Low-speed unmanned vehicle artificial intelligence decision and performance evaluation system

Abstract

The invention discloses a low-speed unmanned vehicle artificial intelligence decision and performance evaluation system, and belongs to the technical field of low-speed unmanned vehicle scheduling, and the system comprises a generation module which is used for generating a scene novelty index through monitoring the instantaneous offset of an environment parameter; the topology module is used for constructing a dynamic scene topology according to the scene novelty index, and the dynamic scene topology comprises nodes and edge weights generated based on the scene novelty index; the decision-making module is used for converting the edge weight into a decision-making rule according to the dynamic scene topology and generating a decision-making strategy; the performance calculation module is used for tracking the deviation between the vehicle motion parameter and the control instruction during the execution of the decision strategy to calculate a performance disturbance coefficient; and a topology calibration module. The method can effectively adapt to unrecorded edge scenes such as sudden congestion and severe weather superposition, improves the generalization ability of a system to complex scenes exceeding a pre-training range, and reduces decision errors caused by insufficient scene adaptation.

Description

Technical Field

[0001] This invention relates to the field of scheduling technology for low-speed unmanned vehicles, and more specifically, to an artificial intelligence decision-making and performance evaluation system for low-speed unmanned vehicles. Background Technology

[0002] With the rapid development of autonomous driving technology, low-speed unmanned vehicles are increasingly widely used in urban short-distance transportation, park commuting, and last-mile delivery. The complexity and uncertainty of their operating environment place stringent requirements on the reliability of artificial intelligence decision-making systems and performance evaluation systems. Currently, the artificial intelligence decision-making and performance evaluation systems for low-speed unmanned vehicles mainly rely on static, predefined scenario libraries for training and evaluation. By pre-setting the environmental characteristics, vehicle motion parameters, and control logic under known scenarios, decision-making strategies are generated and performance is quantitatively evaluated. This can play a certain role in scenarios with a high degree of structure and gradual environmental changes.

[0003] However, in actual operation, low-speed autonomous vehicles frequently face various unrecorded edge scenarios and complex interactive conditions, and the limitations of existing systems gradually become apparent. Due to the lack of a real-time recognition mechanism for dynamic unknown scenarios, the system cannot perceive environmental changes beyond the pre-training range in a timely manner, and the evaluation parameters are fixed for a long time, making it difficult to capture the nonlinear interaction effects between vehicles, the environment, and traffic participants during environmental changes. For example, the superimposed effects of sudden traffic congestion and rainstorms, and changes in traffic rules caused by traffic sign obstruction and temporary traffic diversion in foggy weather. These complex interaction relationships often exceed the coverage of the static scenario library, making it impossible for the system's performance evaluation results to accurately match the actual operating state of the vehicle, thus leading to decision-making errors.

[0004] Existing systems suffer from a closed-loop contradiction between real-time accuracy and generalization ability. In complex scenarios such as extreme weather or special traffic control, in order to achieve real-time output of evaluation results, the system must prioritize calling predefined logic for rapid response. This process compresses the exploration space for unknown interaction relationships, preventing the system's adaptability to new scenarios from improving dynamically. Instead, it further exacerbates the lack of generalization ability. This contradiction causes existing systems to always be passively lagging behind changes in unknown scenarios and unable to stably adapt to complex and ever-changing actual operating environments. This not only affects the scientificity and reliability of decision-making but also seriously restricts the operational safety of low-speed unmanned vehicles. Summary of the Invention

[0005] To address the problems existing in the prior art, the purpose of this invention is to provide an artificial intelligence decision-making and performance evaluation system for low-speed unmanned vehicles. This system can effectively adapt to unrecorded edge scenarios such as sudden traffic jams and severe weather, improve the system's generalization ability to complex scenarios beyond the pre-training range, and reduce decision-making errors caused by insufficient scenario adaptation.

[0006] To solve the above problems, the present invention adopts the following technical solution.

[0007] A low-speed autonomous vehicle AI decision-making and performance evaluation system includes:

[0008] The generation module is used to generate a scene novelty index by monitoring the instantaneous shift of environmental parameters;

[0009] The topology module is used to construct a dynamic scene topology based on the scene novelty index. The dynamic scene topology includes nodes and edge weights generated based on the scene novelty index.

[0010] The decision module is used to convert edge weights into decision rules and generate decision strategies based on the dynamic scene topology.

[0011] The performance calculation module is used to track the deviation between vehicle motion parameters and control commands during the execution of decision-making strategies in order to calculate performance disturbance coefficients;

[0012] The topology calibration module is used to recalibrate the topology of dynamic scenes based on the performance perturbation coefficient, and the recalibration is achieved by adjusting the weights of nodes and edges.

[0013] The scene extraction module is used to extract key topological features from the recalibrated topology and generate scene prototypes;

[0014] The prototype evolution module is used to establish a continuous evolution mechanism based on the scene prototype. It triggers topology reconstruction and updates the prototype library by comparing the similarity between the new scene prototype and the inventory prototype.

[0015] Furthermore, the construction of the dynamic scene topology based on the scene novelty index includes:

[0016] Local feature enhancement is performed based on the scene novelty index to generate weight coefficients for the local environment dimension;

[0017] The weight coefficients of the local environment dimension are mapped to high-dimensional topological nodes, and spatial dependency constraints are applied between the nodes.

[0018] Edge weights are established between the generated high-dimensional topology nodes. The edge weights are generated based on the difference mapping between the historical state changes and instantaneous offsets of the nodes.

[0019] A minimum spanning tree is constructed based on the edge weights between nodes, and the edge weights are dynamically adjusted during the generation process to form the final topology.

[0020] Perform topology feedback calibration on the final topology structure.

[0021] Furthermore, the step of converting edge weights into decision rules and generating decision strategies based on the dynamic scene topology includes:

[0022] The edge weights in the dynamic scene topology are dynamically classified to form multi-level condition groups, and the historical fluctuation range of the edge weights is segmented and mapped with the local nonlinear offset to establish an edge weight condition mapping table.

[0023] The condition rules in the edge weight condition mapping table are combined to form local decision rule groups, and the local rules are prioritized according to the topological path length between nodes and the strength of the edge weight conditions.

[0024] A strategy sequence is generated based on the sorted local decision rule groups, and the order of local rule combination is adjusted through multi-stage perturbation to form a global decision strategy;

[0025] The global decision-making strategy is constrained and validated to identify the consistency of path connectivity and edge weight conditions, and the strategy is fine-tuned and corrected through local rules.

[0026] Furthermore, the step of tracking the deviation between vehicle motion parameters and control commands during the execution of the decision-making strategy to calculate the performance disturbance coefficient includes:

[0027] The deviation between the vehicle's motion parameters and control commands is broken down into multiple motion stages, and a deviation measurement index is set for each stage.

[0028] Based on the deviation value of each motion stage, different weighting coefficients are applied to the motion state of each stage, and the disturbance coefficient is weighted according to the magnitude and trend of the deviation value.

[0029] Based on the generated disturbance coefficients for each stage, a dynamic feedback loop is constructed to adapt to environmental changes by adjusting the disturbance coefficients in real time.

[0030] Based on the adjusted perturbation coefficients, the perturbation coefficients are evaluated over a long period through global optimization.

[0031] Furthermore, the recalibration of the dynamic scene topology based on the performance perturbation coefficient includes:

[0032] During the execution of the decision-making strategy, different stages of the vehicle's movement are identified, and performance calibration indicators are set for each stage;

[0033] The motion state and control deviation at each stage are interactively analyzed, and different weighting coefficients are applied to the disturbance coefficient at each stage according to the current motion state of the vehicle.

[0034] Based on the real-time control execution of the vehicle, the control error fluctuation at each stage is analyzed, and the disturbance coefficient is adjusted.

[0035] After dynamically adjusting the phased disturbance coefficients, a global disturbance correction is implemented, and a global strategy for long-term adjustment of the disturbance coefficients is achieved through multi-period performance evaluation.

[0036] Furthermore, the recalibration achieved by adjusting node and edge weights includes:

[0037] Analyze the flow, activity, and connectivity characteristics of each node and its adjacent edges in the topology to identify the interaction relationships between nodes and edges, and assign preliminary weights to each node and its adjacent edges.

[0038] Construct an optimization function that adjusts the weights of nodes and edges based on the centrality of nodes, the transmission capacity of edges, and the execution efficiency characteristics, and sets constraints to limit abnormal weight distribution.

[0039] Based on the optimization function, the weights of nodes and edges are adjusted in real time through a local feedback mechanism;

[0040] The weight configuration of nodes and edges is adjusted within a multi-objective optimization framework using a global search algorithm.

[0041] Furthermore, the step of extracting key topological features from the recalibrated topology and generating a scene prototype includes:

[0042] Each node, edge, and its relationships in the recalibrated topology are analyzed to identify the features in the topology and to perform parametric modeling. The attributes of nodes and edges are extracted, including connectivity, topology density, link stability, and node centrality.

[0043] The extracted topological features are abstracted and categorized to form a simplified feature set, and weights are assigned to each type of feature.

[0044] A scene prototype is generated based on the categorized set of topological features.

[0045] Furthermore, the establishment of a continuous evolution mechanism based on the scenario prototype includes:

[0046] Monitor key parameters of the scenario prototype and regularly detect and evaluate changes in key parameters;

[0047] Based on the feedback, the key parameters in the scene prototype were corrected, and the connection methods between nodes or the weights of edges were adjusted.

[0048] By evaluating the evolution of multiple system cycles and combining historical data with long-term trends, optimization strategies are adjusted.

[0049] Furthermore, the step of triggering topology reconstruction and updating the prototype library by comparing the similarity between the new scene prototype and the inventory prototypes includes:

[0050] Persistent cohomology analysis is performed on the topological structure of the new scenario prototype and the inventory prototype to extract topological invariants, and difference score vectors are generated based on the Wasserstein distance to compare persistence intervals.

[0051] Using the difference score vector and historical performance perturbation coefficient, the trigger threshold is adjusted through a dynamic threshold function, and the trigger decision is output.

[0052] When the trigger decision instruction is triggered, the key topological difference region is identified based on the difference score vector, and the dynamic scene topology is reconstructed by local gradient optimization using the projection gradient descent algorithm.

[0053] Based on the difference score vector and the reconstructed topology, a persistent homology-guided spectral clustering algorithm is used to cluster and update the prototype library, merging or adding scene prototypes.

[0054] Furthermore, the step of generating a scene novelty index by monitoring the instantaneous shift of environmental parameters includes:

[0055] Continuous wavelet transform is performed on the time series data of environmental parameters to extract instantaneous amplitude and phase shift at different time scales, forming a multi-scale transient feature matrix;

[0056] The phase space is reconstructed by time delay embedding method for multi-scale transient feature matrix, a phase space point matrix is ​​constructed, and singular value decomposition is performed on the phase space point matrix to extract the left singular vectors corresponding to the first k singular values ​​as the dominant mode vectors.

[0057] Based on the dominant mode vector, the information entropy of each mode is calculated, and the information entropy is weighted and summed according to the relative contribution of singular values ​​to generate the scene novelty index.

[0058] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0059] (1) This solution can accurately capture the nonlinear interaction effects between vehicles, environment and traffic participants through dynamic scene topology recalibration, real-time evolution of scene prototypes and unknown scene recognition mechanism. It can effectively adapt to unrecorded edge scenes such as sudden congestion and severe weather, improve the system's generalization ability to complex scenes beyond the pre-training range, and reduce decision-making errors caused by insufficient scene adaptation.

[0060] (2) This scheme relies on the dynamic calculation of performance disturbance coefficient and adaptive adjustment of evaluation parameters, abandoning the long-term fixed parameter settings. It can track vehicle motion deviation and environmental changes in real time, so that the performance evaluation results can accurately match the actual operating state, avoid evaluation lag caused by parameter rigidity, provide reliable data support for decision-making strategy optimization, and improve the scientificity and adaptability of decision-making logic.

[0061] (3) This solution resolves the closed-loop contradiction between real-time response and generalization capability in existing systems by using multi-stage perturbation optimization, global search algorithm and dynamic updates of prototype library. In extreme weather, special traffic control and other scenarios, it ensures that the predefined logic responds quickly to meet real-time requirements, while retaining the exploration space for unknown interaction relationships, thus achieving a dynamic balance between real-time accuracy and generalization capability.

[0062] (4) This scheme continuously improves the scenario prototype library and decision rule system by using techniques such as local gradient optimization to reconstruct the topology and persistent coherence-guided clustering updates, thereby enhancing the system's operational stability and security in complex environments and effectively reducing the collision risk and emergency braking frequency in extreme scenarios. Attached Figure Description

[0063] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0064] Figure 1 This is a flowchart illustrating the various modules of a low-speed unmanned vehicle artificial intelligence decision-making and performance evaluation system according to the present invention. Detailed Implementation

[0065] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0066] Please see Figure 1 A low-speed autonomous vehicle artificial intelligence decision-making and performance evaluation system includes:

[0067] The generation module is used to generate a scene novelty index by monitoring the instantaneous shift of environmental parameters;

[0068] The topology module is used to construct a dynamic scene topology based on the scene novelty index. The dynamic scene topology includes nodes and edge weights generated based on the scene novelty index.

[0069] The decision module is used to convert edge weights into decision rules and generate decision strategies based on the dynamic scene topology.

[0070] The performance calculation module is used to track the deviation between vehicle motion parameters and control commands during the execution of decision-making strategies in order to calculate performance disturbance coefficients;

[0071] The topology calibration module is used to recalibrate the topology of dynamic scenes based on the performance perturbation coefficient, and the recalibration is achieved by adjusting the weights of nodes and edges.

[0072] The scene extraction module is used to extract key topological features from the recalibrated topology and generate scene prototypes;

[0073] The prototype evolution module is used to establish a continuous evolution mechanism based on the scene prototype. It triggers topology reconstruction and updates the prototype library by comparing the similarity between the new scene prototype and the inventory prototype.

[0074] The generation module also includes: performing continuous wavelet transform on the time-series data of environmental parameters, extracting instantaneous amplitude and phase shift at different time scales, and forming a multi-scale transient feature matrix;

[0075] The phase space is reconstructed by time delay embedding method for multi-scale transient feature matrix, a phase space point matrix is ​​constructed, and singular value decomposition is performed on the phase space point matrix to extract the left singular vectors corresponding to the first k singular values ​​as the dominant mode vectors.

[0076] Based on the dominant mode vector, the information entropy of each mode is calculated, and the information entropy is weighted and summed according to the relative contribution of singular values ​​to generate the scene novelty index.

[0077] In this embodiment, firstly, continuous wavelet transform is used to perform feature decomposition on the time-series data of environmental parameters. This transform can capture the instantaneous change features of the data at different time scales through the scaling and translation operations of the mother wavelet. Mathematically, the formula for continuous wavelet transform can be expressed as follows: The derivation of the formula stems from the need for multi-scale analysis of time-series data. The instantaneous shift of environmental parameters exhibits different characteristics at different time scales. The mother wavelet adjusts the analysis scale through the scaling factor 'a' and the translation factor... Conjugate operations covering the entire timeline To ensure that the complex values ​​of the transformation results can simultaneously reflect amplitude and phase information, the transformation results at different scales are finally obtained through integration, where x(t) represents the time-series data of environmental parameters; Inside It refers to the selected mother wavelet function. Represents the conjugate operation, by... The scaling and translation transformations, combined with conjugate operations, enable the extraction of multi-scale instantaneous features from time-series environmental parameter data. Here, 'a' is the scaling factor. The translation factor is... The result is a wavelet transform. Through this transform, the instantaneous amplitude and phase shift at different time scales can be extracted. These features are arranged according to the scale and time dimensions to form a multi-scale transient feature matrix.

[0078] After obtaining the multi-scale transient feature matrix, a time-delay embedding method is used for phase space reconstruction. This method maps the one-dimensional feature matrix to a high-dimensional phase space, restoring the global structure and intrinsic correlation of the dynamic changes in environmental parameters. The reconstructed phase space point matrix contains complete dynamic information of the instantaneous shifts of environmental parameters. To further extract dominant features, singular value decomposition (SVD) needs to be performed on this matrix. The formula for SVD is: The derivation logic is based on the orthogonal decomposition property of matrices. The phase space point matrix contains a large amount of redundant information, and singular value decomposition can decompose it into the product of three orthogonal matrices, where the diagonal matrix... The singular values ​​in the decomposition are arranged from largest to smallest, reflecting the contribution of each component. The left singular vectors corresponding to the top k largest singular values ​​are retained. This allows us to filter out the feature patterns that dominate the instantaneous shift in the environment, which are the dominant pattern vectors. In the formula, X is the phase space point matrix, and U is the left singular vector matrix. Let X be a singular valued diagonal matrix, and V be a right singular vector matrix. The superscript T on V represents the matrix transpose operation, which converts the column vectors of V into row vectors, allowing them to participate in matrix multiplication. This achieves the orthogonal decomposition of the original matrix X, completing the mathematical expression of decomposing the original matrix into the product of three orthogonal matrices, in order to extract the matrix's features, such as the dominant pattern vector. It is the dominant mode vector matrix consisting of the left singular vectors corresponding to the first k singular values.

[0079] Finally, the scene novelty index is calculated based on the extracted dominant pattern vectors. First, the information entropy of each dominant pattern vector is calculated. The formula for calculating information entropy is: The formula originates from the quantification logic of uncertainty in information theory. The element distribution of each dominant pattern vector reflects the dynamic characteristics of that pattern, and the probability... Let represent the normalized frequency of the j-th element in the i-th dominant pattern. The entropy value directly corresponds to the complexity and uncertainty of the pattern features. Let the information entropy be the i-th dominant pattern vector. Let be the normalized probability distribution of the j-th element in the i-th dominant mode vector, where n is the dimension of the dominant mode vector; since the contribution of different dominant modes to scene novelty is determined by their corresponding singular values, the relative contribution of each singular value needs to be calculated as a weight. ,in The weights of the i-th dominant mode vector are... For the i-th singular value, The scene novelty index is obtained by summing the information entropy of the first k singular values, multiplying the information entropy of each dominant mode by its corresponding weight, and then summing the results. The scene novelty index comprehensively reflects the novel characteristics of instantaneous shifts in environmental parameters under multi-scale dominant modes.

[0080] In some embodiments of the present invention, the topology module further includes: performing local feature enhancement based on the scene novelty index to generate weight coefficients for the local environment dimension;

[0081] The weight coefficients of the local environment dimension are mapped to high-dimensional topological nodes, and spatial dependency constraints are applied between the nodes.

[0082] Edge weights are established between the generated high-dimensional topology nodes. The edge weights are generated based on the difference mapping between the historical state changes and instantaneous offsets of the nodes.

[0083] A minimum spanning tree is constructed based on the edge weights between nodes, and the edge weights are dynamically adjusted during the generation process to form the final topology.

[0084] Perform topology feedback calibration on the final topology structure.

[0085] In this embodiment, local feature enhancement is performed first. The scene novelty index reflects the novelty of the environment at the global level. However, the novelty performance of different local environmental regions varies, and these differences directly affect the accuracy of the topology in capturing key environmental features. To quantify the importance of different local environments, weight coefficients for the local environment dimension need to be generated based on the scene novelty index. The calculation of the weight coefficients is based on the distribution characteristics and spatial correlation of the scene novelty index within the local region. The formula can be expressed as: ;

[0086] The formula originates from the fact that the importance of the local environment is determined by two factors: first, the mean level of the novelty index of the scene in that region; a higher mean indicates more significant novel features in that region. Second, the degree of fluctuation of the novelty index in that region; smaller fluctuations indicate more stable features and higher reference value for topology construction. These two dimensions are quantified through integral operations, and then the degree of fluctuation is converted into a decay factor through an exponential function, ultimately yielding the weighting coefficients for the local environment dimension. In the formula… Let l be the weight coefficient of the l-th local environment dimension. The spatial extent of the l-th local environmental region. Spatial location The novelty index of the scene, The average novelty index for the entire scene. Let be the spatial diameter of the l-th local environmental region.

[0087] After obtaining the weight coefficients of the local environment dimension, they need to be mapped to high-dimensional topological nodes. The weight coefficients of the local environment dimension contain the importance information of different regions, and the high-dimensional space can more fully retain the inherent relationship of this information and avoid feature loss caused by low-dimensional mapping. During the mapping process, the weight coefficients of each local environment dimension constitute the coordinate components of the high-dimensional space as independent dimensions, forming the corresponding high-dimensional topological nodes. At the same time, there are natural spatial relationships between different local regions in the actual environment, and the environmental changes of adjacent regions are often related. Therefore, it is necessary to impose spatial dependency constraints between nodes. This constraint is achieved by limiting the distance between nodes in the high-dimensional space. Topological nodes corresponding to spatially adjacent local regions are kept closer in the high-dimensional space, while the distance between nodes corresponding to spatially distant regions is increased accordingly, ensuring that the distribution of topological nodes can indirectly reflect the spatial structure of the actual environment.

[0088] Next, edge weights are established between the generated high-dimensional topological nodes. The role of edge weights is to quantify the strength of the association between nodes. This strength is determined by the dynamic changes in the local environment corresponding to the node. Historical state changes of a node reflect long-term feature evolution trends, while instantaneous offsets reflect current sudden changes. The combined difference between these two can comprehensively characterize the dynamic association between nodes. The formula for calculating edge weights is as follows: This formula is based on the dual attributes of the strength of association between nodes: the cumulative difference of historical state changes and the instantaneous difference of instantaneous offset both affect the degree of association. These two differences are quantified separately by a distance function, and then weighting coefficients are introduced. By balancing the contribution ratios of the two, the final edge weights between nodes are obtained, where... For the first The node and the first Edge weights between nodes The balance coefficient and , and The first The node and the first The historical state sequence of each node and The first The node and the first The instantaneous offset vector of each node. This is a function for calculating Euclidean distance.

[0089] After the edge weights are determined, a minimum spanning tree is constructed based on the edge weights between nodes. The minimum spanning tree minimizes the total edge weights while ensuring connectivity of all nodes, thus highlighting the connections between nodes and simplifying the complexity of the topology. During the generation process, the edge weights are not fixed but need to be dynamically adjusted. The adjustment is based on the real-time updated scene novelty index and node states. When the novelty index of the local environment changes, the corresponding node states and edge weights need to be adjusted accordingly to ensure that the topology can reflect the dynamic changes of the environment in real time, ultimately forming a topology that balances connectivity, simplicity, and timeliness. Finally, the final topology is subjected to topology feedback calibration. The calibration process uses the application effect of historical topologies as a reference and verifies it in combination with the real dynamic characteristics of the actual environment. By comparing the node connections and edge weight distribution depicted by the topology with the interaction relationships and change patterns of local areas in the actual environment, the distribution positions of nodes and the quantitative values ​​of edge weights are adjusted. The purpose of calibration is to eliminate possible deviations in the topology construction process and ensure that the topology can accurately and reliably reflect the dynamic characteristics of the environment.

[0090] In some embodiments of the present invention, the decision module further includes: dynamically classifying the edge weights in the dynamic scene topology to form a multi-level condition group, and segmenting and mapping the historical fluctuation range of the edge weights with the local nonlinear offset to establish an edge weight condition mapping table.

[0091] The condition rules in the edge weight condition mapping table are combined to form local decision rule groups, and the local rules are prioritized according to the topological path length between nodes and the strength of the edge weight conditions.

[0092] A strategy sequence is generated based on the sorted local decision rule groups, and the order of local rule combination is adjusted through multi-stage perturbation to form a global decision strategy;

[0093] The global decision-making strategy is constrained and validated to identify the consistency of path connectivity and edge weight conditions, and the strategy is fine-tuned and corrected through local rules.

[0094] In this embodiment, edge weights are first dynamically classified. The magnitude and variation characteristics of edge weights directly reflect the correlation strength and dynamic trend of the local environment corresponding to the node. Edge weights with different characteristics need to be matched with decision logic at different levels. Therefore, clustering and classification are required based on the dynamic characteristics of edge weights. The classification process is based on the numerical distribution, rate of change, and feature similarity of associated nodes of edge weights. Clusters are formed by calculating the feature similarity between edge weights, and finally, multi-level condition groups are divided. Each condition group corresponds to a set of edge weights with similar dynamic characteristics. To establish a direct correspondence between edge weights and decision conditions, the historical fluctuation range and local nonlinear offset of edge weights need to be segmented and mapped. The historical fluctuation range reflects the long-term stable characteristics of edge weights, while the local nonlinear offset reflects short-term sudden changes. By segmenting these two types of features according to preset intervals, each feature interval corresponds to a specific decision condition, and then integrating them to form an edge weight condition mapping table. The quantitative relationship of this mapping process can be expressed as: ;

[0095] This formula stems from the need for decision-making conditions to simultaneously consider the long-term stability and short-term dynamics of the edge weights. It normalizes the historical fluctuation range and local nonlinear offset separately, and then introduces weighting coefficients. By balancing the influence of both factors on the decision-making conditions, a quantified condition index is ultimately obtained, where... For the first The quantitative index of the q-th decision condition in the hierarchical condition group, where w is the current edge weight. and These are the historical minimum and historical maximum fluctuation values ​​of the edge weight, respectively. This represents the local nonlinear offset of the edge weights. and These are the minimum and maximum thresholds for local nonlinear offset, respectively. The balance coefficient and .

[0096] Based on the edge weight condition mapping table, the interrelated decision conditions in the table are logically combined to form local decision rule groups. Each local decision rule group corresponds to the decision logic in a specific local environment scenario and can independently support the basic decision needs in that scenario. To ensure the rationality and efficiency of decision execution, the local rules need to be prioritized according to the topological path length between nodes and the strength of the edge weight conditions. The shorter the topological path length between nodes, the more direct the connection between the corresponding local environment regions, and the higher the urgency of executing the relevant decision rules. The strength of the edge weight conditions directly reflects the importance of the decision needs; the larger the edge weight value, the stronger the corresponding condition, and the higher the priority of the decision rule should be. The quantitative calculation of priority can be expressed as follows: The formula is based on the fact that priority is jointly determined by execution urgency and demand importance. The reciprocal of the path length quantifies execution urgency, and the normalized value of the edge weight quantifies demand importance. A balance coefficient is used to further define the priority. Integrating the two indicators, a priority value for the local rule is formed, where... For the first The priority of local decision-making rules This represents the length of the topological path between the nodes corresponding to this rule. This represents the edge weight condition strength value corresponding to the rule. It is the maximum value of all edge weight condition strength values. The balance coefficient and .

[0097] Local decision rules are arranged sequentially according to their priority from high to low to generate a preliminary strategy sequence. However, this strategy sequence is based solely on a static priority order and may not fully adapt to real-time changes in dynamic scenarios. Therefore, it is necessary to adjust the combination order of local rules through multi-stage perturbation. Multi-stage perturbation introduces small-scale order adjustments at different decision stages to simulate the decision-making effects under different rule combinations. By comparing the differences in the adaptability of each combination, the rule combination order that is more suitable for the current dynamic scenario is selected. During the perturbation process, the adjustment range at each stage is dynamically adjusted according to the changes in the scenario novelty index. The adjustment range is smaller when the scenario changes slowly and appropriately increased when the scenario changes drastically, ultimately forming a global decision strategy that balances priority and dynamic adaptability.

[0098] In some embodiments of the present invention, the performance calculation module further includes: refining the deviation between the vehicle's motion parameters and control commands into multiple motion stages, and setting a deviation measurement index for each stage;

[0099] Based on the deviation value of each motion stage, different weighting coefficients are applied to the motion state of each stage, and the disturbance coefficient is weighted according to the magnitude and trend of the deviation value.

[0100] Based on the generated disturbance coefficients for each stage, a dynamic feedback loop is constructed to adapt to environmental changes by adjusting the disturbance coefficients in real time.

[0101] Based on the adjusted perturbation coefficients, the perturbation coefficients are evaluated over a long period through global optimization.

[0102] In this embodiment, the deviation between vehicle motion parameters and control commands needs to be refined into multiple motion stages. When the vehicle executes a decision-making strategy, it will experience different motion states such as acceleration, constant speed, turning, and deceleration. The motion characteristics of different stages are different, and the causes and effects of deviations are also different. For each motion stage, corresponding deviation measurement indicators need to be set. These indicators need to fit the motion characteristics of the stage. For example, the acceleration stage focuses on speed deviation and acceleration deviation, and the turning stage focuses on steering angle deviation and angular velocity deviation. Through targeted indicator setting, the deviation of each stage can be accurately captured. Based on the deviation value of each motion stage, different weight coefficients need to be applied to the motion state of different stages. Different motion stages have different degrees of impact on the overall vehicle operation safety and decision execution effect. For example, the risk brought by deviation is higher in the constant speed stage of high-speed driving or the turning stage in complex road conditions, and the corresponding weight coefficient needs to be increased accordingly. At the same time, the calculation of the disturbance coefficient needs to take into account both the magnitude and trend of the deviation value. The larger the deviation value, the more obvious the performance fluctuation. The trend of continuous increase in deviation indicates potential performance risks, which need to be considered in the calculation. The formula for calculating the disturbance coefficient can be expressed as: ;

[0103] The instantaneous magnitude of the deviation directly reflects the current performance status, while the rate of change of the deviation reflects the fluctuation trend. By linearly combining the two and multiplying them by the stage weighting coefficient, the degree of disturbance in a single motion stage can be comprehensively quantified. Let be the performance perturbation coefficient for the m-th motion stage. Let be the weighting coefficient for the m-th motion phase. Let be the deviation value for the m-th motion stage. Let m be the rate of change of the deviation value in the m-th motion stage. The trend weight coefficient and .

[0104] After obtaining the performance disturbance coefficients for each stage, a dynamic feedback loop needs to be constructed to adapt to environmental changes. The dynamic feedback loop takes real-time collected motion parameters, control commands, and deviation data as input and continuously monitors the changes in disturbance coefficients at each stage. When environmental parameters change, causing abnormal fluctuations in the deviation value or rate of change, the feedback loop adjusts the weighting coefficients or trend weighting coefficients of the corresponding stage in real time according to the fluctuation amplitude. This allows the disturbance coefficients to respond promptly to dynamic environmental changes and accurately reflect the performance fluctuations of the decision-making strategy in the new environment. For example, when a sudden obstacle appears in the environment, causing the vehicle to decelerate urgently, the deviation in the deceleration stage may change drastically. The feedback loop will immediately increase the weighting coefficient of that stage, allowing the disturbance coefficients to more accurately characterize the performance fluctuations under this sudden scenario.

[0105] Based on the dynamically adjusted perturbation coefficients, long-term evaluation through global optimization is required. This long-term evaluation uses perturbation coefficient data from multiple decision-making cycles, combined with historical environmental characteristics and decision execution effects, to analyze the changing patterns and long-term trends of the perturbation coefficients. The global optimization process uses statistical analysis of the mean, variance, and correlation of perturbation coefficients at each stage to identify long-term performance fluctuation risks. For example, a persistently high perturbation coefficient at a certain stage may indicate an adaptation problem with the decision-making rules for that stage. By adjusting the global configuration of the weight coefficients and trend weight coefficients at each stage through global optimization, the calculation of the performance perturbation coefficients can accurately respond to real-time environmental changes while also considering long-term execution stability. Ultimately, this yields performance perturbation coefficients that combine real-time performance and reliability, providing support for the recalibration of dynamic scene topologies.

[0106] In some embodiments of the present invention, the topology calibration module further includes: identifying different stages in the vehicle motion process during the execution of the decision strategy, and setting performance calibration indicators for each stage;

[0107] The motion state and control deviation at each stage are interactively analyzed, and different weighting coefficients are applied to the disturbance coefficient at each stage according to the current motion state of the vehicle.

[0108] Based on the real-time control execution of the vehicle, the control error fluctuation at each stage is analyzed, and the disturbance coefficient is adjusted.

[0109] After dynamically adjusting the phased disturbance coefficients, a global disturbance correction is implemented, and a global strategy for long-term adjustment of the disturbance coefficients is achieved through multi-period performance evaluation.

[0110] Analyze the flow, activity, and connectivity characteristics of each node and its adjacent edges in the topology to identify the interaction relationships between nodes and edges, and assign preliminary weights to each node and its adjacent edges.

[0111] Construct an optimization function that adjusts the weights of nodes and edges based on the centrality of nodes, the transmission capacity of edges, and the execution efficiency characteristics, and sets constraints to limit abnormal weight distribution.

[0112] Based on the optimization function, the weights of nodes and edges are adjusted in real time through a local feedback mechanism;

[0113] The weight configuration of nodes and edges is adjusted within a multi-objective optimization framework using a global search algorithm.

[0114] In this embodiment, firstly, during the execution of the decision-making strategy, it is necessary to accurately identify the different stages of the vehicle's motion process. Vehicle operation exhibits differentiated motion states such as acceleration, constant speed, steering, and deceleration. The dynamic characteristics and control requirements of each stage differ significantly, directly affecting the execution effect of the decision-making strategy and the actual meaning of the disturbance coefficient. For each motion stage, corresponding performance calibration indicators need to be set. These indicators must closely align with the stage's motion characteristics and control objectives. For example, the acceleration stage emphasizes speed tracking accuracy and acceleration stability; the steering stage emphasizes steering angle control error and path fit; and the constant speed stage emphasizes speed stability and energy consumption optimization. Through targeted calibration indicators... This model can provide a reference for subsequent optimization of disturbance coefficients and adjustment of topology weights. Based on this, an interactive analysis of the motion state and control deviation at each stage is performed. The motion state directly determines the difficulty of executing control commands and the probability of deviations, while control deviation is the core source of performance disturbance coefficients. The interaction between the two reflects the effectiveness and adaptability of the disturbance coefficients. According to the current motion state of the vehicle, different weighting coefficients are applied to the disturbance coefficients at each stage. The greater the impact of the current motion state on overall operational safety and decision-making effectiveness, the higher the corresponding disturbance coefficient weight, ensuring that performance fluctuations at critical stages are given priority consideration. The weighted stage disturbance coefficient can be expressed as... This formula is based on the difference in importance of different motion stages, and introduces stage weighting coefficients. For the original disturbance coefficient A weighted correction is applied to make the corrected perturbation coefficients better reflect the actual impact of performance fluctuations at each stage on topology calibration; where The weighted disturbance coefficient for the m-th motion stage. The weight coefficients for the m-th motion stage satisfy the following conditions: , Let be the original performance perturbation coefficient for the m-th motion stage, and n be the total number of motion stages.

[0115] Based on the real-time control execution of the vehicle, the control error fluctuations at each stage are continuously analyzed. During real-time control execution, environmental changes and vehicle state drift may cause abnormal fluctuations in control error, which directly affect the accuracy of the disturbance coefficient. By monitoring the amplitude, frequency, and trend of error fluctuations, the corresponding disturbance coefficients are dynamically adjusted. For example, when the control error in a certain stage continues to increase, the disturbance coefficient for that stage is increased accordingly to more accurately reflect the current performance fluctuation state, providing real-time and reliable quantitative input for topology calibration. After completing the dynamic adjustment of the staged disturbance coefficients, global disturbance correction needs to be implemented. The disturbance coefficient of a single cycle may be affected by short-term environmental fluctuations and has a certain degree of randomness. Through multi-cycle performance evaluation, disturbance coefficient data from multiple decision execution cycles are collected. Combining historical environmental characteristics and decision execution effects, the long-term trend and statistical law of the disturbance coefficients are analyzed. Based on these analysis results, the global strategy for long-term adjustment of the disturbance coefficients is implemented. For example, the initial configuration of the weight coefficients of each stage is optimized, and the response threshold of error fluctuations is adjusted so that the calculation of the disturbance coefficients can both respond quickly to real-time changes and maintain long-term stability.

[0116] This paper analyzes the flow, activity, and connectivity characteristics of each node and its adjacent edges in the topology. Flow reflects the data transmission intensity between nodes and edges, activity reflects the dynamic frequency of changes in the corresponding environmental region, and connectivity determines the information transmission efficiency of the topology. By quantifying these characteristics, the interaction relationships between nodes and edges are identified. For example, nodes with high flow are more closely associated with their adjacent edges, and nodes with high activity have a more significant impact on the topology's dynamics. Based on these interaction relationships, preliminary weights are assigned to each node and its adjacent edges, so that the preliminary weights can initially reflect the importance and interaction intensity of nodes and edges. To achieve optimized weight allocation, a dedicated optimization function needs to be constructed. The optimization function takes node centrality, edge transmission capacity, and execution efficiency as optimization objectives. Node centrality reflects the degree of a node in the topology; nodes with higher centrality have a greater impact on the generation of decision rules. Edge transmission capacity reflects the efficiency of information transmission between nodes, and execution efficiency is related to the implementation effect of decision strategies. The expression of the optimization function can be represented as: The derivation of this formula stems from the need to comprehensively improve the core representation capability of the topology, information transmission efficiency, and decision execution effectiveness through weight adjustment. This is achieved by weighting and summing the three objectives and then introducing constraint terms. By restricting abnormal weight allocation, an optimization function that balances multiple objectives is ultimately formed, where J is the value of the objective function. For node centrality, For the transmission capacity of the edge, For the sake of edge execution efficiency, , , These are the weight coefficients for the three optimization objectives, and all are greater than 0. These are the weighting coefficients for the constraint terms. These are constraint terms used to limit weight values ​​from exceeding a preset reasonable range and to prevent the occurrence of isolated nodes or edges.

[0117] Based on the constructed optimization function, the weights of nodes and edges are adjusted in real time through a local feedback mechanism. This mechanism continuously monitors the real-time state of each node and edge, including traffic changes, activity fluctuations, and transmission delays. This state data is input into the optimization function to calculate the weight adjustment. When the state of a node or edge changes, the weight adjustment is triggered in a timely manner. For example, if the activity of a node suddenly increases, its weight is quickly increased through the feedback mechanism to ensure that the topology can respond to changes in local features in real time. Finally, a global search algorithm is used to globally optimize the weight configuration of nodes and edges within a multi-objective optimization framework. This framework simultaneously considers multiple objectives, such as improving node centrality, enhancing edge transmission capabilities, and optimizing execution efficiency, avoiding topology imbalance caused by single-objective optimization. The global search algorithm can traverse the reasonable space of weight configuration to find the optimal or near-optimal solution that satisfies all optimization objectives. Through this global optimization, the weight configuration of the topology can accurately match the characteristics of local nodes and edges while ensuring the stability and adaptability of the overall topology, ultimately achieving recalibration of the dynamic scene topology.

[0118] In some embodiments of the present invention, the scene extraction module further includes: analyzing each node, edge and its correlation in the recalibrated topology, identifying the features in the topology, performing parametric modeling, and extracting the attributes of nodes and edges, including connectivity, topology density, link stability and node centrality.

[0119] The extracted topological features are abstracted and categorized to form a simplified feature set, and weights are assigned to each type of feature.

[0120] A scene prototype is generated based on the categorized set of topological features.

[0121] In this embodiment, the recalibrated topology is first comprehensively analyzed, examining each node, edge, and their interrelationships to identify key elements reflecting the essential characteristics of the topology. These features are then transformed into quantifiable attribute indicators through parametric modeling. For connectivity, the focus is on analyzing the degree to which nodes are effectively connected via edges, quantifying the differences in reachability between different nodes to ensure the features reflect the connectivity and reachability of different regions in the environment. The parameterization of topology density focuses on the density of edge distribution, quantified by the following formula: This formula stems from the fact that topological density needs to objectively reflect the tightness of the association between nodes, for nodes containing... In an undirected topology with n nodes, the theoretical maximum number of possible edges is . The standardized density index can be obtained by comparing the actual number of edges E with the maximum possible number of edges. The density index can intuitively reflect the sparse or dense characteristics of the topology. In the formula, D is the topological density. Let E be the total number of nodes in the topology, and E be the actual number of edges in the topology.

[0122] Parameterizing link stability requires combining time-series data of edge weights to analyze edge duration, weight fluctuation amplitude, and continuous connectivity. Stability is quantified by statistically analyzing the coefficient of variation of edge weights within a set period. Node centrality is used to identify core nodes in the topology. By calculating indicators such as the mediating role of nodes in the topological path and the number of connected nodes, the influence of nodes on information transmission and decision propagation is quantified. Through the above process, attributes such as connectivity, topology density, link stability, and node centrality are fully extracted and parameterized, forming a comprehensive topological feature dataset. Abstracting and classifying the extracted topological features is a key step in simplifying feature dimensions and highlighting core information. The abstraction process removes redundant information and secondary fluctuations from the feature data, retaining the essential information. The core features that reflect the essence of the scene are statistically summarized for the centrality indicators of different nodes, extracting the centrality distribution pattern of the topology rather than the specific value of a single node. The classification is based on the similarity of the features' attributes and the functional correlation, integrating the scattered feature indicators into several major categories. For example, connectivity and topology density are classified as structural correlation features, link stability as dynamic evolution features, and node centrality as functional features. Based on the contribution of each type of feature to the overall representation of the scene, corresponding weights are assigned to each type of feature. The higher the contribution, the greater the weight of the feature category. For example, functional features that directly affect the priority of decision rules will have a higher weight than other auxiliary features, ensuring that the scene prototype generated later can highlight the key features.

[0123] Based on the categorized set of topological features, a scene prototype is generated through structured integration. The scene prototype is not a simple accumulation of individual features, but an organic combination of various features according to their inherent relationships and weight configurations to form a complete feature representation system. The feature representation system can capture the attributes of a type of environmental scene. For example, the prototype of a certain type of urban road scene will integrate key features such as low-density connectivity, medium link stability, and a few highly central nodes in the topological structure of this type of scene. During the generation process, the quantified feature parameters and weight configurations need to be solidified into a standardized prototype structure to ensure that the prototype can clearly and stably represent the scene features of the corresponding category.

[0124] In some embodiments of the present invention, the prototype evolution module further includes: monitoring key parameters of the scene prototype and periodically detecting and evaluating changes in key parameters;

[0125] Based on the feedback, the key parameters in the scene prototype were corrected, and the connection methods between nodes or the weights of edges were adjusted.

[0126] By evaluating the evolution of multiple system cycles and combining historical data and long-term trends, optimization strategies are adjusted; persistent cohomology analysis is performed on the topology of new scenario prototypes and inventory prototypes to extract topological invariants, and difference score vectors are generated based on Wasserstein distance to compare persistence intervals.

[0127] Using the difference score vector and historical performance perturbation coefficient, the trigger threshold is adjusted through a dynamic threshold function, and the trigger decision is output.

[0128] When the trigger decision instruction is triggered, the key topological difference region is identified based on the difference score vector, and the dynamic scene topology is reconstructed by local gradient optimization using the projection gradient descent algorithm.

[0129] Based on the difference score vector and the reconstructed topology, a persistent homology-guided spectral clustering algorithm is used to cluster and update the prototype library, merging or adding scene prototypes.

[0130] In this embodiment, key parameters of the scene prototype need to be continuously monitored. These key parameters include previously extracted core attributes such as connectivity, topology density, link stability, and node centrality. They directly determine the accuracy of the scene prototype in representing environmental features. By setting a fixed detection cycle, real-time data of key parameters is collected periodically and compared with the initial parameters and historical change data of the prototype to evaluate the magnitude, rate, and trend of parameter changes, and to determine whether the changes are due to changes in the essential characteristics of the environment. Based on the feedback information obtained from the monitoring and evaluation, each key parameter in the scene prototype is specifically corrected. When the parameter changes exceed the preset reasonable fluctuation range, it indicates that the current prototype can no longer accurately match the actual environmental features. It is necessary to correct the parameter deviation by adjusting the connection method between nodes or the weight of edges. For example, if the topology density parameter continues to increase, it indicates that the correlation strength of local areas in the environment has increased. The density parameter can be corrected by increasing the number of edges between nodes in the corresponding area or increasing the edge weight. If the link stability parameter decreases, it is necessary to adjust the weight configuration of relevant edges or optimize the node connection logic so that the corrected scene prototype can accurately reflect the dynamic features of the environment again.

[0131] Based on this, by comprehensively evaluating the evolution of multiple system cycles and combining historical evolution data with long-term trends, optimization strategies are dynamically adjusted. Parameter changes in a single cycle may be affected by short-term environmental disturbances and are not representative, while evolution data from multiple cycles can show more stable changes. By analyzing the frequency, magnitude, and effect of parameter corrections in different cycles, the long-term evolution direction of the scenario prototype can be identified. For example, the topology density of certain types of scenarios shows a continuous upward trend, or the node centrality distribution gradually concentrates. Based on these trends, the core objectives and parameter configurations of the optimization strategy are adjusted, enabling the evolution mechanism to proactively adapt to long-term environmental changes and improve the long-term stability and adaptability of the system.

[0132] After a new scene prototype is generated, it needs to be compared with the existing prototypes in the prototype library to determine whether topology reconstruction and prototype library updates are necessary. The core of the similarity comparison is to extract topological invariants through persistent cohomology analysis. Persistent cohomology analysis can capture the core features of the topology at different scales, unaffected by changes in local details. The extracted topological invariants can stably represent the essential properties of the topology. Based on the extracted topological invariants, a persistence interval is constructed. The difference between the persistence intervals of the new scene prototype and the existing prototypes is quantified using Wasserstein distance, generating a difference score vector. The formula for calculating Wasserstein distance is: ;

[0133] The derivation of this formula stems from the need to accurately measure the distance between two probability distributions; the distribution characteristics of the persistence interval can be transformed into a probability measure. and By finding the optimal coupling Minimize the expected distance, then standardize the distance result using p-th power and square root operations, finally obtaining a quantitative index that accurately reflects the essential differences between the two topologies, where... For p-order Wasserstein distance, The distance order is... For the probability measure of the persistence interval of the new scenario prototype, A probability measure of the persistence interval of the inventory prototype. Represents the set of persistent intervals for a new scene prototype. and the set of persistent intervals of the inventory prototype Integrating the Cartesian product, the Cartesian product It contains all possible persistent interval pairs in both sets. ,in , By taking the Euclidean distance of all interval pairs in this Cartesian product space Multiply by transportation plan Integrating allows for a comprehensive quantification of the overall difference between two persistent interval sets; For all couplings The set, and These are sample points in two probability measures, respectively. It is a Euclidean distance.

[0134] Using the generated difference score vector and historical performance perturbation coefficient, a dynamic threshold function is used to adjust the trigger threshold and output the trigger decision. The difference score vector directly reflects the degree of topological difference between the new and old prototypes, while the historical performance perturbation coefficient reflects the performance fluctuation of the decision strategy corresponding to the inventory prototype in actual execution. The combination of the two can comprehensively determine whether topology reconstruction is necessary. The expression of the dynamic threshold function is: ;

[0135] This formula utilizes a trigger threshold and then dynamically adjusts it based on historical performance and the current degree of difference. A higher mean historical performance disturbance coefficient indicates poorer adaptability of the inventory prototype, and the threshold should be lowered to more easily trigger reconstruction. Conversely, a higher magnitude of the current difference score vector indicates more significant differences between the old and new prototypes, and the threshold should be increased to avoid false triggers. This dynamic adjustment logic is achieved through a combination of an exponential function and a linear term. The adjusted trigger threshold, As the initial threshold, and These are adjustment coefficients, and all are greater than 0. For historical cycle number, For the first Performance disturbance coefficient for each historical period For the difference score vector, The magnitude of the difference score vector is used. When the magnitude of the difference score vector is greater than the adjusted trigger threshold, the trigger decision indicates that topology reconstruction will be triggered; otherwise, it will not be triggered.

[0136] When the decision instruction is triggered, key topological difference regions are identified based on the difference score vector. The dimension with a larger value in the difference score vector corresponds to the feature dimension with significant difference in the topological structure. By combining the topological position corresponding to the feature dimension, the key difference region can be located. The dynamic scene topology is reconstructed by local gradient optimization using the projection gradient descent algorithm. This algorithm can make precise local adjustments in the key difference region along the gradient direction of the optimization target, avoiding unnecessary disturbance to the overall topological structure. This ensures that the reconstructed topology can not only adapt to the features of the new scene prototype, but also maintain the stability and coherence of the overall structure. Based on the difference score vector and the reconstructed topology, the prototype library is updated by clustering using a persistent coherence-guided spectral clustering algorithm. Persistent coherence guidance ensures that the clustering process focuses on the essential features of the topology, while spectral clustering algorithms can efficiently group similar scene prototypes into one class. By calculating the cluster similarity between a new scene prototype and an inventory prototype, if the similarity is higher than a preset threshold, the new scene prototype is merged with the corresponding inventory prototype, and the core parameters of that class of prototypes are updated; if the similarity is lower than the preset threshold, the new scene prototype is added as a new category to the prototype library, realizing the dynamic expansion and optimization of the prototype library, and ensuring that the prototype library can continuously cover various environmental scenarios.

[0137] The above description is merely a preferred embodiment of the present invention; however, the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and its improved concept, should be covered within the scope of protection of the present invention.

Claims

1. A low-speed unmanned vehicle artificial intelligence decision-making and performance evaluation system, characterized in that, include: The generation module is used to generate a scene novelty index by monitoring the instantaneous shift of environmental parameters; The topology module is used to construct a dynamic scene topology based on the scene novelty index. The dynamic scene topology includes nodes and edge weights generated based on the scene novelty index. The decision module is used to convert edge weights into decision rules and generate decision strategies based on the dynamic scene topology. The performance calculation module is used to track the deviation between vehicle motion parameters and control commands during the execution of decision-making strategies in order to calculate performance disturbance coefficients; The topology calibration module is used to recalibrate the topology of dynamic scenes based on the performance perturbation coefficient, and the recalibration is achieved by adjusting the weights of nodes and edges. The scene extraction module is used to extract key topological features from the recalibrated topology and generate scene prototypes; The prototype evolution module is used to establish a continuous evolution mechanism based on the scene prototype. It triggers topology reconstruction and updates the prototype library by comparing the similarity between the new scene prototype and the inventory prototype.

2. The low-speed unmanned vehicle artificial intelligence decision-making and performance evaluation system according to claim 1, characterized in that, The construction of dynamic scene topology based on scene novelty index includes: Local feature enhancement is performed based on the scene novelty index to generate weight coefficients for the local environment dimension; The weight coefficients of the local environment dimension are mapped to high-dimensional topological nodes, and spatial dependency constraints are applied between the nodes. Edge weights are established between the generated high-dimensional topology nodes. The edge weights are generated based on the difference mapping between the historical state changes and instantaneous offsets of the nodes. A minimum spanning tree is constructed based on the edge weights between nodes, and the edge weights are dynamically adjusted during the generation process to form the final topology. Perform topology feedback calibration on the final topology structure.

3. The low-speed unmanned vehicle artificial intelligence decision-making and performance evaluation system according to claim 2, characterized in that, The step of converting edge weights into decision rules and generating decision strategies based on dynamic scene topology includes: The edge weights in the dynamic scene topology are dynamically classified to form multi-level condition groups, and the historical fluctuation range of the edge weights is segmented and mapped with the local nonlinear offset to establish an edge weight condition mapping table. The condition rules in the edge weight condition mapping table are combined to form local decision rule groups, and the local rules are prioritized according to the topological path length between nodes and the strength of the edge weight conditions. A strategy sequence is generated based on the sorted local decision rule groups, and the order of local rule combination is adjusted through multi-stage perturbation to form a global decision strategy; The global decision-making strategy is constrained and validated to identify the consistency of path connectivity and edge weight conditions, and the strategy is fine-tuned and corrected through local rules.

4. The low-speed unmanned vehicle artificial intelligence decision-making and performance evaluation system according to claim 3, characterized in that, The step of tracking the deviation between vehicle motion parameters and control commands during the execution of the decision-making strategy to calculate the performance disturbance coefficient includes: The deviation between the vehicle's motion parameters and control commands is broken down into multiple motion stages, and a deviation measurement index is set for each stage. Based on the deviation value of each motion stage, different weighting coefficients are applied to the motion state of each stage, and the disturbance coefficient is weighted according to the magnitude and trend of the deviation value. Based on the generated disturbance coefficients for each stage, a dynamic feedback loop is constructed to adapt to environmental changes by adjusting the disturbance coefficients in real time. Based on the adjusted perturbation coefficients, the perturbation coefficients are evaluated over a long period through global optimization.

5. The low-speed unmanned vehicle artificial intelligence decision-making and performance evaluation system according to claim 4, characterized in that, The recalibration of the dynamic scene topology based on the performance perturbation coefficient includes: During the execution of the decision-making strategy, different stages of the vehicle's movement are identified, and performance calibration indicators are set for each stage; The motion state and control deviation at each stage are interactively analyzed, and different weighting coefficients are applied to the disturbance coefficient at each stage according to the current motion state of the vehicle. Based on the real-time control execution of the vehicle, the control error fluctuation at each stage is analyzed, and the disturbance coefficient is adjusted. After dynamically adjusting the phased disturbance coefficients, a global disturbance correction is implemented, and a global strategy for long-term adjustment of the disturbance coefficients is achieved through multi-period performance evaluation.

6. The low-speed unmanned vehicle artificial intelligence decision-making and performance evaluation system according to claim 5, characterized in that, The recalibration achieved by adjusting node and edge weights includes: Analyze the flow, activity, and connectivity characteristics of each node and its adjacent edges in the topology to identify the interaction relationships between nodes and edges, and assign preliminary weights to each node and its adjacent edges. Construct an optimization function that adjusts the weights of nodes and edges based on the centrality of nodes, the transmission capacity of edges, and the execution efficiency characteristics, and sets constraints to limit abnormal weight distribution. Based on the optimization function, the weights of nodes and edges are adjusted in real time through a local feedback mechanism; The weight configuration of nodes and edges is adjusted within a multi-objective optimization framework using a global search algorithm.

7. The low-speed unmanned vehicle artificial intelligence decision-making and performance evaluation system according to claim 6, characterized in that, The step of extracting key topological features from the recalibrated topology and generating a scene prototype includes: Each node, edge, and its relationships in the recalibrated topology are analyzed to identify the features in the topology and to perform parametric modeling. The attributes of nodes and edges are extracted, including connectivity, topology density, link stability, and node centrality. The extracted topological features are abstracted and categorized to form a simplified feature set, and weights are assigned to each type of feature. A scene prototype is generated based on the categorized set of topological features.

8. The low-speed unmanned vehicle artificial intelligence decision-making and performance evaluation system according to claim 7, characterized in that, The establishment of a continuous evolution mechanism based on the scenario prototype includes: Monitor key parameters of the scenario prototype and regularly detect and evaluate changes in key parameters; Based on the feedback, the key parameters in the scene prototype were corrected, and the connection methods between nodes or the weights of edges were adjusted. By evaluating the evolution of multiple system cycles and combining historical data with long-term trends, optimization strategies are adjusted.

9. The low-speed unmanned vehicle artificial intelligence decision-making and performance evaluation system according to claim 8, characterized in that, The step of triggering topology reconstruction and updating the prototype library by comparing the similarity between the new scene prototype and the inventory prototypes includes: Persistent cohomology analysis is performed on the topological structure of the new scenario prototype and the inventory prototype to extract topological invariants, and difference score vectors are generated based on the Wasserstein distance to compare persistence intervals. Using the difference score vector and historical performance perturbation coefficient, the trigger threshold is adjusted through a dynamic threshold function, and the trigger decision is output. When the trigger decision instruction is triggered, the key topological difference region is identified based on the difference score vector, and the dynamic scene topology is reconstructed by local gradient optimization using the projection gradient descent algorithm. Based on the difference score vector and the reconstructed topology, a persistent homology-guided spectral clustering algorithm is used to cluster and update the prototype library, merging or adding scene prototypes.

10. The low-speed unmanned vehicle artificial intelligence decision-making and performance evaluation system according to claim 1, characterized in that, The method of generating a scene novelty index by monitoring the instantaneous shift of environmental parameters includes: Continuous wavelet transform is performed on the time series data of environmental parameters to extract instantaneous amplitude and phase shift at different time scales, forming a multi-scale transient feature matrix; The phase space is reconstructed by time delay embedding method for multi-scale transient feature matrix, a phase space point matrix is ​​constructed, and singular value decomposition is performed on the phase space point matrix to extract the left singular vectors corresponding to the first k singular values ​​as the dominant mode vectors. Based on the dominant mode vector, the information entropy of each mode is calculated, and the information entropy is weighted and summed according to the relative contribution of singular values ​​to generate the scene novelty index.

Images