and continuous evolution capability in complex and ever-changing scenarios.

[0007] Furthermore, the multimodal environmental perception data includes visual images, radar point clouds, and voice commands.

[0008] Furthermore, the data alignment employs a timestamp synchronization mechanism to map heterogeneous data into a unified vector space.

[0009] Furthermore, the red line detection of the safety fuse layer includes physical damage detection and ethical red line detection.

[0010] Furthermore, in step 4, the matching degree between each candidate decision scheme and the current graph state is calculated based on the mental evolution graph network. The candidate decision is used as temporary input to perform a forward inference on the mental graph network, measuring the change in node activation patterns caused by the decision, and then the confidence level is calculated. Defined as: ; in, For node v and decision The relevance score, Let σ be the node importance weight, and σ be the sigmoid function. Set up a lightweight risk assessment network Where c is the scenario constraint vector, and a lightweight risk assessment network outputs a risk score for each candidate decision. The overall utility of each candidate decision is calculated and expressed as: ; Where λ is the risk aversion coefficient; The optimal candidate decision scheme is output based on the overall utility, and is expressed as: .

[0011] Furthermore, in step 4, the lightweight risk assessment network includes: The feature fusion layer projects multiple inputs through independent linear layers to a set dimension and then adds them together, activated by ReLU. The risk setting layer sets up risk heads for different risk modes that run in parallel. Each risk head consists of two MLP layers. The risk aggregation layer concatenates the outputs of the three risk heads, and then outputs the final risk score through an MLP layer plus a sigmoid function.

[0012] The beneficial effects are: the risk assessment network uses addition for fusion, which can reduce the number of parameters and promote intermodal interaction. Each risk head can learn different types of risk patterns (such as physical risk, rule risk, and strategy risk), thereby enhancing the robustness of risk perception.

[0013] Furthermore, in step 5, new environmental state data is collected as the result of the decision execution in the next perception cycle after the decision is executed or after a predefined waiting time, and the environmental state data is normalized. The environmental state includes task completion indicators, reward signals, safety constraint satisfaction status, and resource consumption. By aggregating multiple environmental state data in a weighted sum manner, the actual assessment value R is obtained, expressed as: ; in, Assigning weights to each environmental state data point. and These are the penalty coefficients for safety and cost, respectively. Task completion rate indicators For safety reasons, This refers to resource consumption. Based on the actual assessment value and the obtained expected value The error is expressed as: ; Error As the starting point of the gradient of the loss function, it propagates backward along the computational graph of the system. For the gradient information required for the evolution of the graph structure, the contribution of each edge to the final evaluation value is obtained by using the policy gradient or by the automatic differentiation of the graph neural network.

[0014] Furthermore, in step 5, it is determined that the decision error occurs over multiple consecutive periods. If the result is consistently positive and the error mainly originates from a certain type of unmodeled cognitive pattern, then a new node is added to the mental evolution graph network for refinement, and the state of the new node is initialized with the feature component in the current mental state vector that is positively correlated with the error. Decision node in multiple errors If the feedback consistently contributes a tiny gradient or has extremely low activation, then pruning is triggered. The edge weights are updated based on the gradient of the error with respect to each edge, as follows: ; in, The learning rate is set as the edge weight, and combined with Heblin learning, the connection between two nodes is strengthened when they are both highly activated in successful decisions, and weakened when they are both activated in failed decisions. When an error signal indicates that the current cognitive structure cannot effectively handle a certain type of scenario, a topology reconnection operation is triggered.

[0015] The beneficial effects are: updating the mental evolution graph network based on decision error can remove redundant nodes, reduce model complexity, and make the network update more accurate.

[0016] Furthermore, in step 6, the evolved mental evolution graph network state is directly used as the initial state for the next round of iteration. After each iteration, check whether the number of nodes and edges in the mental evolution graph network structure exceeds the preset limit. If it does, perform global pruning. At the same time, check for gradient explosion or vanishing phenomena and perform gradient pruning if necessary.

[0017] The beneficial effects are: the state of the mental evolution map network serves as the initial state for the next cycle, which can improve the continuity of cognitive accumulation; the updating of the mental map structure can ensure the efficient operation of the mental map and reduce redundancy.

[0018] According to another aspect of the present invention, a multi-scenario intelligent decision-making system based on mental modeling includes a multimodal sensor, a memory, and a processor. The multimodal sensor is used to collect multimodal environmental perception data and transmit it to the processor. The memory stores computer-executable instructions, and the processor is used to execute the executable instructions stored in the memory. When the executable instructions are executed by the processor, they are executed according to the steps of the method described above. Attached Figure Description

[0019] Figure 1 This is a flowchart of an embodiment of the multi-scenario intelligent decision-making method based on mental modeling, which is Example 1 of the present invention. Detailed Implementation

[0020] The following detailed description provides further details on specific implementation methods. Example 1

[0021] A multi-scenario intelligent decision-making method based on mental modeling, such as Figure 1 As shown, it includes the following steps: Step 1: Collect multimodal environmental perception data in real time using multimodal sensors, perform data alignment and preprocessing on the multimodal environmental perception data, and generate perception input vectors. The original environmental data includes visual images, radar point clouds, and voice commands. The data alignment adopts a timestamp synchronization mechanism to map heterogeneous data into a unified vector space.

[0022] Data alignment is achieved through a timestamp synchronization mechanism. This alignment aims to synchronize sensing data of different frequencies and dimensions in time, forming a unified multimodal sensing frame. Preprocessing includes existing denoising, normalization, and missing value imputation operations to generate standardized sensing input vectors.

[0023] Before data collection, initialization must be performed, including: The mental model is initialized by constructing an initial Mind Evolution Graph Network (MEGN). This graph network contains an initial set of nodes and edges, represented as follows: V represents the set of nodes, and E represents the set of edges. The set of nodes represents cognitive primitives, belief states, or intention representations, while the set of edges represents the strength of associations and information transmission paths between cognitive primitives. During initialization, the node feature vectors are randomly initialized or assigned values ​​based on a pre-trained base model.

[0024] Parameter configuration: Load the hyperparameters required for operation, including the dimension of the high-dimensional orthogonal feature space, the resolution parameter for waveform generation, the iteration step size for mental evolution, the decision confidence threshold, and the optimization objective function for closed-loop iteration.

[0025] Context mapping: Based on the current application scenario, such as autonomous driving, battlefield simulation, human-machine collaboration, etc., load the scenario prior knowledge base and establish the initial mapping relationship between the environment state space and the system internal representation space.

[0026] Step 2 involves mapping the perceptual input vector into a high-dimensional mental state vector using a deep neural network, thus constructing a mental evolution graph. The perceptual input vector is then mapped to a high-dimensional orthogonal feature space, and high-dimensional orthogonal features are extracted to form high-dimensional orthogonal feature vectors. The perceptual input vector is then mapped through a high-dimensional orthogonal feature mapping network, which introduces orthogonality constraints, such as Gram-Schmidt orthogonalization or orthogonal regularization loss functions, to ensure that feature components of different dimensions are independent and decoupled. High-dimensional orthogonal features include: environmental geometric features, semantic entity features, temporal dynamic features, and interaction relationship features.

[0027] High-dimensional orthogonal features are mapped into continuous cognitive waveform signals. These waveforms are encoded using multi-channel time-frequency domain coding, with each channel corresponding to a class of orthogonal feature components. This transforms the numerical changes of high-dimensional orthogonal feature vectors into amplitude, frequency, and phase modulation of the waveform, forming a waveform sequence that can characterize dynamic environmental changes and the internal cognitive state of the system.

[0028] Based on a global or logical clock, a unified timestamp is added to each frame of perceived input, each set of feature waveforms, and the subsequently generated decision records. High-dimensional orthogonal features, cognitive waveform segments, and corresponding state hash values ​​are stored as evidence. Existing blockchain technology is used to put key decision-making steps on the chain, forming an immutable decision log for post-analysis, model auditing, and data playback during iterative optimization.

[0029] Step 3: Perform multi-dimensional operations based on the pre-constructed mental evolution graph network to generate a decision candidate set. Dynamically input the cognitive waveform into the pre-constructed mental evolution graph network for message passing and node state updates. Based on the similarity of node states, changes in information entropy, and decision error feedback, dynamically adjust the structure of the mental evolution graph. Dynamic adjustments include adding new nodes, removing redundant nodes, and strengthening or weakening the connection weights of edges. The evolution of the mental evolution graph yields a mental state vector, and the graph network state of the evolved mental evolution graph is encoded into a compact mental state vector.

[0030] In a mental evolution graph network, nodes represent different cognitive states, such as beliefs, intentions, and knowledge fragments, while edges represent the connections and influence paths between states.

[0031] The transfer function for state updates in the mental evolution graph network is: ; in, Let v be the state of node v at time t. Let w(t) be the set of neighboring nodes, and w(t) be the cognitive waveform input at time t.

[0032] During dynamic adjustments, at each mental evolution cycle (or every T time steps), all node pairs in the mental evolution graph network are calculated. State similarity. Node state vector Taken from the updated value after message passing, the similarity metric uses the negative of the existing cosine similarity or Euclidean distance, and the similarity is expressed as: ; in, Let i be the updated value for node i. This is the updated value for node j.

[0033] The node aggregation process is as follows: Set the aggregation threshold. For example, 0.85. If And the edge weight between the two nodes If it is, then it is determined to be a redundant node. Create a new node. Its state vector is: ; Will and All adjacent edges are migrated to The edge weights are taken as the maximum or average of the original weights. (Delete) and Update the node set V and the edge set E.

[0034] The information entropy of a node is calculated by taking into account its historical activation frequency for each node v. Calculate information entropy, and the historical activation frequency is the frequency in the most recent The number of times a message is passed within a period is represented by the information entropy as: ; in, ,like This indicates that the node is either almost never active or always active. If the threshold is set, then the node has low information content.

[0035] A gradient-based method is used to estimate the contribution of each node to the final decision output. After each decision is generated, the error signal is backpropagated to the graph network nodes, and the cumulative absolute value of the node gradients is recorded. The contribution imp(v) is the weighted sum of the cumulative gradient and the node activation frequency.

[0036] Pruning is performed when the following adjustments are met; a pruning threshold is set. If the node satisfies and If a node is found to be redundant, it is deleted along with all its edges. If deletion results in a disconnected graph, nodes on the critical connecting paths are retained.

[0037] Through error feedback in subsequent step 5, the correlation strength between nodes is strengthened or weakened. The error is decomposed into the error contribution of each node and each edge through backpropagation (using the automatic differentiation mechanism of graph neural networks). The weight of each edge is updated according to the following formula: ; in, The edge weight learning rate is sgn(), the sign function is sgn(), and a gradient-based optimizer (such as Adam) is used for stable updates.

[0038] Introducing supplementary rules from Hebbian learning: if two nodes are simultaneously highly activated during the decision-making process (the product of their activation values ​​exceeds a threshold), then the connection between them is strengthened; if their activation patterns are inconsistent, then the connection is weakened.

[0039] ; in, , α is the node activation value, α is the learning rate, and β is the decay factor.

[0040] When two nodes exhibit high co-activation in multiple decisions and are currently unconnected, add an edge with an initial weight set to a medium value.

[0041] If the weight of a certain side If the edge remains below the threshold, it is deleted, indicating the fading of the connection between concepts.

[0042] The evolved dynamic graph network is compressed into a compact vector of fixed dimensions for use by the decision decoder in step 4. The encoding process is as follows: Hierarchical graph pooling methods are used to aggregate node states into a global graph representation, such as local average / max pooling: averaging or maximizing the values ​​of all node states to obtain a vector. Calculate the importance weight of each node. Where q is a learnable query vector, and then a weighted sum is obtained. The topological features of the graph (such as eigenvalues ​​of the adjacency matrix, node degree distribution, clustering coefficients, etc.) are transformed into low-dimensional vectors through an auxiliary encoder. To preserve the temporal continuity of mental evolution, a graph representation of multiple consecutive cycles is input into a temporal encoder, such as a GRU, which generates a vector containing the evolutionary trend. The final mental state vector is obtained as follows: ; in, This represents vector concatenation. MLP stands for Multilayer Perceptron, which maps the concatenated high-dimensional vector to a compact mental state vector z of a predetermined dimension (e.g., 128 dimensions).

[0043] Step 4: Input the mental state vector into the decision decoder to generate candidate decision schemes. Candidate decision schemes are forcibly intercepted and input into a serial three-layer security decision-making mechanism for verification: First, a red line test of the safety circuit breaker is performed. If it fails, the candidate decision-making scheme is discarded directly. If the security circuit breaker layer is passed, the system will proceed to the rule verification layer for compliance testing. If the system fails to pass the compliance test, the solution will be rejected. If the rule validation layer is passed, the system will proceed to the utility optimization layer for comprehensive utility scoring. Only after a candidate decision-making scheme passes the safety circuit breaker layer and the rule verification layer is the utility-optimized decision allowed to be mapped to a specific output. The safety circuit breaker layer's judgment is like matching a "life safety / ethical red line library." If a red line is hit (such as harming humans or violating the law), the output path is directly cut off physically or logically, the instruction is discarded, and it does not proceed to subsequent levels. The rule verification layer can test whether the instruction conforms to social rules, environmental order, or task constraints. Specific outputs include actions, control variables, and strategy parameters.

[0044] The decision with the highest confidence level is output as the execution instruction, while alternative decisions and their confidence level information are retained for subsequent evaluation. The generated decision instructions are sent to the execution mechanism (such as robot controller, simulation environment API, human-computer interaction interface) for actual execution.

[0045] When making a decision, a confidence score is calculated for each of the multiple candidate decision schemes generated, using the activation level of nodes and the transit confidence of edges in the mental evolution graph network. The candidate decision scheme with the highest confidence score is then used as the decision output.

[0046] Based on the mental evolution graph network, the matching degree between each candidate decision and the current graph state is calculated. The candidate decision is used as temporary input, and a forward inference is performed on the mental graph network. The changes in node activation patterns caused by the decision are measured, and the confidence level is then calculated. Defined as: ; in, For node v and decision The relevance score is calculated using existing attention mechanisms. σ represents the node importance weight, and σ is the sigmoid function to normalize the confidence score to the [0,1] interval; Set up a lightweight risk assessment network Where c is the scenario constraint vector, which includes parameters such as safety boundaries, resource limitations, and ethical norms. A lightweight risk assessment network outputs a risk score for each candidate decision. The overall utility of each candidate decision is calculated and expressed as: ; Wherein, λ is the risk aversion coefficient, which is dynamically adjusted according to the scenario, such as increasing λ in high-risk scenarios; The optimal candidate decision scheme is output based on the overall utility, and is expressed as: .

[0047] If the confidence levels of multiple candidate decisions are similar, it indicates that the mental state is ambiguous. In this case, the confidence threshold needs to be lowered or additional information collection triggered, and the confidence variance should be output. As a supplementary indicator of decision reliability.

[0048] Check the consistency between candidate decisions and historical decision trajectories. If the current decision with the highest confidence level seriously conflicts with the recent decision sequence, a consistency correction is triggered to increase the confidence level of candidate decisions that are consistent with historical decision patterns.

[0049] The lightweight risk assessment network includes: The feature fusion layer projects multiple inputs through independent linear layers to a set dimension, then sums them, and activates them with ReLU. For example, if the input dimension is... The dimension is set to 128; The risk setting layer sets up risk heads for different risk modes in multiple parallel paths. Each risk head consists of two MLP layers (128→64→32). The risk aggregation layer concatenates the outputs of the three risk heads, and then passes them through an MLP layer (96→32→1) plus a Sigmoid function to output the final risk score.

[0050] Step 5: Evaluate the effectiveness of the decision implementation after the decision output, and use the evaluation results as feedback signals.

[0051] In the next perception cycle after the decision is executed or after a predefined waiting time, new environmental state data is collected as the effect of the decision execution, and the environmental state data is normalized. The environmental state includes task completion indicators, reward signals, safety constraint satisfaction status, and resource consumption. By aggregating multiple environmental state data in a weighted sum manner, the actual assessment value R is obtained, expressed as: ; in, Assigning weights to each environmental state data point. and These are the penalty coefficients for safety and cost, respectively. Task completion rate indicators For safety penalties, cost represents the amount of resources consumed; Based on the actual assessment value and the obtained expected value The error is expressed as: ; Using the error δ as the starting point of the gradient of the loss function, the algorithm propagates backward along the system computation graph. For the gradient information required for the evolution of the graph structure, the contribution of each edge to the final evaluation value is obtained by using the policy gradient or by the automatic differentiation of the graph neural network.

[0052] Expected values ​​can be obtained from the following sources: If the system adopts a supervised learning paradigm, the expected value is the labeled expected result.

[0053] If the system adopts a reinforcement learning paradigm, the expected value is the state value output by the value network.

[0054] If the system adopts a self-supervised approach, the expected value can be the self-evaluation value predicted by the model in the preceding steps.

[0055] If it is determined that the decision error δ is consistently positive in multiple consecutive periods and the error mainly originates from a certain type of unmodeled cognitive pattern, then a new node is added to the mental evolution graph network for refinement, and the state of the new node is initialized with the feature component in the current mental state vector that is positively correlated with the error. If a node consistently contributes a small gradient or has extremely low activation in multiple error feedbacks, then a pruning operation is triggered. The edge weights are updated based on the gradient of the error with respect to each edge, as follows: ; in, The learning rate is set as the edge weight, and combined with Heblin learning, the connection between two nodes is strengthened when they are both highly activated in successful decisions, and weakened when they are both activated in failed decisions. When an error signal indicates that the current cognitive structure cannot effectively handle a certain type of scenario, a topology reconnection operation is triggered.

[0056] Step 6: Iterate through steps 1 to 5 based on the feedback signal, specifically as follows: The evolved mental evolution graph network state is directly used as the initial state for the next round of iteration; After each iteration, check whether the number of nodes and edges in the mental evolution graph network structure exceeds the preset limit. If it does, perform global pruning. At the same time, check for gradient explosion or vanishing phenomena and perform gradient pruning if necessary.

[0057] Compared with existing technologies, the solution in this embodiment demonstrates higher decision-making accuracy, faster scene adaptation speed, stronger security risk avoidance capability, and better computational efficiency in complex tasks with multiple scenarios, high dynamism, and strong coupling. Based on mind maps, it constructs an intelligent decision-making system integrating high-dimensional orthogonal feature extraction, dynamic evolution of mind evolution graph networks, and closed-loop iterative optimization. It achieves significant technical effects in multiple technical dimensions such as feature representation capability, model adaptability, decision quality, and system efficiency, reduces scene switching costs, and greatly improves the security of autonomously generated decisions through the chain of data collection-decision generation-circuit breaker review. It supports continuous learning and knowledge accumulation, and provides an innovative solution for the widespread application of intelligent decision-making technology in complex and ever-changing scenarios. Example 2

[0058] A multi-scenario intelligent decision-making system based on mental modeling includes a multimodal sensor, a memory, and a processor. The multimodal sensor is used to collect multimodal environmental perception data and transmit it to the processor. The memory stores computer-executable instructions, and the processor is used to execute the executable instructions stored in the memory. When the executable instructions are executed by the processor, they are executed according to the steps of the method described above.

[0059] The above descriptions are merely embodiments of the present invention, and common knowledge regarding specific structures and characteristics is not elaborated upon here. It should be noted that those skilled in the art can make various modifications and improvements without departing from the structure of the present invention, and these should also be considered within the scope of protection of the present invention. These modifications and improvements will not affect the effectiveness of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A multi-scenario intelligent decision-making method based on mental modeling, comprising the following steps: Step 1: Acquire multimodal environment perception data, perform data alignment and preprocessing on the multimodal environment perception data, and generate perception input vectors; Its characteristic is that it further includes: Step 2: Map the perceptual input vector to a high-dimensional orthogonal feature space, extract the high-dimensional orthogonal features, and map the high-dimensional orthogonal features into a continuous cognitive waveform signal. The cognitive waveform is encoded in a multi-channel time-frequency domain, with each channel corresponding to a class of orthogonal feature components. Step 3: Perform multi-dimensional operations based on the pre-constructed mental evolution graph network to generate a decision candidate set. Dynamically input the cognitive waveform into the pre-constructed mental evolution graph network to perform message passing and node state updates. Based on the similarity of node states, changes in information entropy, and decision error feedback, dynamically adjust the structure of the mental evolution graph. The mental state vector is obtained from the evolution of the mental evolution graph. Step 4: Input the mental state vector into the decision decoder to generate candidate decision schemes, forcibly intercept the candidate decision schemes, and input them into the serial three-layer security decision mechanism for verification. First, a red line test of the safety circuit breaker is performed. If it fails, the candidate decision-making scheme is discarded directly. If it passes, it will proceed to the rule verification layer for compliance testing; if it fails, the proposal will be rejected. If it passes, it will proceed to the utility optimization layer for scoring; Only after a candidate decision solution passes the security circuit breaker layer and the rule verification layer is it allowed to map the utility-optimized decision to a specific output. Step 5: Evaluate the effectiveness of the decision implementation after the decision output, and use the evaluation results as feedback signals; Step 6: Iterate through steps 1 to 5 based on the feedback signal.

2. The multi-scenario intelligent decision-making method based on mental modeling according to claim 1, characterized in that: In step 2, the multimodal environment perception data includes visual images, radar point clouds, and voice commands.

3. The multi-scenario intelligent decision-making method based on mental modeling according to claim 2, characterized in that: The data alignment uses a timestamp synchronization mechanism to map heterogeneous data into a unified vector space.

4. The multi-scenario intelligent decision-making method based on mental modeling according to claim 2, characterized in that: The red line detection of the safety fuse layer includes physical damage detection and ethical red line detection.

5. The multi-scenario intelligent decision-making method based on mental modeling according to claim 4, characterized in that: In step 4, the matching degree between each candidate decision and the current graph state is calculated based on the mental evolution graph network. The candidate decision is used as temporary input, and a forward inference is performed on the mental graph network to measure the change in node activation mode caused by the decision. The confidence level is defined as: ； in, For node v and decision The relevance score, Let σ be the node importance weight, and σ be the sigmoid function. Set up a lightweight risk assessment network Where c is the scenario constraint vector, and a lightweight risk assessment network outputs a risk score for each candidate decision. The overall utility of each candidate decision is calculated and expressed as: ； Where λ is the risk aversion coefficient; The optimal candidate decision scheme is output based on the overall utility, and is expressed as: 。 6. The multi-scenario intelligent decision-making method based on mental modeling according to claim 5, characterized in that: In step 4, the lightweight risk assessment network includes: The feature fusion layer projects multiple inputs through independent linear layers to a set dimension and then adds them together, activated by ReLU. The risk setting layer sets up risk heads for different risk modes that run in parallel. Each risk head consists of two MLP layers. The risk aggregation layer concatenates the outputs of the three risk heads, and then outputs the final risk score through an MLP layer plus a sigmoid function.

7. The multi-scenario intelligent decision-making method based on mental modeling according to claim 6, characterized in that: In step 5, new environmental state data is collected as the effect of decision execution in the next perception cycle after the decision is executed or after a predefined waiting time. The environmental state data is then normalized. The environmental state includes task completion indicators, reward signals, safety constraint satisfaction, and resource consumption. By aggregating multiple environmental state data in a weighted sum manner, the actual assessment value R is obtained, expressed as: ； in, Assigning weights to each environmental state data point. and These are the penalty coefficients for safety and cost, respectively. Task completion rate indicators For safety reasons, This refers to resource consumption. Based on the actual assessment value and the obtained expected value The error is expressed as: ； Error As the starting point of the gradient of the loss function, it propagates backward along the computational graph of the system. For the gradient information required for the evolution of the graph structure, the contribution of each edge to the final evaluation value is obtained by using the policy gradient or by the automatic differentiation of the graph neural network.

8. The multi-scenario intelligent decision-making method based on mental modeling according to claim 7, characterized in that: In step 5, it is determined that the decision error occurs over multiple consecutive periods. If the result is consistently positive and the error mainly originates from a certain type of unmodeled cognitive pattern, then a new node is added to the mental evolution graph network for refinement, and the state of the new node is initialized with the feature component in the current mental state vector that is positively correlated with the error. Decision node in multiple errors If the feedback consistently contributes a tiny gradient or has extremely low activation, then pruning is triggered. The edge weights are updated based on the gradient of the error with respect to each edge, as follows: ； in, The learning rate is set as the edge weight, and combined with Heblin learning, the connection between two nodes is strengthened when they are both highly activated in successful decisions, and weakened when they are both activated in failed decisions. When an error signal indicates that the current cognitive structure cannot effectively handle a certain type of scenario, a topology reconnection operation is triggered.

9. The multi-scenario intelligent decision-making method based on mental modeling according to claim 8, characterized in that: In step 6, the evolved mental evolution graph network state is directly used as the initial state for the next round of iteration. After each iteration, check whether the number of nodes and edges in the mental evolution graph network structure exceeds the preset limit. If it does, perform global pruning. At the same time, check for gradient explosion or vanishing phenomena and perform gradient pruning if necessary.

10. A multi-scenario intelligent decision-making system based on mental modeling, comprising a multimodal sensor, a memory, and a processor, wherein the multimodal sensor is used to collect multimodal environmental perception data and transmit it to the processor, the memory stores computer-executable instructions, and the processor is used to execute the executable instructions stored in the memory; Its features are: When the executable instructions are executed by the processor, they are performed in accordance with the steps of the method as described in any one of claims 1-9.

Images