New energy commercial vehicle adaptive control method and system based on scene perception

Abstract

The invention relates to the field of scene self-adaptive control, in particular to a new energy commercial vehicle self-adaptive control method and system based on scene awareness, and the method comprises the steps: extracting vehicle real-time operation features and inputting the vehicle real-time operation features to a scene recognition model which fuses semi-supervised clustering and a domain knowledge graph; and outputting the current scene category and the associated performance demand. The model has an online updating capability, and when it is detected that a current real-time feature vector significantly deviates from a known scene, iterative optimization is automatically triggered. Based on the recognized scene and performance requirements, a strategy reasoning engine is called, the engine conducts multi-target reasoning and optimization calculation by traversing the logic relation in the knowledge graph and combining real-time data, and an accurate vehicle control parameter set matched with the current scene is generated. And the parameter set is output to each control system of the vehicle, so that self-adaptive closed-loop control of an electric driving system, a battery management system and a thermal management system is realized, and the energy efficiency, the performance and the intelligent level of the whole vehicle in different operation scenes are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of scene adaptive control, in particular to a new energy commercial vehicle adaptive control method and system based on scene perception. BACKGROUND

[0002] New energy commercial vehicles operate in complex and diverse scenes, covering city logistics, port transportation and other application scenarios. The performance requirements of the vehicle for electric drive, battery management and thermal management systems differ significantly in different scenarios. Existing vehicle control methods mostly use fixed control strategies or simple adaptation schemes based on single-dimensional data, lacking accurate perception and dynamic response capabilities for complex scenes.

[0003] Current scene recognition techniques often rely on basic clustering algorithms, which have problems such as lack of semantic information and rigid scene adaptation, making it difficult to accurately match the actual operating state of the vehicle. At the same time, the control strategy reasoning lacks reliable logical support, and cannot meet the key performance requirements in different scenarios.

[0004] In addition, as operation data accumulates and scenarios evolve, control strategies are difficult to adaptively optimize, affecting vehicle operating efficiency. Therefore, there is an urgent need for an adaptive control method that integrates scene perception to achieve precise control and performance improvement of new energy commercial vehicles in complex and changing scenarios. SUMMARY

[0005] To solve the above problems, the present application provides a new energy commercial vehicle adaptive control method and system based on scene perception.

[0006] In a first aspect, the present application provides a new energy commercial vehicle adaptive control method based on scene perception, comprising the following steps: S1, obtaining multi-dimensional real-time operating data of a new energy commercial vehicle, and performing feature extraction on the operating data to form a real-time feature vector; S2, inputting the real-time feature vector into a scene recognition model; the scene recognition model outputs a recognition result of the application scenario category in which the current vehicle is located, and a description of the key performance requirements associated from the knowledge graph; during model application, if the deviation of the current real-time feature vector from the existing scene pattern representation exceeds a preset threshold, an iterative update of the scene recognition model is triggered; S3, based on the application scenario category and the description of the key performance requirements, calling a strategy reasoning engine; the strategy reasoning engine analyzes the description of the key performance requirements and combines the real-time operating data to perform reasoning and optimization calculation through traversing the logical relationships between entities associated with the current scene in the knowledge graph, generating a set of vehicle control parameters adapted to the scene; S4. Output the vehicle control parameter set to dynamically adjust the control parameters of at least one of the vehicle's electric drive system, battery management system, and thermal management system to achieve scenario-based adaptive control.

[0007] Through the two-stage collaboration of scene recognition model and policy reasoning engine, the vehicle's operating status is directly associated with high-level business scenario semantics and performance requirements, and the optimal control parameters are dynamically generated based on this.

[0008] Because the control strategy is dynamically generated and optimized based on the precise performance requirements of the current specific scenario, it can maximize the vehicle's operation within the optimal or near-optimal range of various performance indicators. This results in reduced energy consumption, extended driving range, optimized operating status of key components, and improved driving smoothness.

[0009] By integrating deviation detection and trigger update mechanisms into the model application, the entire system can continuously optimize and improve itself as vehicle operation data accumulates and new scenarios emerge. This allows the system to maintain a high degree of real-world mapping over the long term, avoiding the performance degradation problem of traditional systems due to scenario fixation.

[0010] As a further limitation of the technical solution of the present invention, the scene recognition model is constructed in the following manner: S021. Obtain historical multi-dimensional operating feature data of new energy commercial vehicles, and use a semi-supervised clustering algorithm with introduced constraints to automatically group the historical multi-dimensional operating feature data, mine potential scene patterns, and form an initial scene pattern cluster. S022. Construct a knowledge graph in the field of new energy commercial vehicles. The knowledge graph includes scene types, vehicle components, performance indicators, environmental factor entities, and edges representing the logical relationships between them. S023. Semantically align and fuse the initial scene pattern clusters with the constructed knowledge graph; specifically, this includes mapping each initial scene pattern cluster to a scene type entity in the knowledge graph, and based on the logical relationship of the entity in the knowledge graph, associating a set of key performance requirement features to generate semantic scene patterns. S024. Based on the semantic scene pattern, construct a training dataset; wherein the input sample is the historical multidimensional running feature data in the initial scene pattern cluster, and the output label is the corresponding scene type and associated key performance requirement description; use the training dataset to train a machine learning model to obtain a scene recognition model, which can output the recognized scene category and key performance requirement description according to the input feature data.

[0011] This approach deeply integrates data-driven pattern mining with knowledge-driven semantic fusion. By semantic alignment, data patterns are given contextual meaning and performance requirement associations. This enables the trained model to not only accurately identify scenario categories but also simultaneously output key performance requirements, providing a clear basis for subsequent control strategy inference.

[0012] As a further limitation of the technical solution of the present invention, the steps of S021 include: S021a. Obtain historical time-series operation data of new energy commercial vehicles, including vehicle status parameters, driving energy consumption sequence, geographical location trajectory and environmental perception information; clean, align and standardize the time-series operation data to obtain a standardized historical operation dataset. S021b. Based on the standardized historical operation dataset, a multi-dimensional operation feature vector is constructed to characterize the operation mode through time window sliding and statistical calculation; the multi-dimensional operation feature vector includes at least spatiotemporal distribution features, energy consumption load features, and driving behavior features. S021c. Based on the labeled operation logs, determine the attribution relationship between at least some sample pairs and construct a set of pairwise constraints to guide clustering; the set of pairwise constraints includes: Must-link constraints containing sample pairs that must belong to the same scene pattern and Cannot-link constraints containing sample pairs that cannot belong to the same scene pattern. S021d, The multidimensional running feature vector and the pairwise constraint set are input together into the improved semi-supervised clustering algorithm; the algorithm iteratively updates the cluster centers and the cluster affiliation of samples, so that the objective function... Reaching the minimum value, all samples are eventually divided into k A cluster of initial scene patterns that do not overlap.

[0013] By standardizing time-series data and constructing multi-dimensional features, the validity of the input data is ensured. The introduction of paired constraint sets provides domain knowledge guidance for clustering, avoiding the blindness of unsupervised clustering. The improved semi-supervised clustering algorithm optimizes the objective function iteratively, making the initial scenario pattern cluster division more in line with the actual application scenario.

[0014] As a further limitation of the technical solution of the present invention, the improved semi-supervised clustering algorithm described in S021d iteratively optimizes the following objective function. To achieve clustering:

[0015] In the formula, The preset number of clusters, For the first Clusters, For the first Cluster The cluster center, For sample feature vectors Distance from the cluster center To constrain the penalty weighting coefficient for violations; The number of Must-link constraints violated. The number of Cannot-link constraints that are violated.

[0016] The objective function achieves a quantifiable balance between intra-cluster compactness and constraint satisfaction. The weighting coefficients can flexibly adjust the importance of data aggregation and domain rules, effectively reducing Must-link / Cannot-link constraint violations, making the clustering results fit the logic of the scenario domain, and improving the rationality and reliability of the initial scenario pattern clusters.

[0017] As a further limitation of the technical solution of the present invention, in S022, the steps of constructing a knowledge graph in the field of new energy commercial vehicles include: S022a. Extract text data containing terms and descriptions related to scenarios, components, indicators and environments from multiple heterogeneous knowledge sources; S022b. Based on predefined entity types, identify and extract scene type entities, vehicle component entities, performance index entities, and environmental factor entities from the text data; S022c. Based on a predefined semantic relationship pattern, establish edges representing logical relationships between the extracted entities; the semantic relationships include at least: demand relationships representing specific performance requirements of the scenario, compositional relationships representing membership or connection relationships between components, and influence relationships representing the impact of the environment on performance or components. S022d. Store the entities and relationships in a graph database to form a queryable knowledge graph; and encapsulate a service interface for the knowledge graph, which supports the retrieval of entities, relationships and attributes and rule-based logical reasoning.

[0018] Construct a domain knowledge graph covering multiple types of entities and semantic relationships to achieve structured integration of scattered domain knowledge; design a graph database storage and standardized service interface to support efficient entity retrieval, relationship traversal and logical reasoning, providing solid knowledge support for scene semantic alignment, performance requirement extraction and strategy reasoning.

[0019] As a further limitation of the technical solution of the present invention, step S023 includes: S023a. For each initial scene pattern cluster, calculate the mean of all its sample feature vectors to obtain the cluster center feature vector representing the pattern cluster. S023b. For each scene type entity in the knowledge graph, generate its corresponding entity semantic feature vector based on its associated entities and relationships. S023c, Calculate the similarity between the cluster center feature vector of each initial scene pattern cluster and the entity semantic feature vector of each scene type entity in the knowledge graph; select the scene type entity with the highest similarity as its candidate mapping entity for each pattern cluster; S023d. Determine whether the highest similarity corresponding to the candidate mapped entity exceeds a preset confidence threshold: If so, then execute S023e; If not, mark the initial scene pattern cluster as an unmatched new pattern cluster, and create a new scene type entity in the knowledge graph based on its cluster center feature vector before executing S023e. S023e. For each initial scenario pattern cluster that has completed mapping, in the knowledge graph, starting from the mapped scenario type entity, traverse along the demand relationship edge to obtain all directly related performance indicator entities and their attribute values, which are used as the key performance requirement features associated with the pattern cluster. S023f. The initial scenario pattern cluster that has been mapped and associated with the requirement features is encapsulated into a semantic scenario pattern; the semantic scenario pattern includes at least: the identifier of the pattern cluster, the mapped scenario type entity, and the associated set of key performance requirement features.

[0020] By matching the similarity between cluster centers and scene entities, a precise mapping between data patterns and domain knowledge is achieved. The mechanism for creating new scenes with unmatched patterns ensures the integrity of scene coverage. The automatic association of performance requirement features endows data patterns with clear semantic attributes. The resulting semantic scene patterns combine data features and business meaning, providing high-quality labeled data for model training.

[0021] As a further limitation of the technical solution of the present invention, S024 specifically includes: S024a. For each semantic scene mode, samples are selected from the historical multidimensional running feature data of its corresponding initial scene mode cluster. The feature vector of the sample is used as the input feature, and the scene type label and key performance requirement feature vector contained in the semantic scene mode are used as supervision labels to form a training data. The set of all training data constitutes the training dataset. S024b. Construct a multi-task neural network model, the model including: A shared feature encoding layer is used to encode the input feature vector; A scene classification output layer, connected to the shared feature encoding layer, is used to output the classification result of the scene type; A requirement regression output layer, connected to the shared feature encoding layer, is used to output the regression prediction value of the key performance requirement feature vector; S024c. The multi-task neural network model is trained using the training dataset, and the joint loss function L is optimized through backpropagation algorithm until the model converges, thereby obtaining the scene recognition model.

[0022] As a further limitation of the technical solution of the present invention, the model iterative update steps include: S2a: When the model is running online, features are extracted from the newly input running data to form the current real-time feature vector; S2b: Calculate the Mahalanobis distance between the current real-time feature vector and each existing scene mode representation in the model; take the minimum value among all Mahalanobis distances; if the minimum value is greater than a preset threshold, trigger model update. S2c: When an update is triggered, relevant runtime data is collected to form a cluster of candidate patterns, which is then submitted to the knowledge graph for semantic review. Based on the review results, one of the following operations is performed: If the review identifies a new scenario, the candidate pattern cluster will be used as a new sample to generate a new semantic scenario pattern, which will then be used to fine-tune the model. If the review indicates a known scenario evolution, the data of the candidate pattern clusters will be merged into the corresponding scenario, and the model will be incrementally learned.

[0023] Deviation detection based on Mahalanobis distance can accurately identify the differences between new data patterns and existing scenarios, avoiding scene omissions or misjudgments; combined with semantic review of knowledge graphs to clarify pattern attributes, the model can be iteratively updated through fine-tuning or incremental learning, so that the scene recognition model always keeps in sync with the actual operating scenario.

[0024] As a further limitation of the technical solution of the present invention, step S3 includes: S31. Parse the key performance requirement description to obtain the various performance indicators and quantitative requirements contained therein; take the scenario type entity corresponding to the application scenario category as the starting node, traverse along the requirement relationship edge in the knowledge graph, verify the various performance indicators, and extract all directly related performance indicator entities and their preset constraints or optimization objectives to form an initial performance requirement constraint set. S32. Based on real-time operating data and vehicle status, retrieve environmental factor entities and vehicle component entities that are connected to the current scene and performance indicators through influence relationships and association relationships in the knowledge graph; quantify their real-time status values ​​into boundary parameters, and merge them with the performance requirement constraint set to formally define them as a multi-objective optimization problem containing multiple optimization objectives and constraints. S33. Using a preset vehicle system simulation model, the multi-objective optimization problem is solved; a multi-objective evolutionary algorithm is used to iteratively search in the decision space, and the performance and constraint satisfaction of each set of candidate parameters are evaluated through the simulation model, and finally a Pareto optimal solution set that is uniformly distributed in the objective space is generated, where each solution corresponds to a set of candidate control parameters. S34. Based on the preset driving strategy, select a final control parameter solution from the Pareto optimal solution set; convert the solution into a standardized control instruction set that conforms to the vehicle's underlying controller communication protocol and interface specifications, and output it.

[0025] Secondly, the technical solution of the present invention also provides an adaptive control system for new energy commercial vehicles based on scene perception, comprising: Data acquisition module: used to acquire multi-dimensional real-time operating data of new energy commercial vehicles; Feature extraction module: used to extract features from the multidimensional real-time running data to form a real-time feature vector; Scene recognition module: It has a built-in pre-built scene recognition model; the scene recognition module is used to input the real-time feature vector into the scene recognition model, output the application scenario category of the current vehicle and the description of key performance requirements associated with the knowledge graph; and during the application of the model, if the deviation between the real-time feature vector and the existing scene pattern representation is detected to exceed a preset threshold, it triggers the iterative update of the scene recognition model. Strategy reasoning engine: Based on the application scenario category and key performance requirement description, it traverses the logical relationships between entities in the knowledge graph associated with the current scenario, and combines real-time running data to perform reasoning and optimization calculations to generate a set of vehicle control parameters adapted to the scenario. Control output module: used to output the set of vehicle control parameters for dynamic adjustment of control parameters of at least one of the vehicle's electric drive system, battery management system and thermal management system, so as to realize scene-based adaptive control.

[0026] As can be seen from the above technical solutions, this application has the following advantages: through the two-stage collaboration of scene recognition model and strategy reasoning engine, the vehicle operation status is directly associated with high-level business scenario semantics and performance requirements, and the optimal control parameters are dynamically generated based on this.

[0027] Achieve precise adaptation of control strategies for new energy commercial vehicles to application scenarios. Relying on the fusion training of semi-supervised clustering and domain knowledge graphs, scene recognition combines data objectivity with semantic richness. The dynamic update mechanism ensures that the model adapts to the evolution of scenarios. Finally, through dynamic adjustment of multiple system parameters, the operating efficiency, energy economy and performance stability of vehicles in different scenarios are improved. Attached Figure Description

[0028] To more clearly illustrate the technical solution of this application, the accompanying drawings used in the description will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0029] Figure 1 This is a flowchart illustrating the method provided in an embodiment of the present invention.

[0030] Figure 2 A block diagram of a system provided in an embodiment of the present invention. Detailed Implementation

[0031] To make the purpose, features, and advantages of this application more apparent and understandable, specific embodiments and accompanying drawings will be used to clearly and completely describe the technical solution protected by this application. Obviously, the embodiments described below are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0032] Unless otherwise defined, all technical and scientific terms used in this application have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this application and in the specification of this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0033] like Figure 1 As shown in the figure, this embodiment of the invention provides an adaptive control method for new energy commercial vehicles based on scene perception, including the following steps: S1. Obtain multi-dimensional real-time operating data of new energy commercial vehicles, and extract features from the operating data to form a real-time feature vector; S2. Input the real-time feature vector into the scene recognition model; the scene recognition model outputs the recognition result of the application scene category of the current vehicle and the description of key performance requirements obtained from the knowledge graph; during the application of the model, if the deviation between the current real-time feature vector and the existing scene pattern representation exceeds a preset threshold, iterative updates of the scene recognition model are triggered. S3. Based on the application scenario category and key performance requirement description, the strategy reasoning engine is invoked; the strategy reasoning engine parses the key performance requirement description and, in conjunction with real-time running data, performs reasoning and optimization calculations by traversing the logical relationships between entities associated with the current scenario in the knowledge graph, generating a set of vehicle control parameters adapted to the scenario. S4. Output the vehicle control parameter set to dynamically adjust the control parameters of at least one of the vehicle's electric drive system, battery management system, and thermal management system to achieve scenario-based adaptive control.

[0034] In this embodiment of the invention, the scene recognition model is constructed in the following manner: S021. Obtain historical multi-dimensional operational feature data of new energy commercial vehicles, and automatically group the historical multi-dimensional operational feature data using a semi-supervised clustering algorithm with introduced constraints to mine potential scene patterns and form an initial scene pattern cluster; this step specifically includes: S021a. Obtain historical time-series operation data of new energy commercial vehicles, including vehicle status parameters, driving energy consumption sequence, geographical location trajectory and environmental perception information; clean, align and standardize the time-series operation data to obtain a standardized historical operation dataset. S021b. Based on the standardized historical operation dataset, a multi-dimensional operation feature vector is constructed to characterize the operation mode through time window sliding and statistical calculation; the multi-dimensional operation feature vector includes at least spatiotemporal distribution features, energy consumption load features, and driving behavior features. The specific process for constructing a multidimensional operational feature vector to characterize the operational mode is as follows: Define a fixed-length time window and a sliding step size. On the timeline of standardized historical data, start from the beginning and slide this window according to the step size. The data within each window will be transformed into a feature vector.

[0035] For the standardized data within each window, calculate the three main categories of features: 1. Spatiotemporal distribution characteristics, including: Average vehicle speed: The average speed of vehicles within the window.

[0036] Trip distance: The cumulative distance traveled, calculated using GPS track points.

[0037] Number of stops / duration: Identifies events where the vehicle speed is 0 and continues to exceed a threshold (e.g., 30 seconds).

[0038] Frequently visited area types: Combining geographical location and map data, determine whether the vehicles in the window are located in logistics parks, highways, city centers, etc.

[0039] Runtime segment: The specific time period in which the window is located.

[0040] 2. Energy load characteristics, including: Average energy consumption per 100 kilometers: Total energy consumption within the window / driving distance, then multiply by 100.

[0041] Battery average / peak output power: Calculates the average and maximum battery power.

[0042] Average / Peak Drive Power: A statistical measure used to calculate the power of the drive motor.

[0043] Average load: The average load rate is estimated by back-calculation based on the vehicle's nominal weight, acceleration and driving force relationship (F=ma).

[0044] 3. Driving behavior characteristics, including: Number of rapid accelerations / decelerations: Define a threshold and count the number of times the speed exceeds the threshold.

[0045] Average acceleration / deceleration: Calculates the average of positive and negative acceleration.

[0046] Speed ​​standard deviation: reflects the fluctuation of vehicle speed.

[0047] Steering wheel angle variance: reflects the frequency and intensity of steering operations.

[0048] All the calculated statistics are concatenated in a predetermined order to form a fixed-dimensional numerical vector. This vector is the multidimensional operational feature vector corresponding to that time window.

[0049] Sliding window, repeat the above steps for the next time period to generate the next multidimensional running feature vector. Finally, the entire historical dataset will be transformed into a matrix composed of several multidimensional feature vectors. S021c, Based on the labeled running logs, determine the attribution relationships between at least some sample pairs and construct a set of pairwise constraints to guide clustering; the set of pairwise constraints includes: Must-link constraints containing sample pairs that must belong to the same scene pattern and Cannot-link constraints containing sample pairs that cannot belong to the same scene pattern; S021d, The multidimensional running feature vector and the pairwise constraint set are input together into the improved semi-supervised clustering algorithm; the algorithm iteratively updates the cluster centers and the cluster affiliation of samples, so that the objective function... Reaching the minimum value, all samples are eventually divided into k The initial clusters of scene patterns are mutually exclusive. The improved semi-supervised clustering algorithm iteratively optimizes the following objective function. To achieve clustering:

[0050] In the formula, The preset number of clusters, For the first Clusters, For the first Cluster The cluster center, For sample feature vectors Distance from the cluster center To constrain the penalty weighting coefficient for violations; The number of Must-link constraints violated. The number of violations of the Cannot-link constraint, the first There are clusters that satisfy... ,in, For the entire sample set, For clusters A single sample in the middle, For the first Cluster The cluster center is usually defined as the mean of the samples within the cluster. For the sample Distance from the cluster center To constrain the penalty weighting coefficient for violations.

[0051] The number of must-link constraints violated, if the sample pairs ( If a data point satisfies the Must-link constraint but is assigned to a different cluster, it is included in this statistic. The number of violations of the Cannot-link constraint, if sample pairs ( If an element satisfies the Cannot-link constraint but is classified into the same cluster, it is included in this statistic. The cluster partitioning is updated iteratively. } and cluster center { }, so that the objective function To reach the minimum value, all samples must eventually be divided into... k A set of mutually exclusive initial scene patterns; S022. Construct a knowledge graph for the field of new energy commercial vehicles. The knowledge graph includes scene types, vehicle components, performance indicators, environmental factor entities, and edges representing the logical relationships between them. The steps for constructing the knowledge graph for the field of new energy commercial vehicles in this step include: S022a. Extract text data containing terms and descriptions related to scenarios, components, indicators and environments from multiple heterogeneous knowledge sources; S022b. Based on predefined entity types, identify and extract scene type entities, vehicle component entities, performance index entities, and environmental factor entities from the text data; S022c. Based on a predefined semantic relationship pattern, establish edges representing logical relationships between the extracted entities; the semantic relationships include at least: demand relationships representing specific performance requirements of the scenario, compositional relationships representing membership or connection relationships between components, and influence relationships representing the impact of the environment on performance or components. S022d. Store the entities and relationships in a graph database to form a queryable knowledge graph; and encapsulate a service interface for the knowledge graph, which supports the retrieval of entities, relationships and attributes and rule-based logical reasoning.

[0052] In this embodiment of the invention, domain knowledge data is obtained from multiple heterogeneous knowledge sources, and the obtained knowledge data is preprocessed; the knowledge sources include at least: new energy commercial vehicle industry technical standards and specifications, vehicle system technical manuals and fault diagnosis libraries, publicly available typical operating condition libraries and test reports, and domain expert experience rule libraries; the obtained unstructured text data is segmented, part-of-speech tagging is performed, and preliminary entity recognition is performed; Based on a pre-defined entity type system, core entities are extracted from pre-processed knowledge data. The entity type system includes: scenario type entities, which correspond to specific vehicle operation and work scenarios. Vehicle component entities: corresponding to the physical subsystems and key components of the vehicle; Performance indicators: These are the technical and economic indicators that evaluate vehicle performance. Environmental factors entities: corresponding to external conditions that affect vehicle operation.

[0053] Based on a preset semantic relationship pattern, the semantic relationships between the core entities are identified and extracted; the semantic relationship pattern includes, but is not limited to: Demand Relationship: Connects the scenario type entity and the performance indicator entity, indicating that a certain scenario has specific requirements for a certain performance. Compositional relationships: Connecting vehicle component entities, indicating the hierarchical or subordinate relationships between components; Impact Relationship: Connects the environmental factor entity with the performance index entity or vehicle component entity, representing the impact of environmental factors on performance or component; Association: Connects scene type entities and environmental factor entities, indicating that a scene is often associated with certain environmental factors.

[0054] Based on the entity types and semantic relationship patterns, a schema layer of the knowledge graph is defined, which specifies the types, attributes, and constraints of entities and relationships.

[0055] The extracted entity and relation instances are stored in the graph database according to the definition of the schema layer to form the data layer of the knowledge graph; the same entity from different knowledge sources is aligned and merged to form a unified entity identifier and complete attributes.

[0056] The completed knowledge graph is encapsulated with query and reasoning interfaces for semantic alignment and relationship querying by the scene recognition and policy generation model.

[0057] S023. Semantically align and fuse the initial scene pattern clusters with the constructed knowledge graph; including: mapping each initial scene pattern cluster to a scene type entity in the knowledge graph, and based on the logical relationship of the entity in the knowledge graph, associating a set of key performance requirement features to generate semantic scene patterns; specifically including: S023a. For each initial scene pattern cluster, calculate the mean of all its sample feature vectors to obtain the cluster center feature vector representing the pattern cluster. S023b. For each scene type entity in the knowledge graph, generate its corresponding entity semantic feature vector based on its associated entities and relationships. S023c, Calculate the similarity between the cluster center feature vector of each initial scene pattern cluster and the entity semantic feature vector of each scene type entity in the knowledge graph; select the scene type entity with the highest similarity as its candidate mapping entity for each pattern cluster; S023d. Determine whether the highest similarity corresponding to the candidate mapped entity exceeds a preset confidence threshold: If it exceeds the limit, then execute S023e; If the limit is not exceeded, the initial scene pattern cluster is marked as an unmatched new pattern cluster, and a new scene type entity is created in the knowledge graph based on its cluster center feature vector before S023e is executed. S023e. For each initial scenario pattern cluster that has completed mapping, in the knowledge graph, starting from the mapped scenario type entity, traverse along the demand relationship edge to obtain all directly related performance indicator entities and their attribute values, which are used as the key performance requirement features associated with the pattern cluster. S023f. The initial scenario pattern cluster that has been mapped and associated with the requirement features is encapsulated into a semantic scenario pattern; the semantic scenario pattern includes at least: the identifier of the pattern cluster, the mapped scenario type entity, and the associated set of key performance requirement features.

[0058] S024. Based on the semantic scene patterns, construct a training dataset; wherein the input samples are historical multi-dimensional operational feature data in the initial scene pattern cluster, and the output labels are the corresponding scene type and associated key performance requirement descriptions; use this training dataset to train a machine learning model to obtain a scene recognition model, which can output the recognized scene category and key performance requirement descriptions based on the input feature data. Specifically, this includes: S024a. For each semantic scene mode, samples are selected from the historical multidimensional running feature data of its corresponding initial scene mode cluster. The feature vector of the sample is used as the input feature, and the scene type label and key performance requirement feature vector contained in the semantic scene mode are used as supervision labels to form a training data. The set of all training data constitutes the training dataset. S024b. Construct a multi-task neural network model, the model including: A shared feature encoding layer is used to encode the input feature vector; A scene classification output layer, connected to the shared feature encoding layer, is used to output the classification result of the scene type; A requirement regression output layer, connected to the shared feature encoding layer, is used to output the regression prediction value of the key performance requirement feature vector; S024c. The multi-task neural network model is trained using the training dataset, and the joint loss function L is optimized through the backpropagation algorithm until the model converges, thereby obtaining the scene recognition model. Joint loss function

[0059] in, The classification loss corresponding to the scene classification task. The regression loss is used to predict the characteristics of the demand. This is a coefficient used to balance the weights of the two tasks.

[0060] In some embodiments, the model iterative update steps include: S2a: When the model is running online, features are extracted from the newly input running data to form the current real-time feature vector; S2b: Calculate the Mahalanobis distance between the current real-time feature vector and each existing scene mode representation in the model; take the minimum value among all Mahalanobis distances; if the minimum value is greater than a preset threshold, trigger model update. S2c: When an update is triggered, relevant runtime data is collected to form a cluster of candidate patterns, which is then submitted to the knowledge graph for semantic review. Based on the review results, one of the following operations is performed: If the review identifies a new scenario, the candidate pattern cluster will be used as a new sample to generate a new semantic scenario pattern, which will then be used to fine-tune the model. If the review indicates a known scenario evolution, the data of the candidate pattern clusters will be merged into the corresponding scenario, and the model will be incrementally learned.

[0061] In some embodiments, step S3 includes: S31. Parse the key performance requirement description to obtain the various performance indicators and quantitative requirements contained therein; take the scenario type entity corresponding to the application scenario category as the starting node, traverse along the requirement relationship edge in the knowledge graph, verify the various performance indicators, and extract all directly related performance indicator entities and their preset constraints or optimization objectives to form an initial performance requirement constraint set. S32. Based on real-time operating data and vehicle status, retrieve environmental factor entities and vehicle component entities that are connected to the current scene and performance indicators through influence relationships and association relationships in the knowledge graph; quantify their real-time status values ​​into boundary parameters, and merge them with the performance requirement constraint set to formally define them as a multi-objective optimization problem containing multiple optimization objectives and constraints. S33. Using a preset vehicle system simulation model, the multi-objective optimization problem is solved; a multi-objective evolutionary algorithm is used to iteratively search in the decision space, and the performance and constraint satisfaction of each set of candidate parameters are evaluated through the simulation model, and finally a Pareto optimal solution set that is uniformly distributed in the objective space is generated, where each solution corresponds to a set of candidate control parameters. S34. Based on the preset driving strategy, select a final control parameter solution from the Pareto optimal solution set; convert the solution into a standardized control instruction set that conforms to the vehicle's underlying controller communication protocol and interface specifications, and output it.

[0062] In this embodiment of the invention, the specific implementation process of combining real-time operating data, performing inference and optimization calculations, generating and outputting a set of vehicle control parameters adapted to the scenario is as follows: (1) Quantification of the demand constraint extraction process: In the knowledge graph, obtain the scene type entity corresponding to the application scene category. .

[0063] Traversal from All demand relationships starting from The target performance index entity set { is obtained. }

[0064] For each performance metric entity It reads the quantization target or constraint defined in its attributes and concretizes it into a mathematical form based on real-time running data: If defined as maximizing the objective, it then transforms into an objective function: ,in The efficiency value is calculated from the simulation model, and X is the control parameter vector. That is equivalent to .

[0065] If defined as a constraint, it is transformed into an inequality constraint: Here is the battery temperature. Less than the preset temperature threshold ; The performance requirement constraint set Therefore, it is defined as: That is, the combination of the objective function set and the constraint set.

[0066] (2) Dynamic environmental constraint integration process: In a knowledge graph, from scene entities and performance index entity set { Starting from}, perform multi-hop reasoning along influence and association relationships to obtain relevant environmental factor entities { } and vehicle component state entity { }

[0067] According to the entity { }and{ The type of} is matched with and its corresponding real-time value is read from the real-time running data stream. and .

[0068] These real-time values ​​are then substituted into a preset constraint rule template to generate dynamic boundary constraints. For example, the rule template might be: If... For the road slope, then , For tolerance constraints; Combining dynamic boundary constraints with the vehicle's inherent physical limits (Such as the maximum torque of the motor and the maximum current of the battery) are combined to form a system constraint set. .

[0069] (3) Simulation-based multi-objective evolutionary solution process: Definition of decision space: The decision space The domain is defined by the vehicle control parameter vector X to be optimized, for example, X = [motor torque distribution coefficient, regenerative braking intensity, thermal management fan speed], and X∈ .

[0070] Formal definition of the optimization problem: Integrating the above process, the multi-objective optimization problem is formally defined as follows:

[0071]

[0072] in, For the target vector, and These are inequality and equality constraint vectors, respectively, derived from... and .

[0073] A decomposition-based multi-objective evolutionary algorithm is used to solve the problem, and each generation iterates as follows: Based on the current population, offspring individuals (i.e., new control parameter vectors Xnew) are generated through crossover and mutation operators.

[0074] Simulation model evaluation: Each new individual Xnew is input into the vehicle system simulation model. Based on the current vehicle state and Xnew, the model calculates the predicted state trajectory of the vehicle over a future period through numerical integration or table lookup, and extracts various performance index values ​​from it to calculate F(Xnew) and constraint violation degree.

[0075] Individuals that violate constraints are handled using either a penalty function method or a feasibility-first criterion. For feasible individuals, the penalty function is applied based on their objective vector. The problem is scalarized and assigned to the subproblems using Chebyshev decomposition or weighted summation.

[0076] From the merged parent and offspring populations, the next generation is selected based on the performance of each individual on each sub-problem.

[0077] The iteration terminates when the number of iterations reaches the preset maximum value, or when the rate of change of the distribution index of the Pareto solution set is less than the threshold, and the final non-dominated solution set is output as the Pareto optimal solution set.

[0078] (4) Strategy decision-making and instruction generation process: Based on the preset driving strategy, define a utility function U(F), or set a priority weight vector W for each objective. For each candidate solution in the optimal solution set, calculate its utility value U or weighted sum W·F.

[0079] The candidate solution that maximizes U or minimizes W·F is selected as the final control parameter solution Xfinal.

[0080] According to the target controller's communication protocol, each parameter value in Xfinal is encapsulated into a specific data frame format according to a predefined signal mapping table, scaling factor, and offset. For example, the torque distribution coefficient (0-1) is converted into a 2-byte CAN signal.

[0081] The packaged standardized control instruction set is sent to the corresponding controller via the vehicle bus.

[0082] like Figure 2 As shown, this embodiment of the invention also provides a scene-aware adaptive control system for new energy commercial vehicles, including: Data acquisition module: used to acquire multi-dimensional real-time operating data of new energy commercial vehicles; Feature extraction module: used to extract features from the multidimensional real-time running data to form a real-time feature vector; Scene recognition module: It has a built-in pre-built scene recognition model; the scene recognition module is used to input the real-time feature vector into the scene recognition model, output the application scenario category of the current vehicle and the description of key performance requirements associated with the knowledge graph; and during the application of the model, if the deviation between the real-time feature vector and the existing scene pattern representation is detected to exceed a preset threshold, it triggers the iterative update of the scene recognition model. Strategy reasoning engine: Based on the application scenario category and key performance requirement description, it traverses the logical relationships between entities in the knowledge graph associated with the current scenario, and combines real-time running data to perform reasoning and optimization calculations to generate a set of vehicle control parameters adapted to the scenario. Control output module: used to output the set of vehicle control parameters for dynamic adjustment of control parameters of at least one of the vehicle's electric drive system, battery management system and thermal management system, so as to realize scene-based adaptive control.

[0083] In some embodiments, the scene recognition model of the scene recognition module is constructed using the following units: Historical data processing unit: used to acquire historical multidimensional operating feature data of new energy commercial vehicles, and to automatically group the historical multidimensional operating feature data using a semi-supervised clustering algorithm with introduced constraints, to mine potential scene patterns and form an initial scene pattern cluster; Knowledge Graph Construction Unit: Used to construct a knowledge graph in the field of new energy commercial vehicles (including scene types, vehicle components, performance indicators, environmental factor entities, and edges representing the logical relationships between them). Semantic alignment unit: used to semantically align and fuse the initial scene pattern clusters with the knowledge graph (mapping each initial scene pattern cluster to a scene type entity in the knowledge graph, and associating key performance requirement features based on the logical relationship of the entity to generate semantic scene patterns). Model training unit: used to construct a training dataset based on the semantic scene pattern (input samples are historical multidimensional running feature data of the initial scene pattern cluster, and output labels are scene type and key performance requirement descriptions), and use the training dataset to train a machine learning model to obtain the scene recognition model.

[0084] In some embodiments, the historical data processing unit includes: The data preprocessing subunit is used to acquire historical time-series operation data of new energy commercial vehicles (including vehicle status parameters, driving energy consumption sequence, geographical location trajectory and environmental perception information), and to clean, align and standardize the time-series operation data to obtain a standardized historical operation dataset. Feature construction subunit: Based on the standardized historical operation dataset, it is used to construct a multi-dimensional operation feature vector (including at least spatiotemporal distribution features, energy consumption load features and driving behavior features) to characterize the operation mode through time window sliding and statistical calculation. Constraint Construction Subunit: Based on labeled runtime logs, this unit determines the attribution relationships between at least some sample pairs and constructs a set of pairwise constraints to guide clustering, including Must-link and Cannot-link constraints. Clustering subunit: Used to input the multidimensional running feature vector and the pairwise constraint set into the improved semi-supervised clustering algorithm, and to minimize the objective function by iteratively updating the cluster center and the cluster affiliation of the samples, and finally dividing all samples into k disjoint initial scene pattern clusters.

[0085] The improved semi-supervised clustering algorithm iteratively optimizes the following objective function. To achieve clustering:

[0086] In the formula, The preset number of clusters, For the first Clusters, For the first Cluster The cluster center, For sample feature vectors Distance from the cluster center To constrain the penalty weighting coefficient for violations; The number of Must-link constraints violated. The number of Cannot-link constraints that are violated.

[0087] The knowledge graph construction unit includes: Data extraction subunit: Used to extract text data containing terms and descriptions related to scenarios, components, indicators and environments from multiple heterogeneous knowledge sources; Entity recognition subunit: used to identify and extract corresponding entities from the text data based on predefined entity types (scene type, vehicle component, performance index, environmental factors); Relationship Establishment Subunit: Used to establish edges representing logical relationships between extracted entities based on predefined semantic relationship patterns (demand relationship, composition relationship, influence relationship); The graph storage subunit is used to store the entities and relationships into a graph database to form a queryable knowledge graph; and to encapsulate service interfaces for the knowledge graph (supporting the retrieval of entities, relationships and attributes, and rule-based logical reasoning).

[0088] In some embodiments, the semantic alignment unit includes: Cluster center calculation subunit: used to calculate the mean of all sample feature vectors for each initial scene pattern cluster, and obtain the cluster center feature vector representing the pattern cluster; Entity semantic vector generation subunit: used to generate corresponding entity semantic feature vectors for each scene type entity in the knowledge graph, based on its associated entities and relationships; Similarity matching subunit: used to calculate the similarity between the cluster center feature vector of each initial scene pattern cluster and the entity semantic feature vector of each scene type entity in the knowledge graph, and select the scene type entity with the highest similarity as the candidate mapping entity for each pattern cluster; Mapping and creating sub-units: This is used to determine whether the highest similarity of the candidate mapping entities exceeds the preset confidence threshold; if it does, then demand feature association is performed; if it does not exceed the threshold, then the initial scene pattern cluster is marked as an unmatched new pattern cluster, and after creating a new scene type entity in the knowledge graph based on its cluster center feature vector, demand feature association is performed. Demand Feature Association Subunit: Starting from the mapped scenario type entity, it traverses along the demand relationship edges of the knowledge graph to obtain all directly associated performance indicator entities and their attribute values, which serve as the key performance demand features associated with this pattern cluster; and encapsulates the initial scenario pattern cluster that has completed the mapping and demand feature association into a semantic scenario pattern (containing the pattern cluster identifier, the mapped scenario type entity, and the set of key performance demand features).

[0089] In some embodiments, the model training unit includes: Training data construction sub-unit: For each semantic scene mode, samples are selected from the historical multi-dimensional running feature data of its corresponding initial scene mode cluster. The feature vector of the sample is used as the input feature, and the scene type label and key performance requirement feature vector of the semantic scene mode are used as supervision labels to form training data; the collection of all training data constitutes the training dataset. Model structure definition sub-units: used to build multi-task neural network models (including: shared feature encoding layer - encoding input feature vectors; scene classification output layer - connecting to the shared layer and outputting scene type classification results; demand regression output layer - connecting to the shared layer and outputting regression prediction values ​​of key performance demand feature vectors). Model training subunit: used to train the multi-task neural network model using the training dataset, optimize the joint loss function through backpropagation algorithm until the model converges, and obtain the scene recognition model; Joint loss function

[0090] in, The classification loss corresponding to the scene classification task. The regression loss is used to predict the characteristics of the demand. This is a coefficient used to balance the weights of the two tasks.

[0091] In some embodiments, the iterative update unit of the scene recognition module includes: New data pattern generation subunit: used during online model runtime, to extract features from new input runtime data and form the current real-time feature vector; Deviation calculation subunit: used to calculate the Mahalanobis distance between the current real-time feature vector and each existing scene mode representation in the model; take the minimum value among all Mahalanobis distances, and if the minimum value is greater than a preset threshold, trigger model update; Pattern Review and Update Subunit: When an update is triggered, this subunit collects relevant runtime data to form a cluster of candidate patterns and submits it to the knowledge graph for semantic review; based on the review results, it performs one of the following operations: (1) If the review is a new scenario: the candidate pattern cluster is used as a new sample to generate a new semantic scene pattern, which is then used to fine-tune the model; (2) If the review is a known scenario evolution: merge the data of the candidate pattern cluster into the corresponding scenario and perform incremental learning on the model.

[0092] In some embodiments, the policy reasoning engine includes: Constraint Extraction Subunit: Used to parse the key performance requirement description, obtain the various performance indicators and quantitative requirements contained therein; taking the scenario type entity corresponding to the application scenario category as the starting node, traverse along the requirement relationship edge in the knowledge graph to verify the various performance indicators, and extract all directly related performance indicator entities and their preset constraints or optimization objectives to form an initial performance requirement constraint set. The optimization problem formalization subunit is used to retrieve environmental factor entities and vehicle component entities that are connected to the current scene and performance indicators through influence relationships and correlation relationships in the knowledge graph based on real-time running data and vehicle status; quantify their real-time state values ​​into boundary parameters, and merge them with the initial performance requirement constraint set to formalize them into a multi-objective optimization problem containing multiple optimization objectives and constraints. The optimization solution subunit is used to solve the multi-objective optimization problem using a preset vehicle system simulation model. It employs a multi-objective evolutionary algorithm to iteratively search within the decision space, evaluates the performance and constraint satisfaction of each set of candidate parameters through the simulation model, and finally generates a Pareto optimal solution set (each solution corresponds to a set of candidate control parameters). Instruction generation subunit: used to select a final control parameter solution from the Pareto optimal solution set according to the preset driving strategy; convert the solution into a standardized control instruction set that conforms to the vehicle's underlying controller communication protocol and interface specification, and output it to the control output module.

[0093] Those skilled in the art will clearly understand that the techniques in the embodiments of the present invention can be implemented using software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solutions in the embodiments of the present invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium such as a USB flash drive, mobile hard drive, read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk, or other media capable of storing program code. It includes several instructions to cause a computer terminal (which may be a personal computer, server, or a second terminal, network terminal, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0094] In the embodiments provided by this invention, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0095] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0096] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0097] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A scene-aware adaptive control method for new energy commercial vehicles, characterized in that, Includes the following steps: S1. Obtain multi-dimensional real-time operating data of new energy commercial vehicles, and extract features from the operating data to form a real-time feature vector; S2. Input the real-time feature vector into the scene recognition model; The scene recognition model outputs the recognition result of the application scene category of the current vehicle, as well as the description of key performance requirements obtained from the knowledge graph; during the application of the model, if the deviation between the current real-time feature vector and the existing scene pattern representation exceeds a preset threshold, iterative updates of the scene recognition model are triggered. S3. Based on the application scenario category and key performance requirements description, invoke the strategy reasoning engine; The strategy reasoning engine parses the key performance requirement description and, in conjunction with real-time running data, performs reasoning and optimization calculations by traversing the logical relationships between entities in the knowledge graph associated with the current scenario, generating a set of vehicle control parameters adapted to the scenario. S4. Output the vehicle control parameter set to dynamically adjust the control parameters of at least one of the vehicle's electric drive system, battery management system, and thermal management system, thereby achieving scenario-based adaptive control.

2. The adaptive control method for new energy commercial vehicles based on scene perception according to claim 1, characterized in that, The scene recognition model is constructed in the following way: S021. Obtain historical multi-dimensional operational feature data of new energy commercial vehicles, and use a semi-supervised clustering algorithm with introduced constraints to automatically group the historical multi-dimensional operational feature data, mine potential scene patterns, and form an initial scene pattern cluster. S022. Construct a knowledge graph in the field of new energy commercial vehicles. The knowledge graph includes scene types, vehicle components, performance indicators, environmental factor entities, and edges representing the logical relationships between them. S023. Semantically align and fuse the initial scene pattern clusters with the constructed knowledge graph; specifically, this includes mapping each initial scene pattern cluster to a scene type entity in the knowledge graph, and based on the logical relationship of the entity in the knowledge graph, associating a set of key performance requirement features to generate semantic scene patterns. S024. Based on the semantic scene pattern, construct a training dataset; wherein the input sample is the historical multidimensional running feature data in the initial scene pattern cluster, and the output label is the corresponding scene type and associated key performance requirement description; use the training dataset to train a machine learning model to obtain a scene recognition model, which can output the recognized scene category and key performance requirement description according to the input feature data.

3. The adaptive control method for new energy commercial vehicles based on scene perception according to claim 2, characterized in that, The steps in S021 include: S021a. Obtain historical time-series operational data of new energy commercial vehicles, including vehicle status parameters, driving energy consumption sequences, geographical location trajectories, and environmental perception information; clean, align, and standardize the time-series operational data to obtain a standardized historical operational dataset. S021b. Based on a standardized historical operation dataset, a multi-dimensional operation feature vector is constructed to characterize the operation mode through time window sliding and statistical calculation; the multi-dimensional operation feature vector includes at least spatiotemporal distribution features, energy consumption load features, and driving behavior features. S021c. Based on the labeled operation logs, determine the attribution relationship between at least some sample pairs and construct a set of pairwise constraints to guide clustering; the set of pairwise constraints includes: Must-link constraints containing sample pairs that must belong to the same scene pattern and Cannot-link constraints containing sample pairs that cannot belong to the same scene pattern. S021d: The multidimensional running feature vector and the set of pairwise constraints are input into an improved semi-supervised clustering algorithm; the algorithm iteratively updates the cluster centers and the cluster affiliation of samples, thereby improving the objective function. Reaching the minimum value, all samples are eventually divided into k A cluster of initial scene patterns that do not overlap.

4. The adaptive control method for new energy commercial vehicles based on scene perception according to claim 3, characterized in that, The improved semi-supervised clustering algorithm in S021d iteratively optimizes the following objective function. To achieve clustering: In the formula, The preset number of clusters, For the first Clusters, For the first Cluster The cluster center, For sample feature vectors Distance from the cluster center, To constrain the penalty weighting coefficient for violations; The number of Must-link constraints violated. The number of Cannot-link constraints that are violated.

5. The adaptive control method for new energy commercial vehicles based on scene perception according to claim 2, characterized in that, In S022, the steps for constructing a knowledge graph in the field of new energy commercial vehicles include: S022a. Extract text data containing terms and descriptions related to scenarios, components, indicators and environments from multiple heterogeneous knowledge sources; S022b: Based on predefined entity types, identify and extract scene type entities, vehicle component entities, performance index entities, and environmental factor entities from text data; S022c. Based on the predefined semantic relationship pattern, establish edges representing logical relationships between the extracted entities; the semantic relationships include at least: demand relationships representing specific performance requirements of the scenario, composition relationships representing membership or connection relationships between components, and influence relationships representing the impact of the environment on performance or components. S022d. Store entities and relationships in a graph database to form a queryable knowledge graph; and encapsulate service interfaces for the knowledge graph, which support the retrieval of entities, relationships and attributes and rule-based logical reasoning.

6. The adaptive control method for new energy commercial vehicles based on scene perception according to claim 2, characterized in that, The steps in S023 include: S023a. For each initial scene pattern cluster, calculate the mean of all its sample feature vectors to obtain the cluster center feature vector representing the pattern cluster. S023b. For each scene type entity in the knowledge graph, generate its corresponding entity semantic feature vector based on its associated entities and relationships. S023c, Calculate the similarity between the cluster center feature vector of each initial scene pattern cluster and the entity semantic feature vector of each scene type entity in the knowledge graph; select the scene type entity with the highest similarity as its candidate mapping entity for each pattern cluster; S023d. Determine whether the highest similarity corresponding to the candidate mapped entity exceeds the preset confidence threshold: If so, then execute S023e; If not, mark the initial scene pattern cluster as an unmatched new pattern cluster, and create a new scene type entity in the knowledge graph based on its cluster center feature vector before executing S023e. S023e. For each initial scenario pattern cluster that has completed mapping, in the knowledge graph, starting from the mapped scenario type entity, traverse along the demand relationship edge to obtain all directly related performance indicator entities and their attribute values, which are used as the key performance requirement features associated with the pattern cluster. S023f. The initial scenario pattern cluster that has been mapped and associated with the requirement features is encapsulated into a semantic scenario pattern; the semantic scenario pattern includes at least: the identifier of the pattern cluster, the mapped scenario type entity, and the associated set of key performance requirement features.

7. The adaptive control method for new energy commercial vehicles based on scene perception according to claim 2, characterized in that, S024 specifically includes: S024a. For each semantic scene pattern, select samples from the historical multidimensional running feature data of its corresponding initial scene pattern cluster, use the feature vector of the sample as the input feature, and use the scene type label and key performance requirement feature vector contained in the semantic scene pattern as the supervision label to form a training data; the set of all training data constitutes the training dataset. S024b. Construct a multi-task neural network model, the model including: A shared feature encoding layer is used to encode the input feature vector; A scene classification output layer is connected to a shared feature encoding layer to output the classification result of the scene type; A demand regression output layer is connected to a shared feature encoding layer to output regression predictions of key performance demand feature vectors; S024c. Train the multi-task neural network model using the training dataset, optimize the joint loss function L through the backpropagation algorithm until the model converges, and obtain the scene recognition model.

8. The adaptive control method for new energy commercial vehicles based on scene perception according to claim 1, characterized in that, The steps for model iterative updates include: S2a: When the model is running online, features are extracted from the newly input running data to form the current real-time feature vector; S2b: Calculate the Mahalanobis distance between the current real-time feature vector and each existing scene mode representation in the model; take the minimum value among all Mahalanobis distances; if the minimum value is greater than a preset threshold, trigger model update. S2c: When an update is triggered, relevant runtime data is collected to form a cluster of candidate patterns, which is then submitted to the knowledge graph for semantic review. Based on the review results, one of the following operations is performed: If the review is for a new scenario, the candidate pattern cluster will be used as a new sample to generate a new semantic scenario pattern, which will be used to fine-tune the model. If the review indicates a known scenario evolution, the data of the candidate pattern clusters will be merged into the corresponding scenario, and the model will be incrementally learned.

9. The adaptive control method for new energy commercial vehicles based on scene perception according to claim 1, characterized in that, The steps in S3 include: S31. Parse the key performance requirement description to obtain the various performance indicators and quantitative requirements it contains; take the scenario type entity corresponding to the application scenario category as the starting node, traverse along the requirement relationship edge in the knowledge graph, verify each performance indicator, and extract all directly related performance indicator entities and their preset constraints or optimization objectives to form the initial performance requirement constraint set. S32. Based on real-time operation data and vehicle status, retrieve environmental factor entities and vehicle component entities that are connected to the current scene and performance indicators through influence relationships and association relationships in the knowledge graph; quantify their real-time status values ​​into boundary parameters, and integrate them with the performance requirement constraint set to formally define them as a multi-objective optimization problem containing multiple optimization objectives and constraints. S33. Using a pre-set vehicle system simulation model, solve the multi-objective optimization problem; adopt a multi-objective evolutionary algorithm to perform iterative search in the decision space, evaluate the performance and constraint satisfaction of each set of candidate parameters through the simulation model, and finally generate a Pareto optimal solution set that is uniformly distributed in the objective space, where each solution corresponds to a set of candidate control parameters. S34. Based on the preset driving strategy, select a final control parameter solution from the Pareto optimal solution set; convert the solution into a standardized control instruction set that conforms to the vehicle's underlying controller communication protocol and interface specifications, and output it.

10. A scene-aware adaptive control system for new energy commercial vehicles, characterized in that, include: Data acquisition module: used to acquire multi-dimensional real-time operating data of new energy commercial vehicles; Feature extraction module: used to extract features from multidimensional real-time running data to form real-time feature vectors; Scene recognition module: It has a built-in pre-built scene recognition model; the scene recognition module is used to input real-time feature vectors into the scene recognition model and output the application scenario category of the current vehicle and the description of key performance requirements associated with the knowledge graph. During the application of the model, if the deviation between the real-time feature vector and the existing scene pattern representation exceeds a preset threshold, iterative updates to the scene recognition model will be triggered. Strategy reasoning engine: Based on the application scenario category and key performance requirements description, it traverses the logical relationships between entities in the knowledge graph associated with the current scenario, and combines real-time running data to perform reasoning and optimization calculations to generate a set of vehicle control parameters adapted to the scenario. Control output module: Used to output a set of vehicle control parameters for dynamic adjustment of control parameters of at least one of the vehicle's electric drive system, battery management system and thermal management system, to achieve scenario-based adaptive control.

Images