CN-121768390-B - Self-adaptive voice vehicle control method, device, vehicle and storage medium

CN121768390BCN 121768390 BCN121768390 BCN 121768390BCN-121768390-B

Abstract

The invention relates to the technical field of vehicle control and discloses a self-adaptive voice vehicle control method, a device, a vehicle and a storage medium, wherein the method comprises the steps of obtaining first load information, wherein the first load information is used for reflecting the load of multidimensional hardware in the vehicle; the method comprises the steps of matching corresponding current performance grades according to first load information, compressing the reasoning capacity of a large language model according to the current performance grades, wherein the reasoning capacity is used for representing the inputtable information quantity and the outputtable control message quantity of the large language model, and processing voice instructions of a user based on the compressed large language model to obtain control messages for controlling a vehicle. The invention solves the problems of large model reasoning efficiency and hardware resource dynamic balance in the vehicle field.

Inventors

HE XINYUAN

Assignees

重庆长安汽车股份有限公司

Dates

Publication Date: 20260508
Application Date: 20260304

Claims (9)

1. An adaptive voice control method, the method comprising: The method comprises the steps of acquiring first load information, counting inter-process communication calling time of voice conversion semantics when receiving user voice, and taking the inter-process communication calling time of the voice conversion semantics and the load data as the first load information; Matching corresponding current performance grades according to the first load information, and compressing the reasoning capacity of a large language model according to the current performance grades, wherein the reasoning capacity is used for representing the input information quantity and the output control message quantity of the large language model; the method comprises the steps of correspondingly determining the current performance level through the similarity of communication call time consumption between processes in each load scene in the first load information and an experience database, unloading the large language model if the current performance level is an unloading level, judging whether the performance level matched with the preset continuous times is the unloading level if the current performance level is the compression level when the large language model is not loaded, and loading the large language model and limiting the inputtable information quantity and the outputtable control message quantity of the large language model according to the corresponding compression quantity if the performance level matched with the preset continuous times is not the unloading level; and processing the voice command of the user based on the compressed large language model to obtain a control message for controlling the vehicle.
2. The method of claim 1, wherein prior to the acquiring the first load information, the method further comprises: Determining a multi-dimensional hardware experimental sample with the same hardware capacity as in a vehicle; loading pressure test data into the multi-dimensional hardware experimental sample, and defining corresponding performance grades for different load scenes reached by the multi-dimensional hardware experimental sample, wherein the pressure test data is used for applying load to the hardware experimental sample so as to reach various load scenes; executing an interprocess communication task and a large language model reasoning task through the pressurized hardware experiment sample; And creating an experience database based on the performance level corresponding to each load scene, the communication call time consumption between the processes corresponding to each load scene and the reasoning time consumption corresponding to each load scene.
3. The method of claim 1, wherein the processing the voice command of the user based on the compressed large language model to obtain the control message for controlling the vehicle comprises: when the large language model is unloaded, calling a pre-stored instruction-message comparison table; If the target control message corresponding to the voice command is queried through the command-message comparison table, outputting the target control message; And when the large language model is not unloaded, generating an inference prompt word of the voice instruction according to the input information quantity limited by the compression level, and inputting the inference prompt word into the large language model for inference to obtain an inference control message.
4. The method according to claim 1, wherein the method further comprises: counting the current reasoning time when the reasoning is finished, and taking the current reasoning time as second load information; Correspondingly updating the current performance level through the second load information and the similarity of time consumption reasoning under each load scene in the experience database; And adjusting the quantity of the control messages according to the updated current performance level.
5. The method of claim 1, wherein the multidimensional hardware comprises a central processor, a neural network processor, and a memory.
6. An adaptive voice control vehicle device, the device comprising: The load acquisition module is used for acquiring first load information, wherein the first load information is used for reflecting the load of multidimensional hardware in a vehicle, and the first load information comprises the steps of acquiring load data of the multidimensional hardware, counting the inter-process communication calling time of voice conversion semantics when receiving user voice, and taking the inter-process communication calling time of the voice conversion semantics and the load data as the first load information; The performance level evaluation module is used for matching the corresponding current performance level according to the first load information, and compressing the reasoning capacity of the large language model according to the current performance level, wherein the reasoning capacity is used for representing the inputtable information quantity and the outputtable control message quantity of the large language model; the method comprises the steps of correspondingly determining the current performance level through the similarity of communication call time consumption between processes in each load scene in the first load information and an experience database, unloading the large language model if the current performance level is an unloading level, judging whether the performance level matched with the preset continuous times is the unloading level if the current performance level is the compression level when the large language model is not loaded, and loading the large language model and limiting the inputtable information quantity and the outputtable control message quantity of the large language model according to the corresponding compression quantity if the performance level matched with the preset continuous times is not the unloading level; And the decision module is used for reasoning the voice command of the user based on the processed large language model to obtain a control message for controlling the vehicle.
7. A vehicle, characterized by comprising: A memory and a processor in communication with each other, the memory having stored therein computer instructions which, upon execution, cause the processor to perform the method of any of claims 1 to 5.
8. A computer readable storage medium having stored thereon computer instructions for causing a computer to perform the method of any one of claims 1 to 5.
9. A computer program product comprising computer instructions for causing a computer to perform the method of any one of claims 1 to 5.

Description

Self-adaptive voice vehicle control method, device, vehicle and storage medium Technical Field The invention relates to the technical field of vehicle control, in particular to a self-adaptive voice vehicle control method, a device, a vehicle and a storage medium. Background With the wide application of large language models in edge computing scenes, the voice meaning of a user is inferred through the large models on vehicles so as to assist the user in controlling the vehicle, and the method is a popular technology at present. However, heterogeneous computing platforms (such as a central processing unit, a neural network processor, a memory, a graphics processor and the like) of the vehicle are in shortage of resources, and reasoning on the heterogeneous computing platforms with limited resources provides challenges for controlling hardware resources of the vehicle, so that the problem of high model reasoning efficiency and dynamic balance of the hardware resources is needed to be solved for the vehicle field. Disclosure of Invention The invention provides a self-adaptive voice vehicle control method, a device, a vehicle and a storage medium, which are used for solving the problems that the large model reasoning efficiency and hardware resources are difficult to dynamically balance in the vehicle field. The invention provides a self-adaptive voice vehicle control method, which comprises the steps of obtaining first load information, matching corresponding current performance grades according to the first load information, compressing reasoning capacity of a large language model according to the current performance grades, wherein the reasoning capacity is used for representing the inputtable information quantity and the outputtable control message quantity of the large language model, and processing voice instructions of a user based on the compressed large language model to obtain a control message for controlling a vehicle. According to the technical means, the resource and efficiency balance problem of the voice control vehicle at the lower end side of the limited resource environment is solved from the core level by acquiring the first load information of the vehicle multidimensional hardware, matching the corresponding performance level and dynamically compressing the large language model reasoning capability. Compared with the traditional single-processor load judgment scheme, the multi-dimensional hardware load state is covered, so that reasoning delay or resource waste caused by misjudgment of single hardware load is avoided, and the overall hardware bearing capacity of large-model reasoning adaptation is ensured. The intelligent degree of the voice control vehicle is guaranteed when the hardware resources are sufficient, the real-time response requirement is preferentially met when the resources are tense through accurately matching the performance level to adjust the input information quantity and the output control message quantity, the problem that the time effectiveness of the control vehicle is affected by the reasoning task jamming under high load is avoided, the problem of functional failure caused by the unbalance of the resource allocation is solved, and the stability, the instantaneity and the intelligent dynamic unification of the voice control vehicle under different hardware load scenes are realized. In some alternative embodiments, before the first load information is acquired, the method further comprises determining a multidimensional hardware experiment sample with the same hardware capacity as that in the vehicle, loading pressure test data into the multidimensional hardware experiment sample, defining corresponding performance levels for different load scenes reached by the multidimensional hardware experiment sample, applying loads to the hardware experiment sample by the pressure test data to reach various load scenes, executing inter-process communication tasks and large language model reasoning tasks through the pressurized hardware experiment sample, and creating an experience database based on the performance levels corresponding to the load scenes, the corresponding inter-process communication call time consumption under the load scenes and the corresponding reasoning time consumption under the load scenes. According to the technical means, the performance level is defined through pressure test and the experience database is created by constructing the multidimensional hardware experiment sample consistent with the vehicle hardware capacity, so that an accurate and efficient reference basis is provided for subsequent load judgment and reasoning capacity adjustment. The method has the advantages that various load scenes are covered through the pressurization experiments in advance, the corresponding relation between the performance level and the communication time consumption and the reasoning time consumption is established, the additional system overhead caused by monito