An unmanned intelligent decision-making method based on pre-trained models

Abstract

This invention discloses an unmanned intelligent decision-making method based on a pre-trained model. In the field of artificial intelligence technology, this invention addresses three core problems encountered when deploying ultra-large-scale pre-trained models in unmanned systems: uncontrollable decision-making due to alignment bounce, cascading amplification of hallucination outputs in the closed-loop decision chain, and broken causal reasoning chains in long-term tasks. A three-pronged closed-loop decision-making architecture is proposed. Through a dynamic constraint mechanism using alignment elastic tensors, the Mahalanobis distance between the model's hidden states and the safe alignment distribution is calculated in real-time during the inference phase, triggering Logit rescaling to suppress the probability of generating dangerous actions. A knowledge graph-based hallucination propagation blocking mechanism extracts action triples from candidate instructions and matches them with the ontology subgraph, performing factual consistency verification before execution. Finally, a causal state graph persistent reasoning architecture maintains the global causal memory of decisions in the form of a directed acyclic graph.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to an unmanned intelligent decision-making method based on a pre-trained model. Background Technology

[0002] With the rapid development of ultra-large-scale pre-trained language models and visual language models represented by GPT series and Grok, artificial intelligence agents have become a core development direction recognized by academia and industry. Pre-trained models are evolving from single dialogue engines to embodied intelligent decision-making brains for unmanned systems such as drones, ground robots, underwater unmanned platforms, and autonomous vehicles.

[0003] The basic architecture of large models is based on Transformer, and its core mechanism is multi-head self-attention mechanism. When the model is inferring, it propagates forward layer by layer according to the input context, and finally generates the probability distribution of the next token in the Decoder output layer and samples it to obtain the output.

[0004] The closest existing technical solution to this invention is a combination of a large model as a task planner for unmanned systems, which combines a thought chain approach with enhanced retrieval and generation, plus prompt word constraints. Its typical workflow is as follows: the unmanned system's sensors collect environmental information and, after preprocessing, concatenate historical dialogue records or knowledge fragments obtained from vector retrieval into the prompt word context; the large model generates a reasoning process step by step in a thought chain manner, and finally outputs action command text; the unmanned system's underlying control module executes the command and returns the execution result as a new perceptual input, forming a closed loop.

[0005] This solution has three core flaws: First, alignment bounce leads to uncontrollable decision-making behavior. After training and aligning large models based on reinforcement learning and human feedback, when encountering complex and extreme long-tail scenarios of unmanned systems, the input distribution deviates from the coverage of the alignment dataset, causing the probability distribution of the model output to drift in the opposite direction to the dangerous distribution in the pre-training stage, generating dangerous decision instructions that break through physical constraints or ethical boundaries. Existing technologies cannot intervene in real time during the inference stage.

[0006] Second, the illusory output is amplified in a cascading manner in the closed-loop decision chain. The output of the large model is essentially a probability sampling, which has inherent uncertainty. In the perception-cognition-execution closed loop of the unmanned system, after an erroneous action decision is executed by the underlying controller, the execution result changes the real environmental state. After being re-perceived by the sensor, it enters the next round of reasoning as the real input, causing the error to be amplified round by round, eventually leading to systemic and catastrophic cascading failure. Existing technologies lack an online verification mechanism for the factual consistency of the action chain.

[0007] Third, in long-term complex tasks, causal reasoning breaks down. Large models have limited effective context windows, and the linearly concatenated historical dialogue records exceed the window length as the number of steps increases, resulting in the truncation and loss of early key information. Although semantic retrieval from vector databases can recall relevant information, it cannot retain the ordered causal relationships between action steps. When the model executes the Nth step, it cannot effectively associate the initial task objective with the causal constraints established in the intermediate steps, leading to target drift or logical breaks, causing long-term tasks to be abandoned halfway or fall into meaningless loops. Summary of the Invention

[0008] This invention aims to overcome the aforementioned deficiencies of existing technologies and provide an unmanned intelligent decision-making method based on a pre-trained model. By organically integrating the alignment elastic tensor dynamic constraint mechanism, the illusion propagation blocking mechanism based on knowledge graphs, and the causal state graph persistent reasoning architecture, a closed-loop decision-making architecture integrating reasoning security, output verification, and memory causality is formed. Thus, without modifying the original weights of the pre-trained model, real-time security intervention in the reasoning stage, illusion blocking before execution, and global causal consistency maintenance across long time-series steps are achieved.

[0009] The technical problems to be solved by this invention are: first, how to detect and intervene in alignment bounce in real time during the inference stage without retraining or fine-tuning the pre-trained model, so as to prevent dangerous decision commands from being generated; second, how to verify the factual consistency of the large model output before the execution of action commands of unmanned systems, so as to block the propagation of illusions to subsequent decision chains from the source; and third, how to maintain structured decision state memory and causal anchoring mechanism in long-term tasks, maintain global causal consistency, and prevent target drift and logical chain breakage.

[0010] The technical solution adopted in this invention is as follows: An unmanned intelligent decision-making method based on a pre-trained model, running between the perception sensor and the underlying motion controller as a safety intelligent decision-making middleware, includes the following steps: The system receives raw data from multimodal sensors of unmanned systems, and after processing such as timestamp alignment and fusion, visual target detection, state vector extraction, and spatial geometry estimation of passable areas, it generates structured perception summary text, which serves as the perception-side input for large model inference. During task initialization, target anchor nodes are created with the initial task objective, the causal state graph is initialized, and before the start of each decision loop, the current complete causal state graph is serialized into structured text, which together with the perceptual summary constructs a complete reasoning context. During the forward propagation of the pre-trained large model, the current hidden state vector is extracted in real time from the output of the pre-specified Decoder layer. Based on the pre-stored safe distribution center vector and the inverse matrix of the alignment elastic tensor, the Mahalanobis distance of the current hidden state vector relative to the safe alignment distribution is calculated. When the distance exceeds the preset safety threshold, a punitive rescaling is performed on the positions in the original Logit vector of the output layer that belong to the dangerous token set to suppress the generation probability of dangerous actions, generate a safe Logit vector, and sample the output. Action triples are extracted from the action candidate command text output by the large model. Using the action triples and the current system state parameters as query conditions, subgraph matching is performed in the pre-built unmanned system ontology knowledge graph to retrieve relevant constraint rules. If there is a constraint conflict, it is determined to be a hallucination action. The conflict constraint details are injected into the prompt word constraint area, forcing the large model to regenerate candidate commands and verify them in a loop. When the maximum number of retries is exceeded, a safety policy is triggered. If there is no constraint conflict, a legal action command is output. The legal action command is sent to the underlying controller for execution. Based on the execution feedback, decision nodes and status nodes are created and written into the causal state graph. The causal state graph is driven to continuously extend the topology along the task progress direction. It is determined whether the task is completed. If it is not completed, it returns to the perception and acquisition step and starts the next decision loop.

[0011] The offline construction method for the aligned elastic tensor includes: Several safe and legal decision samples are collected from the unmanned system safety operation procedures, simulation verification data, and legally labeled mission trajectories to form a safety alignment dataset. Each sample in the safety alignment dataset is fed into a pre-trained large model, and the corresponding hidden state vector is extracted at the output of a specified Decoder layer to form a set of safe hidden state vectors. The mean of the set of safe hidden state vectors is taken to obtain the safety distribution center vector. The covariance matrix of the set of safe hidden state vectors, i.e., the alignment elasticity tensor, is calculated. The safety distribution center vector and the inverse matrix of the alignment elasticity tensor are persistently stored in a lightweight Adapter layer inserted at the output of a specified Decoder layer.

[0012] The method for calculating the distance to Maharanobis is as follows: ; in, The distance to Maharanobis is a scalar value. For the current reasoning step in the th... The hidden state vector extracted from the output of the layer decoder. For the safe distribution center vector, The inverse matrix of the alignment elastic tensor is the precision matrix. For the hidden layer dimension.

[0013] The calculation method for Logit rescaling is as follows: ; in, For the first The security Logit value of each token after rescaling. This is the original Logit value. To punish the intensity overparameter, For the first The danger mask value for each token is 1 if the token belongs to the set of dangerous tokens determined by the dangerous action dictionary, and 0 otherwise. The Mahalanobis distance calculated for the current step.

[0014] The unmanned system ontology library is organized in the form of a knowledge graph, containing three types of constraint knowledge nodes: physical law constraint nodes, sensor limit nodes, and equipment mechanical constraint nodes. All constraints are stored in the form of triples of preconditions and implied relationships between prohibited or permitted actions. The format of the action triple is (subject, action, object). The subgraph matching rule is: if there is a node in the ontology library that satisfies its preconditions and current state parameters, and its restricted actions match the action type and action object semantics of the current action triple, then a constraint conflict is determined to exist.

[0015] The causal state graph is a directed acyclic graph data structure, consisting of four types of elements: target anchor nodes, decision nodes, state nodes, and causal edges. The target anchor node represents the initial task objective and serves as the root node of the entire causal state graph. The decision node represents a legitimate sub-decision generated by the large model at a certain moment, and its attributes include a time stamp, a perception summary snapshot, action instruction text, and a summary of the decision basis. The state node represents the new state reached by the environment after performing an action, and its attributes include a time stamp, a description of the execution result, and changes in key state parameters. The causal edges are directed edges that connect adjacent decision nodes and state nodes, representing the causal relationship between the execution of an action and the resulting state change. During the serialization of the causal state graph, the summaries of each node and the edge relationships are output in causal time sequence and injected into the large model's inference context in the form of structured text. Mandatory instructions are attached to constrain all decisions of the large model to extend the topology based on the current end state node of the causal state graph and to maintain global consistency with the task objective of the target anchor node.

[0016] The beneficial effects of this invention are as follows: 1. This invention uses an aligned elastic tensor dynamic constraint mechanism to intercept dangerous decisions in real time within millisecond-level inference latency, without the need for model readjustment, effectively ensuring the physical and ethical safety boundaries of unmanned decision-making; 2. This invention uses a knowledge graph-based illusion propagation blocking mechanism to immediately verify fact consistency after the output of a single sub-decision and before its execution, thus intercepting the illusion before execution and blocking the cascading failure chain from the source. 3. This invention maintains the global decision-making causal memory network through a causal state graph persistent reasoning architecture in a highly compressed form of structured graph summarization. It maintains global causal consistency in long-term tasks spanning dozens or even hundreds of steps, which is significantly better than the traditional method of linearly splicing historical dialogues. Attached Figure Description

[0017] Figure 1 This is a data flow diagram of the overall system architecture of the present invention; Figure 2 Flowchart of the preprocessing module for perceived information; Figure 3 Workflow diagram for the inference phase of the alignment elastic tensor dynamic constraint module; Figure 4 Workflow diagram for the hallucination blocking and confidence calibration module; Figure 5 This is a flowchart for the dynamic updating and injection of the causal state diagram; Figure 6 This is a complete workflow diagram of the entire process of this invention. Detailed Implementation

[0018] The following is in conjunction with the appendix Figure 1 - Appendix Figure 6 The specific embodiments of the present invention will be further described below. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention.

[0019] Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0020] The unmanned intelligent decision-making method proposed in this invention runs on a software system architecture. This architecture is located between the perception sensors and the underlying motion controller of the unmanned system, and acts as a middleware for safe intelligent decision-making. The system consists of five core modules: a perception information preprocessing module, an alignment elastic tensor dynamic constraint module, an illusion blocking and confidence calibration module, a causal state graph management module, and a pre-trained large model inference kernel. The specific implementation methods are described in detail below in conjunction with each module.

[0021] I. Sensing Information Preprocessing Module The perception information preprocessing module receives raw data from the multimodal sensors of the unmanned system, including camera image frame sequences, lidar point clouds, inertial measurement unit attitude data, and global positioning system positioning data.

[0022] First, the asynchronous multimodal data from various sensors are timestamped and fused to eliminate inconsistencies in the time dimension of the data from each sensor. Then, visual target detection and state vector extraction are performed in parallel: visual target detection outputs perception information such as obstacle category, position, and size; state vector extraction outputs numerical states such as the human body's position, speed, and posture.

[0023] Based on the above detection results, the spatial geometric information of the passable area is estimated. Finally, the natural language summarization generation submodule converts the above structured information into a standardized text description, such as the current location: 39.5 degrees north latitude, 116.2 degrees east longitude, 120 meters altitude; There is a building obstacle about 15 meters high 60 meters directly in front; Current battery level: 78%; Wind speed: 5 knots east of north, for use in constructing context for subsequent reasoning.

[0024] II. Alignment Elastic Tensor Dynamic Constraint Module This module is divided into two parts: an offline preprocessing stage and an online inference stage.

[0025] The offline preprocessing stage completes the construction of the alignment elastic tensor: From the safe operating procedures, simulation verification data, and expert-annotated legal mission trajectories of unmanned systems, several decision samples confirmed as safe and legal are collected to form a safety alignment dataset, denoted as . .

[0026] Will Each sample in the dataset is fed into a pre-trained large model, and at the specified... Extract the corresponding hidden state vector from the output of the layer Decoder. ,in The hidden layer dimension of this layer constitutes the set of secure hidden state vectors. , The total number of samples in the dataset for safe alignment.

[0027] Taking the mean of the set of hidden security state vectors yields the center vector of the security distribution: ; in, The center vector of the securely aligned distribution.

[0028] Calculate the covariance matrix of the set of secure hidden state vectors, i.e., the aligned elasticity tensor: ; in, This refers to the alignment elasticity tensor, which characterizes the morphology and variance structure of the secure alignment distribution in the high-dimensional hidden state space. (The superscript indicates that the tensor is not part of the main text.) This indicates the transpose operation.

[0029] Will and (Precision matrix) is persistently stored in the insertion point. Offline preprocessing is performed in the lightweight Adapter layer at the output of the Decoder layer.

[0030] Real-time safety constraint intervention during the online inference phase: During the forward propagation of the input sequence for the current decision step in a large model, from the first... The output of the layer decoder extracts the hidden state vector at the current time step in real time. .

[0031] by and Using this as a baseline, calculate the current hidden state vector. Maharanobis distance from the center of safe distribution : ; The calculated results With preset safety threshold Compare them.

[0032] Safety threshold The value is determined based on the specific safety level of the unmanned system, and is usually determined through ROC analysis on the validation set.

[0033] like If the current model output is within the safe alignment range, no intervention is triggered, forward propagation is allowed to continue normally, and the original Logit vector is directly output. like If an alignment bounce trend is detected, the Logit rescaling intervention process will be initiated immediately.

[0034] During Logit rescaling intervention, let the original Logit vector be... ,in Given the vocabulary size, the danger mask vector is... The position with a value of 1 corresponds to the set of dangerous tokens determined by the dangerous action dictionary, and the rescaled safe logit vector. Calculated according to the following rules: ; in, For the first The security Logit value of each token after rescaling. This is the original Logit value. This is a penalty intensity hyperparameter used to control the intensity of intervention. For the first The danger mask value of each token, The Mahalanobis distance calculated for the current step; the greater the distance, the heavier the penalty. After Softmax normalization, token sampling is performed to ensure that the probability of generating dangerous actions is effectively suppressed.

[0035] III. Hallucination Blocking and Confidence Calibration Module After the large model completes a sub-decision output and before the underlying controller executes, this module performs fact-consistency cross-validation on the output action instructions through a structured ontology library.

[0036] Offline construction of the ontology: The ontology is organized in the form of a knowledge graph, containing three types of constraint knowledge nodes. Physical law constraint nodes describe the physical constraints of the environment in which the unmanned system operates, such as prohibiting high-energy-consuming flight maneuvers when the battery level is below 10%; Sensor limit nodes describe the boundaries of a sensor's detection capabilities. For example, when the maximum detection range of a lidar is 200 meters, information about obstacles beyond 200 meters is unreliable. The equipment mechanical constraint node describes the kinematic and dynamic constraints of the unmanned system, such as prohibiting the generation of ascent commands exceeding the ascent rate when the maximum ascent rate is 5 meters per second.

[0037] All constraints are stored in the form of triples of preconditions and implication relations of prohibited or permitted actions, forming a structured ontology graph that can be used for fast subgraph matching.

[0038] The online hallucination detection and blocking process is as follows: Receive the raw action candidate instruction text output by the large model. Lightweight entity and relation extraction is performed on it, and it is parsed into a set of action triples. in Indicates the subject of the action. Indicates the action type. Indicates the object of the action, subscript The first one extracted in the current step The index of a triplet.

[0039] Simultaneously, current system state parameters are extracted from the current perception summary and causal state graph to form a state context. .

[0040] by Each triple and state context in Based on the query criteria, subgraph matching is performed in the ontology knowledge graph to retrieve the set of all constraint rules related to the current action. The matching rule is: if a node exists in the ontology that satisfies its preconditions and... If the current state parameters match, and the restricted action matches the action type and action object semantics of the triple, then the constraint rule is added. .

[0041] right Each triple in the set is determined to be consistent with the retrieved set of constraint rules. There is a conflict, if If the triplet is not empty and the current system state satisfies the triggering condition of the constraint, then the action corresponding to the triplet is determined to be a phantom action, the conflict constraint details are recorded, and the entire action instruction is marked as verification failed. If all triplets have no constraint conflicts, then the verification is marked as successful. Output as a valid action command.

[0042] When validation fails, the detected conflict constraint details are injected into the large model cue word constraint region as structured text, and... If marked as illegal, the large model is forced to regenerate candidate action instructions under the new constraints and re-enter the verification process until the verification is passed or the maximum number of retries is reached.

[0043] When the maximum number of retries is exceeded, a safety hover, shutdown, or return-to-home action is triggered, an alarm log is reported, and manual takeover is awaited to ensure that the system maintains physical safety under any circumstances.

[0044] IV. Cause-and-Effect Diagram Management Module The causal state graph is a directed acyclic graph data structure used to persistently record all decision-making causal relationships of an unmanned system during the execution of long-term tasks. It consists of four types of elements: target anchor nodes, decision nodes, state nodes, and causal edges.

[0045] The target anchor node represents the initial task objective and serves as the root node of the entire causal state graph. All subsequent decision nodes can be traced back to this node, ensuring the consistency of the global objective. The decision node represents a legitimate sub-decision generated by the large model at a certain moment. Its attributes include the time identifier, perception summary snapshot, action instruction text, and decision basis summary. The state node represents the new state reached by the environment after performing an action. Its attributes include the time identifier, execution result description, and changes in key state parameters. Causal edges are directed edges that connect adjacent decision nodes and state nodes, representing the causal relationship between the execution of an action and the state change. Their attributes include causal type and edge weight.

[0046] The dynamic update process of the causal state graph is as follows: During task initialization, a target anchor node is created as the root node of the causal state graph based on the initial task objective, and initialization is completed. Whenever the hallucination blocking and confidence calibration module outputs a legal action instruction, key information is extracted from the instruction and the current perception summary to create a new decision node, and a directed causal edge is added from the end state node of the current causal state graph to the decision node. When the underlying controller executes the action instruction and returns the execution result feedback, a new state node is created based on the feedback information, and a directed causal edge is added from the decision node to the new state node, completing the writing of a complete decision, execution, and state causal chain node group. The causal state graph continues to extend its topology along the task advancement direction.

[0047] The serialization injection and contextual constraint method for causal state graphs is as follows: Before the start of each new decision loop, the current complete causal state graph is serialized into structured text in causal time sequence. The format includes the task target description of the target anchor node, the action description and state description of each decision node and state node, and the advancement relationship between nodes represented by each causal edge. A causal state graph context string is generated and injected into the complete inference context of the large model as a fixed component of the system prompt words. It is located before the perception summary and a mandatory instruction is attached to constrain all decisions of the large model to be topologically extended based on the end state node of the current causal state graph and to maintain global consistency with the task target of the target anchor node. It is forbidden to generate action instructions that contradict the recorded causal chain. Since the causal state graph is injected in the form of a structured graph summary rather than the original dialogue history, the occupation of the context window is highly compressed, and the complete causal memory of dozens to hundreds of steps can be maintained within a limited context budget.

[0048] V. The complete implementation process, taking UAV reconnaissance mission as an example The complete operation process is illustrated by taking a drone as an example to illustrate the task of heading to target point A and completing a designated reconnaissance mission.

[0049] During task initialization, the causal state graph management module creates target anchor nodes, records task objectives, and completes the causal state graph initialization.

[0050] First Decision Loop: The perception information preprocessing module collects data from cameras, LiDAR, inertial measurement units, and GPS to generate a structured perception summary, such as describing the current location coordinates, altitude, information about obstacles ahead, and current battery level. The perception summary is then combined with the serialized causal state graph context (initially containing only the target anchor node) to construct a complete inference context, which is then fed into the large model. The alignment elastic tensor dynamic constraint module extracts the Lth layer hidden state vector during forward propagation and calculates the Mahalanobis distance. If the distance does not exceed the safety threshold, the original Logit is output. Candidate action commands, such as takeoff to an altitude of 100 meters, are sampled. The illusion blocking and confidence calibration module extracts the action triple (drone, takeoff to, altitude 100 meters) and performs subgraph matching in the ontology library, combining parameters such as the current battery level of 96% and the maximum flight altitude limit. If no constraint conflict is found, a valid action command is output. The underlying controller executes the takeoff action and returns the execution result feedback (altitude 100 meters, battery level 96%). The causal state graph management module creates decision nodes and state nodes and writes them into the causal state graph, adds causal edges, and completes the first loop.

[0051] Subsequent decision loop: As the steps progress, the causal state graph continues to extend. When the large model generates a flight over a valley action with only 5% battery remaining in a certain step, the illusion blocking and confidence calibration module detects a conflict between the constraint rule in the ontology library that prohibits high-energy-consuming flight actions when the battery is below 10% and the current action triplet through subgraph matching. It is determined to be an illusion action, and the conflict details are injected into the prompt word constraint area. The large model is forced to regenerate a legal candidate command, such as landing at the nearest location and requesting charging support. After verification, it is executed.

[0052] Meanwhile, if the alignment elastic tensor dynamic constraint module detects that the Mahalanobis distance exceeds the safety threshold at a certain step, that is, if the alignment rebound trend is detected, it immediately performs punitive rescaling on the positions in the output layer Logit vector that belong to the dangerous token set, effectively suppressing the generation probability of dangerous actions that break through physical constraints, and ensuring that the sampled candidate instructions are within the safe range.

[0053] VI. Description of Alternative Implementation Methods For the alignment offset metric in the alignment elastic tensor dynamic constraint module, an alternative based on an energy function can be adopted: train a lightweight energy network offline, use the hidden state of the secure alignment data as low-energy samples and the data outside the alignment as high-energy samples for comparative learning training, calculate the energy value of the current hidden state in real time during online inference, and use the energy value to replace the Mahalanobis distance as the metric for alignment offset. Intervention is also achieved through Logit rescaling, which is suitable for scenarios where the distribution of secure alignment data is highly nonlinear.

[0054] For causal state management in the causal state graph management module, an alternative based on a traditional rule-based symbolic reasoning engine can be adopted: the task objectives, action prerequisites, and action effects of the unmanned system are formally defined in the planning domain definition language format, the symbolic planning engine maintains the global state machine, the state machine is updated after each decision, and the key state quantities of the current state machine are injected into the large model prompt words as structured text. This is suitable for semi-structured unmanned task scenarios where the task structure is relatively regular and the state space can be fully defined in advance.

[0055] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of the present invention. Therefore, any modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the protection scope defined by the claims of the present invention.

Claims

1. An unmanned intelligent decision-making method based on a pre-trained model, operating between the perception sensors and the underlying motion controller of an unmanned system, characterized in that, Includes the following steps: The system receives raw data from multimodal sensors of unmanned systems, and after processing such as timestamp alignment and fusion, visual target detection, state vector extraction, and spatial geometric estimation of passable areas, it generates structured perception summary text. During task initialization, target anchor nodes are created with the initial task objective, the causal state graph is initialized, and before the start of each decision loop, the current complete causal state graph is serialized into structured text, which together with the perceptual summary text constructs a complete reasoning context and is injected into the pre-trained large model. During the forward propagation of the pre-trained large model, the current hidden state vector is extracted in real time from the output of the pre-specified Decoder layer. Based on the pre-stored safe distribution center vector and the inverse matrix of the alignment elastic tensor, the Mahalanobis distance of the current hidden state vector relative to the safe alignment distribution is calculated. When the Mahalanobis distance exceeds the preset safety threshold, the positions in the original Logit vector of the output layer that belong to the dangerous token set are penalized and rescaled. After generating a safe Logit vector, token sampling is performed, and the action candidate instruction text is output. Action triples are extracted from candidate action command texts. Using action triples and current system state parameters as query conditions, subgraph matching is performed in the pre-built unmanned system ontology knowledge graph to retrieve relevant constraint rules. When there is a constraint conflict, it is determined to be a hallucination action. The conflict constraint details are injected into the prompt word constraint area, and the pre-trained large model is forced to regenerate candidate commands and verify them in a loop. When there is no constraint conflict, a legal action command is output. When the number of loop verifications exceeds the preset maximum number of retries, a safety policy is triggered. The legal action command is sent to the underlying controller for execution. Based on the execution result feedback, decision nodes and status nodes are created and written into the causal state diagram. The causal state diagram is driven to continuously extend the topology along the task advancement direction. It is determined whether the task is completed. If it is not completed, it returns to the perception data acquisition step and starts the next decision loop.

2. The method according to claim 1, characterized in that, The method for constructing the alignment elastic tensor includes the following steps: Several decision samples that have been confirmed as safe and legal were collected from the safe operating procedures of unmanned systems, simulation verification data and legal mission trajectories annotated by experts to form a safe alignment dataset. Each sample in the secure alignment dataset is fed into a pre-trained large model, and the corresponding hidden state vector is extracted at the output of the specified Decoder layer to form a set of secure hidden state vectors. The mean of the set of hidden security states is taken to obtain the center vector of the security distribution; The covariance matrix is ​​calculated for the set of secure hidden state vectors to obtain the aligned elasticity tensor; The inverse matrix of the safe distribution center vector and the alignment elastic tensor is persistently stored in a lightweight Adapter layer that is inserted into the output of the specified Decoder layer.

3. The method according to claim 1, characterized in that, The Mahalanobis distance is calculated as follows: using the safe distribution center vector and the inverse matrix of the aligned elastic tensor as a reference, the difference vector between the hidden state vector extracted in the current inference step and the safe distribution center vector is subjected to the following operations in sequence: left multiplication by the inverse matrix of the aligned elastic tensor, inner product with the difference vector, and taking the arithmetic square root, to obtain the Mahalanobis distance in scalar form. The Logit rescaling method is as follows: for the original Logit value of each Token in the output layer Logit vector, subtract the product of the penalty intensity hyperparameter, the danger mask value of the Token, and the Mahalanobis distance to obtain the rescaled safe Logit value. The danger mask value is 1 for Tokens belonging to the dangerous Token set determined by the dangerous action dictionary, and 0 for the rest of the Tokens. The larger the Mahalanobis distance, the heavier the penalty for the dangerous Token.

4. The method according to claim 1, characterized in that, The unmanned system ontology library is organized in the form of a knowledge graph, which includes three types of constraint knowledge nodes: physical law limit nodes, sensor limit nodes, and equipment mechanical limit nodes. All constraints are stored in the form of triplet pairs of preconditions and implied relationships between prohibited or permitted actions. The format of the action triple is an ordered combination of three elements: subject, action, and object. The subgraph matching rule is: when there is a node in the ontology that satisfies its preconditions and matches the current state parameters, and its restricted action matches the action type and action object semantics of the current action triple, it is determined that there is a constraint conflict, and the conflict constraint rule is included in the search results.

5. The method according to claim 1, characterized in that, The causal state graph is a directed acyclic graph data structure, consisting of four types of elements: target anchor nodes, decision nodes, state nodes, and causal edges. The target anchor node represents the initial task objective and serves as the root node of the causal state graph; A decision node represents a valid sub-decision generated by a pre-trained large model at a certain moment. Its attributes include a moment identifier, a perception summary snapshot, action instruction text, and a summary of the decision basis. A state node represents the new state that the environment reaches after an action is performed. Its attributes include a time marker, a description of the execution result, and changes in key state parameters. Causal edges are directed edges that connect adjacent decision nodes and state nodes, representing the causal relationship where an action leads to a state change. Their attributes include causal type and edge weight. All subsequent decision nodes can be traced back to the target anchor node along the directed edge to ensure global goal consistency.

6. The method according to claim 1, characterized in that, During the causal state graph serialization injection, the causal state graph is output with node summaries and edge relationships in causal temporal order to generate a causal state graph context string. This string is injected as a fixed component of the system prompt words into the complete inference context of the pre-trained large model, located before the perception summary text. A mandatory instruction is added to the system prompt words to constrain all decisions of the pre-trained large model to extend the topology based on the current causal state graph terminal state node, and to maintain global consistency with the task objective of the target anchor node. It is prohibited to generate action instructions that contradict the recorded causal chain.

7. The method according to claim 1, characterized in that, The multimodal sensor raw data includes camera image frame sequences, LiDAR point clouds, inertial measurement unit attitude data, and GPS positioning data. The visual target detection outputs obstacle category, location, and size information. The state vector extraction outputs the position, velocity, and attitude values ​​of the human body. The method for generating structured perception summary text is to convert the above structured information into a standardized natural language text description.

8. The method according to claim 1, characterized in that, The preset safety threshold is determined by performing ROC analysis on the validation set and calibrated according to the safety level of the specific unmanned system. When the Mahalanobis distance does not exceed the preset safety threshold, the current model output is determined to be within the safe alignment range, no intervention is triggered, forward propagation is allowed to continue normally, and the original Logit vector is directly output.

9. The method according to claim 1, characterized in that, The safety guarantee strategy includes triggering the unmanned system to perform safe hovering, stopping at the nearest location, or returning to its original route, reporting an alarm log, and waiting for manual takeover; after the mission is terminated, a complete cause-and-effect state graph is persisted for subsequent review and analysis.

10. An unmanned intelligent decision-making system, characterized in that, include: The perception information preprocessing module is used to receive raw data from multimodal sensors of unmanned systems, and after processing such as timestamp alignment and fusion, visual target detection, state vector extraction and spatial geometric estimation of passable areas, it generates structured perception summary text. The Alignment Elastic Tensor Dynamic Constraint Module is used to extract the hidden state vector of a specified Decoder layer in real time during the forward propagation of a pre-trained large model, calculate the Mahalanobis distance, and perform Logit rescaling security intervention when the distance exceeds a preset safety threshold to suppress the generation probability of dangerous action tokens. The illusion blocking and confidence calibration module is used to extract action triples from the action candidate instruction text output by the large model. It performs constraint conflict detection by matching with the subgraph of the knowledge graph of the unmanned system ontology library. Instructions with conflicts are identified as illusion actions and are forced to be regenerated. Instructions without conflicts are output as legal action instructions. The causal state graph management module is used to maintain a causal state graph in the form of a directed acyclic graph with the target anchor node as the root node. Before each decision loop, it is serialized and injected into the inference context. After each decision is executed, the decision node, state node and causal edge are written into the causal state graph. Pre-trained large model inference kernel, used to generate action candidate instruction text based on complete inference context; The five modules work collaboratively in the order of data flow: the perceptual information preprocessing module, the causal state graph management module, the aligned elastic tensor dynamic constraint module, the pre-trained large model inference kernel, and the hallucination blocking and confidence calibration module, forming a closed-loop decision architecture that integrates inference security, output verification, and memory causality.

Images