CN-121978978-A - Multichannel intelligent fragrance control method and system based on large language model

CN121978978ACN 121978978 ACN121978978 ACN 121978978ACN-121978978-A

Abstract

The invention discloses a multichannel intelligent fragrance control method and a multichannel intelligent fragrance control system based on a large language model, which relate to the technical field of intelligent voice fragrance control and have the core that the attention head pruning proportion of the model is dynamically adjusted according to an environment disturbance index; when the environment is disturbed, the pruning proportion is reduced according to the instruction space association degree so as to focus on the core semantics, and invalid fine control instructions are avoided being generated. In addition, the method can also process continuous instructions of users, conflict of multi-user instructions and linkage with entertainment content, and further adaptively adjust model configuration and generate a fragrance control strategy of space decoupling or content fusion by calculating equivalent characteristics of semantic deviation degree and instruction scene coordination degree. According to the invention, intelligent distribution of model computing resources is realized in a dynamic environment, the accuracy of fragrance control instructions is ensured, and the reliability of personal immersive olfactory experience is remarkably improved.

Inventors

Fang Jiaan
LI XINYUAN
LI JIANMIN

Assignees

杭州明炬文化创意有限公司

Dates

Publication Date: 20260505
Application Date: 20260209

Claims (10)

1. The multichannel intelligent fragrance control method based on the large language model is characterized by comprising the following steps of: Acquiring first original voice of a first user, and processing according to a first training large language model through a first attention head pruning proportion to obtain a first language instruction; Acquiring continuous airflow speed data of a first user, and calculating a root mean square value of the continuous airflow speed data through a sliding time window according to the continuous airflow speed data as an environment disturbance index; If the environmental disturbance index is smaller than or equal to a preset threshold, calculating to obtain a semantic dispersion index according to the first language instruction, increasing the first attention head pruning proportion according to the semantic dispersion index, obtaining a second attention head pruning proportion, obtaining a first language instruction according to the first attention head pruning proportion, marking the first language instruction as a second language instruction, outputting a first fragrance control instruction according to the second language instruction, and executing the first fragrance control instruction; If the environmental disturbance index is greater than a preset threshold, calculating to obtain an instruction space association degree according to the first language instruction, and accordingly reducing the first attention head pruning proportion to obtain a third attention head pruning proportion; after the first time window, if the environmental disturbance index is still greater than a preset threshold value, the first language instruction is obtained again according to the pruning proportion of the third attention head, and is marked as a third language instruction, and a second fragrance control instruction is output and executed according to the third language instruction; If the environmental disturbance index is smaller than or equal to the preset threshold, obtaining the extreme difference of the environmental disturbance index in the first time window, increasing the third attention head pruning proportion with the preset threshold according to the extreme difference, obtaining the fourth attention head pruning proportion, obtaining the first language instruction according to the fourth attention head pruning proportion, recording the first language instruction as the fourth language instruction, outputting the third fragrance control instruction according to the fourth language instruction, and executing the third fragrance control instruction.
2. The method for controlling a multichannel intelligent fragrance according to claim 1, wherein after obtaining the second language instruction or the third language instruction or the fourth language instruction, if the second original voice of the first user is received in the second time window and the fifth language instruction is obtained according to the second original voice, the following steps are executed: if the second language instruction is obtained, the current attention head pruning proportion is the second attention head pruning proportion; if the third language instruction is obtained, the current attention head pruning proportion is the third attention head pruning proportion; If the fourth language instruction is obtained, the current attention head pruning proportion is a fourth attention head pruning proportion; calculating a second large language model which is input by the current first language instruction and the fifth language instruction and is pre-trained to obtain a first olfactory intention vector and a second olfactory intention vector; calculating Euclidean distance between the first olfactory intention vector and the second olfactory intention vector, and recording the Euclidean distance as semantic deviation degree between the first instructions; Outputting a first fragrance control instruction and executing if the semantic deviation degree among the first instructions is smaller than or equal to a first deviation threshold value; If the semantic deviation degree between the first instructions is larger than a first deviation threshold value, increasing the current attention head pruning proportion according to the difference value between the semantic deviation degree between the first instructions and the first deviation threshold value to obtain a fifth attention head pruning proportion, reprocessing the fifth language instruction according to a pre-trained first large language model through the fifth attention head pruning proportion to obtain a fifth language instruction, outputting a fourth fragrance control instruction according to the fifth language instruction, and executing the fourth fragrance control instruction.
3. The method for controlling multi-channel intelligent fragrance according to claim 2, wherein after obtaining the second language instruction or the third language instruction or the fourth language instruction, if the third original voice of the second user is received in the third time window and the sixth language instruction is obtained according to the third original voice, the following steps are performed: Calculating the semantic deviation degree between the second instruction of the first language instruction and the sixth language instruction; Outputting a first fragrance control instruction and executing if the semantic deviation degree between the second instructions is smaller than or equal to a preset multiuser conflict threshold value; if the semantic deviation degree between the second instructions is larger than a preset multiuser conflict threshold value, multiuser space decoupling processing is executed; the multi-user spatial decoupling process is as follows: Acquiring the distance from the preset position to the position of the first user and the distance from the preset position to the position of the second user, comparing the distances of the preset position and the second user, acquiring a larger value, and recording the larger value as the distance between the positions; Based on the distance between the separated positions and the semantic deviation degree between the second instructions, increasing the current attention head pruning proportion to obtain a sixth attention head pruning proportion, and obtaining a first language instruction and a sixth language instruction again according to the sixth attention head pruning proportion, and respectively marking the first language instruction and the sixth language instruction as a seventh language instruction and an eighth language instruction; According to the distance between the positions of the first user and the second user and according to the topological rule of the position of the release space of the preset fragrance, a first preset release channel group is allocated for the first user, and a second preset release channel group is allocated for the second user; generating a first sub-control strategy for controlling the first preset release channel group according to the seventh language instruction; Generating a second sub-control strategy for controlling a second preset release channel group according to the eighth language instruction; and merging the first sub-control strategy with the second sub-control strategy, generating a fifth fragrance control instruction and executing the fifth fragrance control instruction.
4. The method for controlling multichannel intelligent fragrance according to claim 3, wherein the specific obtaining process of the sixth attention head pruning proportion is as follows: Acquiring the current pruning proportion of the attention head, the semantic deviation degree among the first instructions, a multiuser conflict threshold value and a preset gain coefficient; The specific calculation formula of the pruning proportion of the sixth attention head is as follows: , wherein, The sixth attention head pruning proportion is shown, Indicating the current proportion of the pruning of the attention head, Representing the degree of semantic deviation between the second instructions, For the multi-user collision threshold value, Representing a preset gain factor.
5. The method for controlling the multichannel intelligent fragrance based on the large language model of claim 2, wherein the step of obtaining the first language instruction further comprises the steps of synchronously obtaining the digital content metadata of the current interaction of the first user, wherein the digital content metadata at least comprises a content scene tag and a language scene tag of the current interaction; Respectively inputting the first language instruction and the digital content metadata into a pre-trained third large language model to obtain corresponding instruction feature vectors and scene feature vectors, and calculating cosine similarity between the instruction feature vectors and the scene feature vectors as the instruction scene synergy; if the instruction scene coordination degree is greater than or equal to the coordination degree threshold value, judging that the instruction scene coordination degree is in a coordination state, and executing according to a first fragrance control instruction; If the instruction scene coordination degree is smaller than the coordination degree threshold value, judging that the instruction scene coordination degree is in a conflict state, triggering content linkage fusion processing, obtaining a sixth fragrance control instruction and executing the sixth fragrance control instruction.
6. The large language model based multichannel intelligent fragrance control method of claim 5, wherein the specific acquisition process of the sixth fragrance control instruction is: Generating an initial fusion text by the first language instruction and the content scene label according to a preset formatting and splicing rule and a difference value of the degree of collaboration of the instruction scene threshold; And generating a ninth language instruction according to the current attention head pruning proportion by the initial fusion text through the pre-trained first large language model, and outputting a sixth fragrance control instruction according to the ninth language instruction.
7. The large language model based multichannel intelligent fragrance control method of claim 1, wherein the specific calculation of the semantic dispersion index comprises: word segmentation and stop word filtering are carried out on the first language instruction through a preset language filtering tool, and an effective word sequence is obtained; inquiring a preset smell semantic database to obtain an inverse document frequency value of each effective word in the effective word sequence; And calculating the standard deviation of the inverse document frequency value of each effective word in the effective word sequence, and normalizing the standard deviation to obtain the semantic dispersity index.
8. The large language model based multichannel intelligent fragrance control method of claim 1, wherein the calculation of the instruction space association degree comprises the following steps: obtaining and calling a pre-trained interdependence analysis model to carry out dependency syntactic analysis on the first language instruction to obtain syntactic structure data; extracting all dependency relationship types characterizing azimuth, direction or time sequence relation from the syntactic structure data; Counting and extracting the total number of dependency relationship types as the number of the spatial semantic relationships; Processing the first language instruction through a preset word segmentation tool to obtain the total word number of the instruction; dividing the number of the space semantic relations by the total word number of the first language instruction after word segmentation, and taking the obtained quotient as the instruction space association degree.
9. The large language model-based multichannel intelligent fragrance control method of claim 1, wherein the specific acquisition process of the fourth attention head pruning proportion is as follows: according to the difference value between the extreme difference of the environmental disturbance index in the first time window and the preset threshold value, a recovery gain coefficient is calculated, and a specific calculation formula of the recovery gain coefficient is as follows: , wherein, Representing the coefficient of the recovery gain, Being a very poor index of environmental disturbance in the first time window, Representing a preset threshold value; Calculating a fourth attention head pruning proportion according to the recovery gain coefficient, the third attention head pruning proportion and the preset attention head pruning proportion, The specific calculation formula of the fourth attention head pruning proportion is as follows: , wherein, The third attention head pruning proportion is represented, Representing a preset attention head pruning proportion.
10. Multichannel intelligent fragrance control system based on big language model, characterized by comprising: the language instruction acquisition module is used for acquiring first original voice of a first user and processing according to a first large pre-trained language model through a first attention head pruning proportion to obtain a first language instruction; The environment disturbance acquisition module is used for acquiring continuous airflow speed data of the first user, and calculating a root mean square value of the continuous airflow speed data through a sliding time window according to the continuous airflow speed data as an environment disturbance index; The first fragrance control instruction module is used for calculating to obtain a semantic dispersion index according to the first language instruction if the environmental disturbance index is smaller than or equal to a preset threshold value, increasing the first attention head pruning proportion according to the semantic dispersion index, obtaining a second attention head pruning proportion according to the first attention head pruning proportion, obtaining the first language instruction again according to the second attention head pruning proportion, marking the second language instruction, outputting the first fragrance control instruction according to the second language instruction, and executing the first fragrance control instruction; the judging module is used for calculating to obtain the instruction space association degree according to the first language instruction if the environment disturbance index is larger than a preset threshold value, and reducing the first attention head pruning proportion accordingly to obtain a third attention head pruning proportion; The second fragrance control instruction module is used for obtaining the first language instruction again according to the pruning proportion of the third attention head after the first time window passes, and marking the first language instruction as the third language instruction, outputting the second fragrance control instruction according to the third language instruction and executing the second language instruction; And the third fragrance control instruction is used for acquiring the extremely poor environmental disturbance index in the first time window if the environmental disturbance index is smaller than or equal to a preset threshold value, increasing the third attention head pruning proportion with the preset threshold value according to the extremely poor environmental disturbance index, obtaining a fourth attention head pruning proportion, obtaining the first language instruction again according to the fourth attention head pruning proportion, recording the fourth language instruction, outputting the third fragrance control instruction according to the fourth language instruction, and executing the third fragrance control instruction.

Description

Multichannel intelligent fragrance control method and system based on large language model Technical Field The application relates to the technical field of intelligent voice fragrance control, in particular to a multichannel intelligent fragrance control method and system based on a large language model. Background With the deep integration of smart home and immersive entertainment experience, environmental fragrance is controlled through natural language instructions to enhance the feeling of reality of individuals in video and audio, games and other scenes, and the smart home and immersive entertainment experience become an important technical trend. The prior art scheme generally deploys a large language model with fixed computing configuration as an interaction core, which parses all instructions of a user in a constant complexity and resource consumption mode and drives the fragrance equipment to execute. However, the real physical space in which the user is located is a dynamic system, in which environmental disturbances caused by factors such as ventilation equipment, personnel activities, etc. can cause a profound contradiction between a fixed instruction resolution mode and dynamic environmental execution conditions. The deficiencies of the prior art root in the statics of its core processing logic. The problem becomes particularly pronounced when environmental disturbances are exacerbated, resulting in fine spatial or timing control complaints at the physical level, where the model still consumes significant computational resources to resolve deeply and attempt to meet fine space-time details in the instruction. This not only results in a waste of instantaneous computing resources, but more importantly, such over-resolution can produce highly sophisticated control signals that rely on stable environments to achieve. Because the environmental disturbance breaks down the basis of accurate execution, these signals cannot be effectively implemented, and finally, a significant deviation occurs between the system output and the odor effect actually perceived by the user, which seriously impairs the consistency and credibility of the experience. In view of the foregoing, there is a need in the art for a solution that enables the intelligent core of a fragrance control system to be provided with environmental adaptation capabilities. The essential problem of the prior art is that the fragrance control method based on the large language model adopts a fixed calculation strategy decoupled from an environment state, and can not intelligently balance the depth of instruction analysis, the calculation efficiency and the control signal executability under the dynamically-changing environment interference, so that stable and accurate personalized olfactory experience is difficult to continuously provide in a real and changeable personal use scene. Disclosure of Invention In order to solve the technical problems, the technical scheme solves the problems in the background technology by providing a multichannel intelligent fragrance control method and a multichannel intelligent fragrance control system based on a large language model. In a first aspect, the embodiment of the application provides a multichannel intelligent fragrance control method based on a large language model, which comprises the following steps of obtaining first original voice of a first user, and processing according to a pre-trained first large language model through a first attention head pruning proportion to obtain a first language instruction; the method comprises the steps of obtaining continuous air flow speed data of a first user, calculating a root mean square value of the continuous air flow speed data through a sliding time window to serve as an environment disturbance index, calculating according to the environment disturbance index to obtain a semantic dispersion index according to a first language instruction if the environment disturbance index is smaller than or equal to a preset threshold value, increasing a first language instruction according to the first language instruction, obtaining a second language instruction according to the first language instruction, recording the second language instruction, outputting a first fragrance control instruction according to the second language instruction, executing the first fragrance control instruction, calculating according to the first language instruction to obtain an instruction space association degree if the environment disturbance index is larger than the preset threshold value, reducing the first attention instruction according to the first language instruction, obtaining a third language instruction according to the third attention instruction after the environment disturbance index is smaller than or equal to the preset threshold value, obtaining a first language instruction according to the third attention instruction, recording the third language instruction according t