Semantic understanding model training and adaptive fine-tuning method for automobile vertical field

Abstract

This invention provides a method for training and adaptive fine-tuning a semantic understanding model for the automotive vertical domain, belonging to the field of in-vehicle semantic understanding model training technology. The method includes: constructing a vehicle function state transition graph based on acquired multi-source data from the automotive domain; determining a minimum state repair set to convert a target vehicle function from non-executable to executable; forming boundary state pairs between the vehicle states before and after applying the minimum state repair set; training a basic automotive domain model with a vehicle function adapter using initial pairwise training tasks to obtain a semantic understanding model; and generating error correction pairwise training tasks based on the error function boundaries and updating the vehicle function adapter corresponding to the error function boundaries. This invention solves for the minimum state repair set based on the vehicle function state transition graph and generates initial pairwise training tasks using the boundary state pairs before and after repair, enabling the basic automotive domain model to distinguish vehicle states with the same semantics but different processing strategies.

Description

Technical Field

[0001] This invention relates to the field of vehicle-mounted semantic understanding model training technology, specifically to a method for training and adaptive fine-tuning semantic understanding models for the automotive vertical domain. Background Technology

[0002] With the rapid development of intelligent connected vehicle technology, in-vehicle voice assistants and intelligent cockpit interaction systems have become core configurations of modern automobiles. These systems rely on large-scale artificial intelligence (AI) language models to understand users' natural language commands, enabling more natural and intelligent human-vehicle interaction. Currently, the industry commonly uses large models pre-trained on massive general-purpose corpora and applies them to the automotive vertical domain through transfer learning techniques. However, such methods struggle to learn different processing strategies corresponding to the same user command on both sides of the boundary. When a decision error occurs at a specific functional boundary, it usually requires overall fine-tuning of the model, which can easily alter the existing semantic representations and processing results of other vehicle functions. Summary of the Invention

[0003] The purpose of this invention is to provide a method for training and adaptive fine-tuning a semantic understanding model for the automotive vertical domain, which enables the in-vehicle semantic understanding model to accurately learn the state dependencies of vehicle functions and automatically correct decision boundary errors without interfering with other functions.

[0004] To achieve the above objectives, this invention provides a method for training and adaptive fine-tuning a semantic understanding model for the automotive vertical domain, comprising: constructing a vehicle function state transition diagram based on acquired multi-source data from the automotive domain; determining a minimum state repair set based on the vehicle function state transition diagram to convert a target vehicle function from non-executable to executable, and forming boundary state pairs by combining the vehicle states before and after applying the minimum state repair set; combining the same user instruction with the boundary state pairs to generate initial pairwise training tasks with the same semantic label but different processing strategy labels; training an automotive domain basic model equipped with a vehicle function adapter using the initial pairwise training tasks, so that the semantic representation of the same user instruction is consistent under the boundary state pairs, and the output processing strategy matches the corresponding state transition rule, thereby obtaining a semantic understanding model; comparing the processing strategy output by the semantic understanding model with the processing strategy determined based on the vehicle function state transition diagram, locating the erroneous function boundary based on the comparison result, generating an error correction pairwise training task based on the erroneous function boundary, and updating the vehicle function adapter corresponding to the erroneous function boundary.

[0005] Optionally, the step of constructing a vehicle function state transition diagram based on the acquired multi-source data in the automotive field includes: extracting vehicle functions, vehicle state variables, execution conditions, control actions, and target states from the multi-source data in the automotive field; constructing vehicle state nodes using the state combinations of the vehicle state variables; establishing function transition edges between the vehicle state nodes based on the execution conditions; and associating the corresponding vehicle functions, control actions, and target states with the function transition edges.

[0006] Optionally, determining the minimum state repair set that enables the target vehicle function to be converted from non-executable to executable based on the vehicle function state transition diagram includes: obtaining the current vehicle state and the execution conditions of the target vehicle function; determining the state conditions that the current vehicle state does not meet from the execution conditions; searching for candidate repair combinations based on the vehicle state variables corresponding to the state conditions to obtain multiple candidate state repair sets; and determining the candidate state repair set containing the fewest vehicle state variables as the minimum state repair set from the candidate state repair set that enables the target vehicle function to be converted to executable and meets the vehicle state constraints.

[0007] Optionally, the boundary state pair includes a first state and a second state; the first state does not satisfy the execution conditions of the target vehicle function, the second state satisfies the execution conditions, and when the transition from the first state to the second state occurs, only the vehicle state variables within the minimum state repair set change.

[0008] Optionally, generating initial paired training tasks with the same semantic labels but different processing strategy labels includes: combining the same user instruction with the first state and the second state respectively to generate a first training task and a second training task; setting the same intent label and slot label for the first training task and the second training task; determining the processing strategy labels for the first training task and the second training task respectively based on the executable status of the target vehicle function in the first state and the second state, wherein the processing strategy labels for the first training task and the second training task are different; and associating the first training task and the second training task as the initial paired training tasks.

[0009] Optionally, the automotive domain basic model includes a shared semantic encoding layer, a semantic projection layer, a decision projection layer, and multiple vehicle function adapters, each vehicle function adapter corresponding to at least one vehicle function.

[0010] Optionally, training the automotive domain base model equipped with a vehicle function adapter using the initial pairwise training task includes: encoding each training task in the initial pairwise training task through the shared semantic encoding layer; generating semantic representations and decision representations for each training task through the semantic projection layer and the decision projection layer; determining semantic consistency loss based on the differences between the semantic representations, determining decision separation loss based on the differences between the decision representations, and determining state transition consistency loss based on the differences between the processing strategy output by the automotive domain base model and the corresponding state transition rule; updating the automotive domain base model based on the semantic consistency loss, decision separation loss, and state transition consistency loss to match the output processing strategy with the corresponding state transition rule, and to reduce the differences between the semantic representations and increase the differences between the decision representations.

[0011] Optionally, when training the automotive domain basic model, the corresponding vehicle function adapter is selected for parameter update based on the target vehicle function corresponding to the initial pairwise training task, and the parameters of vehicle function adapters unrelated to the target vehicle function are frozen.

[0012] Optionally, locating the erroneous functional boundary based on the comparison result includes: obtaining the current vehicle state and the output processing strategy when the semantic understanding model processes the user instruction; determining the corresponding state transition rule from the vehicle functional state transition diagram based on the current vehicle state, and determining the expected processing strategy based on whether the current vehicle state meets the execution conditions of the state transition rule; when the output processing strategy is inconsistent with the expected processing strategy, determining the functional boundary corresponding to the state transition rule as the erroneous functional boundary.

[0013] Optionally, the step of generating error correction pairwise training tasks and updating the corresponding vehicle function adapters based on the error function boundary includes: combining the user instruction with boundary state pairs that cross the error function boundary to generate error correction pairwise training tasks; keeping the parameters in the automotive domain base model unchanged except for the vehicle function adapters corresponding to the error function boundary, and updating the corresponding vehicle function adapters using the error correction pairwise training tasks; validating the semantic understanding model configured with the updated vehicle function adapters using a target boundary validation set, a target function preservation set, and a non-target function protection set; deploying the updated vehicle function adapters when the validation passes, and retaining the vehicle function adapters before the update when the validation fails.

[0014] The above technical solution solves the minimum state repair set based on the vehicle function state transition graph, and uses the boundary state pairs before and after repair to generate initial pairwise training tasks. This enables the basic model in the automotive domain to distinguish vehicle states with the same semantics but different processing strategies. It also locally updates the vehicle function adapter according to the erroneous function boundary, which improves the accuracy of the semantic understanding model in recognizing the function execution boundary and the efficiency of fine-tuning.

[0015] Other features and advantages of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the following detailed description to explain the invention, but do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a semantic understanding model training and adaptive fine-tuning method for the automotive vertical industry.

[0017] Figure 2 This is a flowchart of the generation process for the minimum state repair set and boundary state pairs. Detailed Implementation

[0018] The following is in conjunction with the appendix Figure 1 -Appendix Figure 2 The specific implementation methods of the embodiments of the present invention will be described in detail below. It should be understood that the specific implementation methods described herein are only for illustrating and explaining the embodiments of the present invention, and are not intended to limit the embodiments of the present invention.

[0019] It should be noted that the acquisition, transmission, storage, use, and processing of data in the technical solution of this application all comply with the relevant provisions of national laws and regulations. In the embodiments of this application, certain existing industry solutions such as software, components, and models may be mentioned. These should be considered exemplary, intended only to illustrate the feasibility of implementing the technical solution of this application, and do not imply that the applicant has already used or necessarily used such solutions.

[0020] In the process of realizing this invention, the inventors of this application discovered that existing semantic understanding models in the automotive field have difficulty distinguishing different processing strategies for the same user instruction under different vehicle states on both sides of the function execution boundary. When errors occur, overall fine-tuning is usually required, resulting in low adaptation efficiency.

[0021] Example 1 Reference Figures 1-2 This is the first embodiment of the present invention, which provides a method for training and adaptive fine-tuning a semantic understanding model for the automotive vertical domain, including: S100: Construct a vehicle function state transition diagram based on the acquired multi-source data from the automotive field.

[0022] In the embodiments of this application, vehicle functions, vehicle state variables, execution conditions, control actions, and target states are extracted from multi-source data in the automotive field. The acquisition, storage, and use of multi-source data in the automotive field comply with relevant laws and regulations, and user authorization has been obtained and necessary anonymization and desensitization processing has been implemented.

[0023] In a preferred embodiment of this application, vehicle state variables are selected and combined to construct vehicle state nodes. Functional transition edges between vehicle state nodes are established based on execution conditions, and the corresponding vehicle functions, control actions, and target states are mapped and associated with these functional transition edges. Specifically, for each vehicle function, only vehicle state variables related to the execution conditions, control actions, and target states of that vehicle function are collected to construct vehicle state nodes. Vehicle state variables unrelated to that vehicle function are stored as state context and do not participate in the combination of vehicle state nodes. The vehicle function state transition graph only records combinations of vehicle states belonging to the legal state domain, and the combination explosion problem caused by full state traversal is avoided through a variable pruning strategy.

[0024] In practical implementation, the legal state domain Ω is constructed using a constraint satisfaction problem (CSP) solver: the value range of state variables is defined by the vehicle signal dictionary, and the constraint set is defined by the rules of state mutual exclusion, linkage and vehicle configuration; the CSP solver engine is called to traverse and solve the problem, filtering out illegal combinations that violate physical logic, and combined with the state snapshots collected from the actual vehicle road to verify, generating a legal state database as the legal state domain Ω.

[0025] In one specific embodiment, multi-source data in the automotive field is uniformly acquired by the data processing center. The data sources are divided into two categories: one is business structured data such as vehicle function specifications, vehicle signal dictionaries, and control interface documents; the other is unstructured text data such as automotive vertical platform text, official corporate social media interaction text, historical human-vehicle interaction logs, and corporate-owned labeled data.

[0026] Among them, platform text and social media interaction text are used to supplement automotive terminology, colloquial expressions, and function aliases, expanding the coverage of semantic understanding; historical human-vehicle interaction logs record the function triggering logic and state associations in real-world vehicle usage scenarios, providing real-vehicle evidence for valid state domain verification and erroneous function boundary location; enterprise-owned labeled data provides labeled samples of intents, slots, and processing strategies, serving as supervisory signals for model training; vehicle function specifications, signal dictionaries, and control interface documents are used to clarify the execution conditions, control actions, and target states of each function. Any content requiring confirmation due to rule ambiguity or semantic vagueness must be manually reviewed and verified before being included in the subsequent modeling process.

[0027] Furthermore, the data processing center performs a series of standardized preprocessing operations on the acquired multi-source data, including: unified encoding, deduplication, invalid character cleanup, sensitive field anonymization, vehicle model version marking, and entity alignment. Through entity alignment, the data processing center maps heterogeneous content with the same meaning but different names to a unified automotive terminology database.

[0028] Specifically, the unified automotive terminology database adopts a structured design, recording key attributes such as function identifiers, function aliases, vehicle state variables and their value ranges, control actions, target states, and vehicle models, while retaining original fields for traceability. For example, in the actual mapping process, the data processing center will uniformly identify the colloquial term "tailgate," the term "rear hatch" in the repair manual, and the term "trunk lid" in the user interface as the same functional entity, thereby unifying the functional entity identifiers from different data sources and ensuring that subsequent model training is based on a consistent and standardized semantic foundation.

[0029] For any vehicle function f, the vehicle state node that meets the execution conditions is connected to the target vehicle state node after the control action is executed through a function transition edge. Each function transition edge records the vehicle function f and the execution conditions. Controlling actions and target state This forms the state transition rules. Execution conditions It consists of atomic state conditions and the logical relationships between them. Each atomic state condition is associated with a condition boundary identifier when it transitions from being unsatisfied to being satisfied. The vehicle functional state transition diagram also records the processing strategy and target state preservation rules when execution conditions are not satisfied. For vehicle state s, if If it is established, then =1; if If not, then =0.

[0030] In one specific embodiment, taking the trunk opening function as an example, the enterprise function rules stipulate that the execution conditions for this function include the vehicle being in a gear that allows trunk opening, the vehicle speed meeting the opening conditions, and the trunk not being in a fault-locked state. The vehicle function state transition diagram constructs vehicle state nodes by combining the state of vehicle state variables related to the execution conditions, control actions, and target states of the trunk opening function. Vehicle state nodes that meet all execution conditions are connected to the target state after the trunk is opened via function transition edges, and the trunk opening control action is associated on the function transition edges. For vehicle state nodes that do not meet the execution conditions, a non-execution control action and a corresponding reason prompt are associated. All of the above execution conditions are derived from the enterprise function rules for the corresponding vehicle model.

[0031] The above solution organizes multi-source data in the automotive field into a vehicle functional state transition diagram, transforming scattered functional rules into state transition criteria that can be jointly invoked by the training and decision verification processes.

[0032] S200: Based on the vehicle function state transition diagram, determine the minimum state repair set that transforms the target vehicle function from non-executable to executable, and form boundary state pairs by combining the vehicle states before and after applying the minimum state repair set.

[0033] In the embodiments of this application, the current vehicle state and the execution conditions of the target vehicle function are obtained; the state conditions that the current vehicle state does not meet are determined from the execution conditions; candidate repair combinations are searched based on the vehicle state variables corresponding to the state conditions to obtain multiple candidate state repair sets; from the candidate state repair sets that can convert the target vehicle function into an executable state that satisfies the vehicle state constraints, the candidate state repair set containing the fewest vehicle state variables is determined as the minimum state repair set.

[0034] The boundary state pair includes a first state and a second state; the first state does not meet the execution conditions of the target vehicle function, while the second state meets the execution conditions, and when the transition from the first state to the second state occurs, only the vehicle state variables within the minimum state repair set change.

[0035] It should be noted that, to avoid generating state combinations that cannot simultaneously occur in real vehicles, the search process must be conducted within the legal state domain Ω pre-constructed in S100. The vehicle state obtained after applying a candidate repair combination to the first state is only included in the comparison as a candidate state repair set if it belongs to the legal state domain Ω. For vehicle state variables restricted by vehicle model version, only the state values ​​allowed for the corresponding vehicle model version are used.

[0036] In one optional implementation, let the first state where the target vehicle's function execution conditions are not met be... The candidate repair status participating in the search is , and Let represent the values ​​of the j-th vehicle state variable in the candidate repair states of the j-th vehicle state variable in the first state, respectively. The second state is as follows:

[0037] The minimum state repair set is as follows:

[0038] in, This indicates the first state where the conditions for executing the target vehicle's functions are not met. Indicates the status of candidate repairs participating in the search. The second state represents the state with the fewest vehicle state variables that meet the execution conditions of the target vehicle function and have changed relative to the first state; Ω represents the legal state domain established based on the value range of vehicle state variables, state mutual exclusion relationships, state linkage relationships, and vehicle model version constraints; f represents the vehicle function. Indicates the execution conditions. Indicates candidate repair status The execution conditions for the target vehicle function f are met, where N represents the total number of vehicle state variables participating in the state repair search, and j represents the index of the vehicle state variable. This represents the value of the j-th vehicle state variable in the first state. This represents the value of the j-th vehicle state variable in the candidate repair states. This represents the value of the j-th vehicle state variable in the second state. The function is denoted as argmin, which takes the value 1 when the condition in parentheses is true and 0 when the condition in parentheses is false; argmin represents the candidate repair state with the smallest objective function value selected from the candidate repair states that meet the constraints. This represents the minimum state repair set corresponding to the target vehicle function f. The minimum state repair set consists of the vehicle state variable indices whose values ​​differ between the first state and the second state.

[0039] It should be noted that the above calculation yields the candidate repair state set that minimizes the number of changes in vehicle state variables. When the candidate repair state set contains only one candidate repair state, that candidate repair state is designated as the second state, and the second state is paired with the first state to form a boundary state pair. When the candidate repair state set contains multiple candidate repair states, each candidate repair state is designated as the second state, and each second state is paired with the same first state to form a corresponding boundary state pair. For each boundary state pair, the vehicle state variables whose values ​​differ between the second state and the first state form the corresponding minimum state repair set.

[0040] In a preferred embodiment of this application, to reduce the search range of candidate repair states, it is necessary to first determine the vehicle state variables corresponding to the unmet state conditions of the current vehicle state, as well as the associated vehicle state variables with state linkage relationships. These determined vehicle state variables are then used as search variables, while the remaining vehicle state variables retain their values ​​from the first state. During the search process, constraint propagation is performed based on the value range of vehicle state variables, state mutual exclusion relationships, state linkage relationships, and vehicle model version constraints. When the vehicle state formed by the candidate search branch does not belong to a legal state domain, the further expansion of the candidate search branch is terminated. Similarly, when, based on the allowed values ​​of the remaining searchable vehicle state variables, it is determined that further expansion of the candidate search branch still cannot satisfy the execution conditions of the target vehicle function, the further expansion of the candidate search branch is also terminated. A set of candidate repair states is obtained within the local state space related to the unmet state conditions, without traversing all state combinations of vehicle state variables. Candidate repair states can be obtained using graph search, branch and bound search, or constraint candidate set search, and simultaneously satisfy the execution conditions of the target vehicle function and legal state domain constraints.

[0041] For any pair of boundary states, the first and second states have the same vehicle model version and target vehicle function identifier. Except for the vehicle state variables within the corresponding minimum state repair set, the values ​​of the other vehicle state variables remain consistent. The execution conditions corresponding to the target vehicle function f are used. The first and second states are determined separately to obtain the following results: =0、 =1. Where, This indicates the first state where the conditions for executing the target vehicle's functions are not met. The second state represents the state with the fewest vehicle state variables that meet the target vehicle function execution conditions and have changed relative to the first state. =0 indicates that the first state does not meet the execution conditions of the target vehicle function. =1 indicates that the second state meets the execution condition. Therefore, each first state and its corresponding second state form a boundary state pair that spans the same vehicle function execution boundary.

[0042] In a specific embodiment, taking the aforementioned trunk opening function as an example, when the vehicle speed condition and fault lock condition in the first state are both met, and only the gear position condition is not met, the searched candidate repair states only change the gear position state variable. Each candidate repair state is determined as a second state, and each second state forms a boundary state pair with the first state. The minimum state repair set corresponding to each boundary state pair contains only the gear position state variable, while the vehicle speed, trunk fault lock state, and other vehicle state variables remain unchanged. When there are multiple unmet state conditions in the first state, according to the logical relationship between the state conditions, the search is conducted to find the candidate repair state set with the fewest vehicle state variables that can convert the target vehicle function into an executable state and change relative to the first state. Each candidate repair state in the candidate repair state set is determined as a second state, and each forms a boundary state pair with the first state.

[0043] The above scheme solves for the minimum state repair set that makes the target vehicle function change from unexecutable to executable, constructs boundary state pairs that only require changes in the vehicle state variables, and establishes a correspondence between changes in functional executability and minimum state changes, thereby obtaining training samples that can accurately characterize the functional execution boundary.

[0044] S300: Combine the same user instruction with the boundary state pair to generate initial paired training tasks with the same semantic label but different processing policy labels.

[0045] In embodiments of this application, generating initial paired training tasks with the same semantic labels but different processing strategy labels includes: combining the same user instruction with a first state and a second state respectively to generate a first training task and a second training task; setting the same intent label and slot label for the first training task and the second training task; determining the processing strategy labels for the first training task and the second training task respectively based on the executable status of the target vehicle function in the first state and the second state, wherein the processing strategy labels for the first training task and the second training task are different; and associating the first training task and the second training task as initial paired training tasks.

[0046] In a preferred embodiment of this application, the training task generation unit obtains user instructions corresponding to the target vehicle function based on the vehicle function identifier, associated historical user statements, standard function expressions, and manually confirmed synonyms. For the same user instruction, a first state and a second state are concatenated to form a first training task. Second training task Each training task includes not only user instructions and supervision labels, but also task instructions, vehicle state context, vehicle model version, functional boundary identifiers, paired task association identifiers, output field constraints, and loss routing identifiers.

[0047] The first and second training tasks can be represented as follows:

[0048]

[0049] in, This indicates the first training task. This represents the second training task, where p represents the task instruction, used to specify the processing requirements and output content of the automotive semantic understanding task, and u represents the user instruction used by both the first and second training tasks. Semantic labels are used to represent intent labels and slot labels. The same semantic labels are used for the first and second training tasks. This label represents the processing strategy label corresponding to the target vehicle function in the first state. The label indicates the processing strategy label corresponding to the target vehicle function in the second state; the two processing strategy labels are different. This represents the target state label formed after executing the processing strategy corresponding to the first training task. This represents the target state label formed after executing the processing strategy corresponding to the second training task; The identifier represents the functional boundary corresponding to the target vehicle function f, used to indicate the vehicle function execution boundary crossed by the first and second states. This indicates a pairwise task association identifier, used to mark the first training task and the second training task as the same initial pairwise training task.

[0050] In one optional implementation, the task instruction p includes a description of the vehicle semantic understanding role, a definition of the vehicle state field, output field constraints, and prohibited output content. The output fields include a function identifier, intent, slot, processing strategy, processing reason, control action, and target state. The training task generation unit, based on the vehicle function state transition diagram, combines the task instruction p, the user instruction u used in both the first and second training tasks, and the first state that does not meet the target vehicle function execution conditions. The model input for the first training task is formed by combining the input according to the preset input format. This includes the task instruction p, the user instruction u used by both the first and second training tasks, and the second state, which has the fewest number of vehicle state variables that meet the target vehicle function execution conditions and have changed relative to the first state. The model input for the second training task is formed by combining the same input format, and the corresponding semantic labels, processing strategy labels, processing reasons, control actions and target states are automatically generated. This avoids the tedious process of manually judging the executability of each function and improves the efficiency of training data construction.

[0051] In a specific embodiment, taking the user command "open the trunk" as an example, the function identifiers of both the first and second training tasks are "open trunk," the intent labels are both "open," and the slot labels both point to the trunk. If the first state does not meet the gear position condition, the processing strategy label for the first training task is "do not execute," the reason being that the current gear position does not meet the opening condition, the control action is not to issue the trunk opening command, and the target state is to keep the trunk currently closed. If the second state meets all execution conditions, the processing strategy label for the second training task is "execute," the control action is to issue the trunk opening command, and the target state is the trunk open state. The user command, intent label, and slot label remain the same for both training tasks. In the input vehicle state information, except for the gear position state variable included in the minimum state repair set, all other irrelevant variables saved as state context and unchanged relevant variables remain consistent. The processing strategy label, control action, and target state change accordingly depending on the executable status of the target vehicle function.

[0052] The above scheme automatically generates initial paired training tasks with the same semantic labels but different processing strategy labels by placing the same user instruction on both sides of the function execution boundary. This makes the model training no longer dependent on isolated semantic samples, but directly learns the processing strategy changes that occur when the same semantics cross the function boundary as the vehicle state changes.

[0053] S400: Train the basic model of the automotive domain with a vehicle function adapter using the initial pairwise training task, so that the semantic representation of the same user instruction is consistent under the boundary state pair, and the output processing strategy matches the corresponding state transition rule, thus obtaining the semantic understanding model.

[0054] In the embodiments of this application, the automotive domain foundation model includes a shared semantic encoding layer, a semantic projection layer, a decision projection layer, and multiple vehicle function adapters, each vehicle function adapter corresponding to at least one vehicle function. The semantic understanding model is the trained automotive domain foundation model, retaining the shared semantic encoding layer, semantic projection layer, decision projection layer, and vehicle function adapters of the automotive domain foundation model.

[0055] In embodiments of this application, training an automotive domain foundation model equipped with a vehicle function adapter using initial pairwise training tasks includes: encoding each training task in the initial pairwise training tasks through a shared semantic encoding layer; generating semantic representations and decision representations for each training task through a semantic projection layer and a decision projection layer, respectively; determining semantic consistency loss based on the differences between semantic representations, determining decision separation loss based on the differences between decision representations, and determining state transition consistency loss based on the differences between the processing strategy output by the automotive domain foundation model and the corresponding state transition rules; and updating the automotive domain foundation model based on the semantic consistency loss, decision separation loss, and state transition consistency loss to match the output processing strategy with the corresponding state transition rules, and to reduce the differences between semantic representations and increase the differences between decision representations.

[0056] When training the basic model in the automotive domain, the corresponding vehicle function adapter is selected for parameter update based on the target vehicle function corresponding to the initial pairwise training task, and the parameters of vehicle function adapters that are unrelated to the target vehicle function are frozen.

[0057] In a preferred embodiment of this application, the automotive domain foundation model employs a pre-trained language model based on a Transformer encoder. Task instructions, user instructions, vehicle model version, vehicle state fields and their values, and functional boundary identifiers are converted into labeled sequences and embedded into vectors. These vectors are then input into a shared semantic coding layer to obtain a context-hidden representation containing semantic information of user instructions and vehicle state information.

[0058] The semantic projection layer extracts semantic representations corresponding to intents and slots from the context hiding representation; the decision projection layer fuses the context hiding representation, vehicle state representation, and functional boundary representation to generate decision representations for determining processing strategies, control actions, and target states.

[0059] The semantic projection layer maps semantic representations to intent prediction results and slot prediction results, respectively; the decision projection layer maps decision representations to processing strategy prediction results, control action prediction results, and target state prediction results, respectively. The intent prediction results, slot prediction results, and processing strategy prediction results include the prediction probabilities corresponding to each candidate label, and the prediction probabilities are used to calculate the corresponding label prediction loss.

[0060] The vehicle function adapter employs a bottleneck-type residual bypass structure, inserted after each Transformer sublayer of the shared semantic coding layer. It sequentially performs dimensionality reduction, nonlinear transformation, and dimensionality increase on the hidden representation output by the Transformer layer, then adds the dimensionality increase result to the residual of the original hidden representation, achieving low-overhead injection of function-specific features. The vehicle function adapter sequentially performs dimensionality reduction, nonlinear transformation, and dimensionality increase on the hidden representation output by the corresponding Transformer layer, adding the dimensionality increase result to the hidden representation before inputting it into the next Transformer layer. The corresponding vehicle function adapter is selected based on the target vehicle function identifier in the initial pairwise training task, ensuring that the state boundary features and processing strategy features corresponding to different vehicle functions are represented by the parameters of the corresponding vehicle function adapter.

[0061] For the first and second training tasks in the same initial pairwise training task, semantic representations are obtained through the semantic projection layer, respectively. , The decision representation and the decision projection layer are obtained respectively. , In this context, the subscript sem represents semantics, the subscript dec represents decision, and the superscripts 1 and 2 represent the first and second training tasks, respectively, and do not represent exponentiation.

[0062] Before calculating the distances between semantic representations and between decision representations, L2 norm normalization is performed on both semantic and decision representations, and the normalized results are denoted as... , , and This is to reduce the impact of changes in the magnitude of the representation vector on the distance calculation results.

[0063] In one specific embodiment, the semantic consistency loss is calculated as follows:

[0064] in, This represents the semantic consistency loss. This represents the semantic label prediction loss determined based on the intent labels, slot labels, and model prediction results from the first and second training tasks. It represents the squared Euclidean distance between two semantic representations.

[0065] The semantic label prediction loss is determined using cross-entropy loss. Since the first and second training tasks use the same user instructions, intent labels, and slot labels, reducing the squared Euclidean distance enables the model to maintain consistent semantic representation under different boundary states.

[0066] It should be noted that since the semantic representation has been normalized by the L2 norm, the squared Euclidean distance between two semantic representations is equal to twice the cosine distance, which can reflect the directional difference between the two semantic representations without being affected by the magnitude of the representation vector.

[0067] In one specific embodiment, the decision separation loss is calculated as follows:

[0068] in, Indicates the loss due to decision separation. This represents the processing strategy label prediction loss determined based on the processing strategy labels of the first and second training tasks and the model prediction results; m represents the preset separation interval between the decision representations. This represents the Euclidean distance between two decision representations.

[0069] Since the first training task and the second training task correspond to different processing strategy labels, when the distance between the two decision representations is less than the preset separation interval m, the decision separation loss generates a separation penalty to increase the distance between the two decision representations, so that the automotive domain basic model with vehicle function adapter can understand the model and distinguish the changes in processing strategy caused by boundary state differences.

[0070] When calculating the state transition consistency loss, based on the vehicle states and functional boundary markers in the training task, the corresponding expected processing strategy, permitted control actions, and target states are determined from the vehicle functional state transition diagram. The processing strategy classification loss is determined based on the processing strategy probabilities output by the automotive domain baseline model and the expected processing strategy. An inconsistency penalty is added when the model's output control action is not a permitted control action or the output target state cannot be reached from the current vehicle state via a control action. The sum of the processing strategy classification loss and the inconsistency penalty is determined as the state transition consistency loss. .

[0071] After obtaining the semantic consistency loss, decision separation loss, and state transition consistency loss, the three losses are weighted and combined to obtain the joint training loss:

[0072] Where L represents the joint training loss; This represents the loss of semantic consistency. Indicates the loss due to decision separation; This represents the consistency loss during state transitions; These represent the loss weights corresponding to semantic consistency loss, decision separation loss, and state transition consistency loss, respectively.

[0073] In one optional implementation, to facilitate model convergence and ensure training stability, the loss weights and training hyperparameters can be preset to fixed reference values ​​based on experience. For example, in a preferred embodiment, semantic consistency loss weights are set. =1.0, decision separation loss weight =0.5, weight of state transition consistency loss =2.0. Wherein, The value is relatively high to strengthen the priority of vehicle engineering constraints and ensure that the model output conforms to the vehicle's physical logic; A moderate value is used to ensure the effective separation of decision boundaries.

[0074] The preset separation interval m is set to 0.3. Since both the semantic representation and the decision representation have undergone L2 normalization and their vector magnitude is 1, the Euclidean distance between them ranges from [0,2]. Setting m=0.3 can provide a reasonable penalty threshold for decision separation and ensure that the decision feature distance between positive and negative sample pairs has sufficient discriminative margin.

[0075] In a preferred embodiment of this application, the basic model for the automotive domain is trained using the AdamW optimizer, with an initial learning rate set to... In conjunction with a warm-up training strategy, the number of warm-up steps was set to 1000, followed by a linear decay strategy to adjust the learning rate. For the phased training strategy, the training rounds were set to 10 rounds for the semantic stabilization phase, 20 rounds for the boundary decision phase, and 10 rounds for the joint training phase.

[0076] In a preferred embodiment of this application, in order to reduce the mutual interference between semantic consistency training and decision separation training, the basic model in the automotive domain is trained in stages.

[0077] During the semantic stabilization phase, the parameters of all vehicle function adapters are frozen, and gradient updates are not performed. The parameters of the decision projection layer are also frozen. At this stage, decision separation loss and state transition consistency loss are not calculated. Instead, based solely on the differences between the intent labels, slot labels, and semantic representations of the first and second training tasks, the parameters of the shared semantic encoding layer and semantic projection layer are optimized using semantic consistency loss, ensuring that the same user instruction forms a stable semantic representation under boundary state pairs.

[0078] During the boundary decision training phase, the vehicle function adapter corresponding to the target vehicle function is invoked, and the parameters of the shared semantic encoding layer and semantic projection layer are frozen. Based on the differences between the processing strategy labels, decision representations, and corresponding state transition rules of the two training tasks, the decision projection layer and the vehicle function adapter corresponding to the target vehicle function are updated, and the parameters of other vehicle function adapters are frozen.

[0079] During the joint training phase, the parameters of vehicle function adapters that are unrelated to the target vehicle's functions are frozen, while the parameters of the shared semantic coding layer are kept frozen. Only the semantic projection layer, decision projection layer, and corresponding vehicle function adapters are unfrozen. Based on the joint training loss, the semantic projection layer, decision projection layer, and vehicle function adapters corresponding to the target vehicle's functions are simultaneously optimized.

[0080] During joint training, the corresponding loss weights are adjusted based on the magnitude of change of the three losses and the validation set results. When the value of any weighted loss in relation to the joint training loss is consistently higher than that of the other weighted losses, the loss weight corresponding to that loss is decreased; when the corresponding result of any loss on the validation set does not meet the training requirements, the loss weight corresponding to that loss is increased. The adjusted loss weights are then used to calculate the joint training loss in the next training phase.

[0081] The above scheme constrains the semantic and decision representations of boundary state pairs through semantic projection layers and decision projection layers, respectively, and introduces vehicle function state transition rules to constrain the model output. This enables the basic model in the automotive domain to form mutually separate boundary decisions while maintaining the semantic stability of the same user instruction, thereby solving the problem that vehicle state changes are easily misinterpreted as semantic changes by the model.

[0082] S500: Compare the processing strategy output by the semantic understanding model with the processing strategy determined based on the vehicle function state transition diagram, locate the erroneous function boundary based on the comparison result, generate error correction pair training tasks based on the erroneous function boundary, and update the vehicle function adapter corresponding to the erroneous function boundary.

[0083] In the embodiments of this application, locating the erroneous functional boundary based on the comparison results includes: obtaining the current vehicle state and the output processing strategy when the semantic understanding model processes user instructions; determining the corresponding state transition rule from the vehicle functional state transition diagram based on the current vehicle state, and determining the expected processing strategy based on whether the current vehicle state meets the execution conditions of the state transition rule; when the output processing strategy is inconsistent with the expected processing strategy, the functional boundary corresponding to the state transition rule is determined as the erroneous functional boundary.

[0084] In embodiments of this application, generating error correction pairwise training tasks based on error function boundaries and updating the corresponding vehicle function adapters includes: combining user instructions with boundary state pairs that cross error function boundaries to generate error correction pairwise training tasks; keeping the parameters in the automotive domain base model unchanged except for the vehicle function adapters corresponding to the error function boundaries, and updating the corresponding vehicle function adapters using the error correction pairwise training tasks; validating the semantic understanding model configured with the updated vehicle function adapters using a target boundary validation set, a target function preservation set, and a non-target function protection set; deploying the updated vehicle function adapters when the validation passes, and retaining the vehicle function adapters before the update when the validation fails.

[0085] In a preferred embodiment of this application, the semantic understanding model is deployed on the vehicle-side controller. During runtime, the controller receives user commands and synchronously collects the current vehicle bus signals as the current vehicle state. After inputting these signals into the model, it parses out the function identifier, intent, slot, processing strategy, control action, and target state. Based on the function identifier and vehicle model version, it retrieves the vehicle function state transition diagram and matches the current state with the rule execution conditions. If the conditions are met, the expected strategy is to execute, and the control action and target state are based on the rule record; if the conditions are not met, the expected strategy is not to execute, control actions are prohibited, and the current state is maintained.

[0086] The vehicle function state transition diagram is retrieved based on the function identifier and vehicle model version. The corresponding state transition rule is then called, and it is verified whether the current vehicle state meets the execution conditions of the rule. If it does, the expected processing strategy is determined to be executed, and the expected control action and expected target state are based on the record of the rule. If it does not meet the conditions, the expected processing strategy is determined to be not executed, the expected control action is not to issue control commands, and the expected target state is to maintain the current vehicle state.

[0087] Furthermore, the output of the semantic understanding model is compared with the expected processing strategy. If the model misjudges it as executable when the current state does not meet the execution conditions, it is defined as an erroneous execution sample; if the model misjudges it as non-executable when the current state meets the execution conditions, it is defined as an erroneous rejection sample. If the processing strategy is correct, but the control action or target state does not match the expected value, it is classified as an abnormal state transition output sample, directly transferred to the manual review queue, and does not participate in the subsequent error function boundary localization and adapter update process.

[0088] In a preferred embodiment of this application, abnormal samples generated due to unclear user commands, incorrect voice recognition results, missing vehicle status signals, mismatched vehicle models or versions, or inconsistent state transition rule versions are transferred to a manual review queue and are not directly used to update the vehicle function adapter. Samples with incorrect function identifiers, intents, or slot recognition are marked as semantic recognition error samples and are not processed as functional boundary error samples.

[0089] For erroneous execution samples, the unmet state conditions of the current vehicle state are determined based on the execution conditions in the corresponding state transition rules. When the current vehicle state has only one unmet state condition, the condition boundary corresponding to the transition from unmet to met state condition is determined as the erroneous function boundary.

[0090] When the current vehicle state of an erroneous execution sample has multiple unmet state conditions, a minimum state repair set is determined according to S200. If the minimum state repair set contains only one vehicle state variable, the condition boundary corresponding to that vehicle state variable is determined as the erroneous function boundary; if the minimum state repair set contains multiple vehicle state variables, based on the logical relationship between the atomic state conditions in the execution conditions, it is determined whether each vehicle state variable needs to change individually or collectively, and the condition boundary or combination of condition boundaries that transforms the target vehicle function from non-executable to executable is determined as the erroneous function boundary.

[0091] For incorrectly rejected samples, the current vehicle state at the time of the error is taken as the second state. For each atomic state condition in the corresponding state transition rule, the vehicle state variable corresponding to the atomic state condition is changed within the legal state domain, changing the atomic state condition from satisfied to unsatisfied, while keeping the vehicle state variables corresponding to other atomic state conditions unchanged, resulting in candidate first states corresponding to each atomic state condition. The same user instruction is combined with the candidate first and second states respectively and input into the semantic understanding model; when the processing strategy output by the model for the candidate first and second states does not match the non-execution strategy and the execution strategy respectively, the condition boundary crossed by the candidate first state is determined as an erroneous functional boundary. When multiple condition boundaries do not meet the above matching requirements, each condition boundary is determined as an erroneous functional boundary.

[0092] After determining the erroneous function boundary, boundary state pairs that cross the boundary are generated. For erroneous execution samples, the current vehicle state at the time of the error is taken as the first state, and a second state that satisfies the execution conditions is determined based on the minimum state repair set. For erroneous rejection samples, the candidate first state is taken as the first state, and the current vehicle state at the time of the error is taken as the second state. Except for the vehicle state variables involved in the erroneous function boundary, the values ​​of all other vehicle state variables in the first and second states remain consistent.

[0093] Furthermore, the user commands that triggered the erroneous decisions are extracted and concatenated with the first and second states respectively to construct the first and second error correction training tasks. Both tasks share intent labels and slot labels, but are configured with differentiated processing strategy labels, processing reasons, control actions, and target states based on the vehicle function state transition diagram. These two tasks are then linked as paired error correction training tasks, and the corresponding vehicle model version, erroneous function boundary identifier, and state transition rule version are labeled.

[0094] Based on the target vehicle function corresponding to the error function boundary, the vehicle function adapter to be updated is selected. During the update phase, the parameters of the shared semantic coding layer, semantic projection layer, decision projection layer, and other vehicle function adapters are locked and dedicated to updating the selected adapter parameters through backpropagation in the error correction pairwise training task. After the update is complete, the optimized adapter is attached to the transformer layer corresponding to the target vehicle function in the shared semantic coding layer. During inference, user commands and the current vehicle state are fed into the shared semantic coding layer, which is then parsed sequentially by the semantic projection layer and the decision projection layer, outputting the fine-tuned processing strategy, control actions, and target state.

[0095] After locating the erroneous functional boundary, a target boundary validation set is constructed in the bench simulation environment on the training server based on the state transition rules corresponding to that boundary. This validation set covers the boundary state pairs on both sides of the boundary and their associated user commands. The core purpose is to verify whether the updated adapter can accurately capture the state differences on both sides of the boundary and output correct and distinct processing strategies.

[0096] The target function maintains a set of historical verification tasks extracted offline from the target vehicle function library, covering all verified function boundaries and corresponding tasks except for the current error boundary. This is used for regression testing to ensure that local hot updates do not impair the existing semantic understanding and decision-making capabilities of the function.

[0097] The non-target function protection set is selected from the cross-function verification task library and consists of other functional tasks that are distinct from the target vehicle function. In particular, if a single vehicle function adapter serves multiple functions simultaneously, the protection set must cover other functional tasks that share the same adapter as the target function, thereby strictly isolating the update impact domain and preventing cross-function regression defects.

[0098] Furthermore, regression verification was performed on the semantic understanding model equipped with the updated adapter based on the three types of datasets mentioned above. The verification logic is as follows: If the processing strategies on both sides of the error function boundary are consistent with the state transition rules, and on the functionally executable side, the control chain output by the model (including control actions and target states) is completely consistent with the rule records; at the same time, the inference results of the target function holding set and the non-target function protection set have not deteriorated compared with the previous version, and the false alarm rate has not increased, then the verification is deemed successful, and the updated adapter is allowed to be deployed. Conversely, if any verification item fails to meet the standard, the verification is deemed unsuccessful, the adapter version before the update is immediately rolled back, and the relevant error samples and corresponding state transition rules are packaged and submitted to the manual review queue.

[0099] Taking the aforementioned trunk opening function as an example, when the current gear does not meet the opening conditions, but the semantic understanding model still outputs an execution strategy for the user command "open the trunk," this sample is identified as an erroneous execution sample, and the error is located at the gear condition boundary of the trunk opening function. The vehicle state at the time of the error is taken as the first state, and only the gear state variable is changed to obtain the second state that meets the opening conditions. Training tasks are then generated for the first state (not executing) and the second state (executing), respectively. The vehicle function adapter corresponding to the trunk opening function is updated using the error correction pairwise training task. During the update process, the vehicle function adapters corresponding to the air conditioning, window, and seat controls remain unchanged.

[0100] The above scheme constructs a fully adaptive optimization chain from error perception and boundary attribution to local fine-tuning by back-locating the error handling strategy to the specific functional boundary in the vehicle function state transition diagram, generating error correction pairwise training tasks around this boundary, and updating only the corresponding vehicle function adapter. Through parameter isolation design that updates only the corresponding adapter, coupled with dual verification of the target function hold-out set and the non-target function protection set, the scheme effectively prevents unexpected disturbances to verified functions during fine-tuning, ensuring the complete stability of the semantic understanding accuracy and decision compliance of existing functions. While correcting specific decision boundaries, it does not cause performance degradation to the existing processing capabilities of other vehicle functions.

[0101] In summary, this embodiment determines the minimum state repair set that transforms the target vehicle function from non-executable to executable through a vehicle function state transition graph. It constructs boundary state pairs where only necessary vehicle state variables change, and places the same user instruction on both sides of the function execution boundary. Initial paired training tasks with the same semantic label but different processing strategy labels are automatically generated. During training, semantic representation, decision representation, and processing strategy are constrained to ensure the model maintains semantic consistency with user instructions. The model outputs corresponding processing strategies based on whether the vehicle state crosses the function execution boundary. When the model output is inconsistent with the state transition rules, the corresponding erroneous function boundary is located, and the corresponding vehicle function adapter is updated using only the error correction paired training task. This reduces the number of model parameters involved in fine-tuning and minimizes the impact of adapter updates on the processing results of other vehicle functions.

[0102] The present invention also provides a system for implementing a semantic understanding model training and adaptive fine-tuning method for the automotive vertical domain. The system includes a control module, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the computer program to implement the semantic understanding model training and adaptive fine-tuning method for the automotive vertical domain.

[0103] This invention provides a storage medium storing a program that, when executed by a processor, implements a method for training and adaptive fine-tuning a semantic understanding model for the automotive vertical domain.

[0104] This invention provides a processor for running a program, wherein the program executes a method for training and adaptive fine-tuning a semantic understanding model for the automotive vertical domain.

[0105] This invention provides a device including a processor, a memory, and a program stored in the memory and executable on the processor. When the processor executes the program, it implements a method for training and adaptive fine-tuning a semantic understanding model for the automotive vertical domain. The device described herein can be a server, PC, tablet, mobile phone, etc.

[0106] This application also provides a computer program product that, when executed on a data processing device, is suitable for performing a semantic understanding model training and adaptive fine-tuning method for the automotive vertical industry.

[0107] Those skilled in the art will understand that embodiments of this application can provide methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0108] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0109] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0110] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0111] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0112] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0113] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0114] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0115] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for training and adaptive fine-tuning a semantic understanding model for the automotive vertical domain, characterized in that, include: A vehicle function state transition diagram is constructed based on the acquired multi-source data in the automotive field; Based on the vehicle function state transition diagram, determine the minimum state repair set that transforms the target vehicle function from non-executable to executable, and form a boundary state pair by combining the vehicle states before and after applying the minimum state repair set. The same user instruction is combined with the boundary state pair to generate an initial pair training task with the same semantic label but different processing strategy labels. The initial pairwise training task is used to train the basic model of the automotive domain with a vehicle function adapter, so that the semantic representation of the same user instruction is consistent under the boundary state pair, and the output processing strategy matches the corresponding state transition rule, thus obtaining a semantic understanding model. The processing strategy output by the semantic understanding model is compared with the processing strategy determined according to the vehicle function state transition diagram. The erroneous function boundary is located based on the comparison result. An error correction pair training task is generated based on the erroneous function boundary, and the vehicle function adapter corresponding to the erroneous function boundary is updated.

2. The method for training and adaptive fine-tuning a semantic understanding model for the automotive vertical domain according to claim 1, characterized in that, The process of constructing a vehicle functional state transition diagram based on acquired multi-source data from the automotive field includes: Vehicle functions, vehicle state variables, execution conditions, control actions, and target states are extracted from multi-source data in the automotive field. Vehicle state nodes are constructed by combining the states of the vehicle state variables; Based on the execution conditions, functional transition edges are established between the vehicle state nodes, and the corresponding vehicle functions, control actions, and target states are associated with the functional transition edges.

3. The method for training and adaptive fine-tuning a semantic understanding model for the automotive vertical domain according to claim 1, characterized in that, The determination of the minimum state repair set based on the vehicle function state transition diagram to convert the target vehicle function from non-executable to executable includes: Obtain the current vehicle status and the execution conditions of the target vehicle's functions; Determine the state conditions that are not met in the current vehicle state from the execution conditions; Based on the vehicle state variables corresponding to the state conditions, candidate repair combinations are searched to obtain multiple candidate state repair sets; From the candidate state repair set that can convert the target vehicle function into an executable state that satisfies the vehicle state constraints, the candidate state repair set containing the fewest vehicle state variables is determined as the minimum state repair set.

4. The method for training and adaptive fine-tuning a semantic understanding model for the automotive vertical domain according to claim 1, characterized in that, The boundary state pair includes a first state and a second state; The first state does not meet the execution conditions of the target vehicle function, the second state meets the execution conditions, and when the transition from the first state to the second state occurs, only the vehicle state variables within the minimum state repair set change.

5. The method for training and adaptive fine-tuning a semantic understanding model for the automotive vertical domain according to claim 4, characterized in that, The initial pairwise training task, which generates semantic labels with the same meaning but different processing strategy labels, includes: The same user instruction is combined with the first state and the second state respectively to generate the first training task and the second training task. Set the same intent label and slot label for the first training task and the second training task; Based on the executable status of the target vehicle function in the first state and the second state, the processing strategy labels for the first training task and the second training task are determined respectively, and the processing strategy labels for the first training task and the second training task are different. The first training task and the second training task are associated as the initial pairwise training task.

6. The method for training and adaptive fine-tuning a semantic understanding model for the automotive vertical domain according to claim 1, characterized in that, The automotive domain basic model includes a shared semantic encoding layer, a semantic projection layer, a decision projection layer, and multiple vehicle function adapters, with each vehicle function adapter corresponding to at least one vehicle function.

7. The method for training and adaptive fine-tuning a semantic understanding model for the automotive vertical domain according to claim 6, characterized in that, The training of the automotive domain basic model with a vehicle function adapter using the initial pairwise training task includes: Each training task in the initial paired training task is encoded using the shared semantic coding layer. The semantic projection layer and the decision projection layer generate semantic representations and decision representations for each training task, respectively. The semantic consistency loss is determined based on the differences between the semantic representations, the decision separation loss is determined based on the differences between the decision representations, and the state transition consistency loss is determined based on the differences between the processing strategy output by the automotive domain basic model and the corresponding state transition rule. The automotive domain basic model is updated based on the semantic consistency loss, decision separation loss, and state transition consistency loss to match the output processing strategy with the corresponding state transition rule, and to reduce the differences between the semantic representations and increase the differences between the decision representations.

8. The method for training and adaptive fine-tuning a semantic understanding model for the automotive vertical domain according to claim 7, characterized in that, When training the automotive domain basic model, the corresponding vehicle function adapter is selected for parameter update based on the target vehicle function corresponding to the initial pairwise training task, and the parameters of vehicle function adapters that are unrelated to the target vehicle function are frozen.

9. The method for training and adaptive fine-tuning a semantic understanding model for the automotive vertical domain according to claim 1, characterized in that, The step of locating the boundary of the erroneous function based on the comparison results includes: The current vehicle state and output processing strategy are obtained when the semantic understanding model processes user commands; Based on the current vehicle state, the corresponding state transition rule is determined from the vehicle function state transition diagram, and the expected processing strategy is determined based on whether the current vehicle state meets the execution conditions of the state transition rule. When the output processing strategy is inconsistent with the expected processing strategy, the functional boundary corresponding to the state transition rule is determined as an erroneous functional boundary.

10. The method for training and adaptive fine-tuning a semantic understanding model for the automotive vertical domain according to claim 9, characterized in that, The step of generating error correction pairwise training tasks based on the error function boundary and updating the corresponding vehicle function adapter includes: The user instructions are combined with boundary state pairs that cross the boundary of the error function to generate error correction pairwise training tasks. Keeping all parameters in the automotive domain basic model unchanged except for the vehicle function adapter corresponding to the error function boundary, the corresponding vehicle function adapter is updated using the error correction pairwise training task. The semantic understanding model configured with the updated vehicle function adapter is validated using the target boundary validation set, the target function preservation set, and the non-target function protection set. If the validation passes, the updated vehicle function adapter is deployed; if the validation fails, the vehicle function adapter before the update is retained.

Images