Multi-channel intelligent fragrance control method and system based on large language model

Abstract

The invention discloses a multi-channel intelligent fragrance control method and system based on a large language model, and relates to the technical field of intelligent voice fragrance control, the core of the method is to dynamically adjust the attention head pruning proportion of a model according to an environment disturbance index: when the environment is stable, the pruning proportion is increased according to a semantic dispersion index so as to improve the efficiency; when the environment is disturbed, the pruning proportion is reduced according to the instruction space correlation degree so as to focus on core semantics, and invalid fine control instructions are prevented from being generated. In addition, user continuous instructions, multi-user instruction conflicts and entertainment content linkage can be processed, model configuration is further adjusted in a self-adaptive mode by calculating quantitative characteristics such as semantic deviation degree and instruction scene collaboration degree, and a fragrance control strategy of space decoupling or content fusion is generated. According to the method, intelligent distribution of model computing resources in a dynamic environment is realized, the accuracy of a fragrance control instruction is ensured, and the reliability of personal immersive olfactory experience is remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent voice fragrance control technology, and more specifically, this application relates to a multi-channel intelligent fragrance control method and system based on a large language model. Background Technology

[0002] With the deepening integration of smart homes and immersive entertainment experiences, controlling ambient fragrance through natural language commands to enhance the sense of presence in scenarios such as movies and games has become an important technological trend. Existing technical solutions typically deploy a large language model with fixed computing configuration as the core of the interaction. This model parses all user commands with a constant complexity and resource consumption pattern and drives the fragrance device to execute them.

[0003] However, the real physical space in which users reside is a dynamic system, where environmental disturbances caused by factors such as ventilation equipment and human activities can create a profound contradiction between fixed instruction parsing patterns and dynamic environmental execution conditions.

[0004] The shortcomings of existing technologies stem from the static nature of their core processing logic. This problem becomes particularly pronounced when environmental disturbances intensify, causing the need for fine spatial or temporal control to fail at the physical level: the model still consumes significant computational resources to perform deep analysis and attempt to satisfy the fine spatiotemporal details in the instructions. This not only wastes instantaneous computational resources, but more importantly, this over-analysis generates highly precise control signals that rely on a stable environment. Since environmental disturbances have disrupted the foundation for accurate execution, these signals cannot be effectively implemented, ultimately leading to a significant discrepancy between the system output and the odor effect actually perceived by the user, severely damaging the consistency and credibility of the experience.

[0005] To address the aforementioned issues, there is an urgent need in this field for a solution that enables the intelligent core of a fragrance control system to possess environmental adaptability. The fundamental problem with existing technologies is that their fragrance control methods, based on large language models, employ a fixed computational strategy decoupled from environmental states. This strategy fails to intelligently balance the depth of instruction parsing, computational efficiency, and the executability of control signals under dynamically changing environmental disturbances, making it difficult to continuously provide a stable and accurate personalized olfactory experience in real-world, ever-changing personal usage scenarios. Summary of the Invention

[0006] To address the aforementioned technical problems, a multi-channel intelligent fragrance control method and system based on a large language model is provided. This technical solution solves the problems mentioned in the background section.

[0007] In a first aspect, embodiments of this application provide a multi-channel intelligent fragrance control method based on a large language model, comprising the following steps: acquiring the first original speech of a first user, processing it according to a pre-trained first large language model through a first attention head pruning ratio to obtain a first language instruction; acquiring the continuous airflow velocity data of the first user, and calculating its root mean square value through a sliding time window as an environmental disturbance index; if the environmental disturbance index is less than or equal to a preset threshold, calculating the semantic dispersion index according to the first language instruction, increasing the first attention head pruning ratio accordingly to obtain a second attention head pruning ratio, and re-obtaining the first language instruction, recording it as a second language instruction, and outputting and executing the first fragrance control instruction accordingly; if the environmental disturbance index is less than or equal to a preset threshold, calculating the semantic dispersion index according to the first language instruction, increasing the first attention head pruning ratio accordingly to obtain a second attention head pruning ratio, and re-obtaining the first language instruction, recording it as a second language instruction, and outputting and executing the first fragrance control instruction accordingly; if the environmental disturbance index is less than or equal to a preset threshold, calculating the semantic dispersion index according to the first language instruction, and ... processing it through a first attention head pruning ratio through a first attention head pruning ratio through a first attention head pruning ratio through a first attention head pruning ratio through a second attention head pruning ratio through a second attention head pruning ratio through a second attention head pruning ratio through a second attention head pruning ratio through a second attention head pruning ratio through a second attention head pruning ratio through a second attention head pruning ratio through a second attention head pruning ratio through a second attention head pruning ratio through a second attention head pruning ratio through a second attention head pruning ratio through a second attention head pruning ratio If the disturbance index is greater than a preset threshold, the instruction space correlation degree is calculated based on the first language instruction, and the first attention head pruning ratio is reduced accordingly to obtain the third attention head pruning ratio. After the first time window, if the environmental disturbance index is still greater than the preset threshold, the first language instruction is re-obtained based on the third attention head pruning ratio, recorded as the third language instruction, and the second fragrance control instruction is output and executed accordingly. If the environmental disturbance index is less than or equal to the preset threshold, the range of the environmental disturbance index within the first time window is obtained, and the third attention head pruning ratio is increased based on this and the preset threshold to obtain the fourth attention head pruning ratio. The first language instruction is re-obtained based on this, recorded as the fourth language instruction, and the third fragrance control instruction is output and executed accordingly.

[0008] Secondly, embodiments of this application provide a multi-channel intelligent fragrance control system based on a large language model, comprising: a language instruction acquisition module: used to acquire the first original speech of a first user, and process it according to a pre-trained first large language model through a first attention head pruning ratio to obtain a first language instruction; an environmental disturbance acquisition module: used to acquire the continuous airflow velocity data of the first user, and calculate its root mean square value through a sliding time window as an environmental disturbance index; a first fragrance control instruction module: used to calculate a semantic dispersion index based on the first language instruction if the environmental disturbance index is less than or equal to a preset threshold, and increase the first attention head pruning ratio accordingly to obtain a second attention head pruning ratio, and re-obtain the first language instruction, record it as a second language instruction, and output and execute the first fragrance control instruction accordingly; and a judgment module. The first attention head pruning module is used to calculate the instruction space correlation degree based on the first language instruction if the environmental disturbance index is greater than the preset threshold, and reduce the first attention head pruning ratio accordingly to obtain the third attention head pruning ratio. The second fragrance control instruction module is used to obtain the first language instruction again based on the third attention head pruning ratio after the first time window if the environmental disturbance index is still greater than the preset threshold, record it as the third language instruction, and output and execute the second fragrance control instruction accordingly. The third fragrance control instruction is used to obtain the range of the environmental disturbance index within the first time window if the environmental disturbance index is less than or equal to the preset threshold, and increase the third attention head pruning ratio according to it and the preset threshold to obtain the fourth attention head pruning ratio, and obtain the first language instruction again according to it, record it as the fourth language instruction, and output and execute the third fragrance control instruction accordingly.

[0009] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages:

[0010] 1. By calculating the environmental perturbation index, the dynamic changes in the external physical environment are transformed into a processable quantifiable signal, which serves as the primary basis for decision-making. When the environment is stable, the complexity of instructions is judged based on the semantic dispersion index, and the attention head pruning ratio is increased to focus on understanding with higher computational efficiency. When the environment is perturbed, the attention head pruning ratio is reduced based on the instruction space correlation to avoid the model consuming resources on ineffective details. This mechanism makes the large language model no longer a static "black box," and its computational depth can respond in real time and match the feasibility conditions of the external physical world, realizing the dynamic coupling of model computational resource allocation and physical environment constraints.

[0011] 2. Through the aforementioned dynamic adjustment mechanism, a balance can be intelligently achieved between computational accuracy and computational overhead. More resources are allocated to pursue optimal resolution when the environment is stable and the instructions are complex; when the environment is disturbed or the instructions contain invalid details, the model is proactively simplified, suppressing deep computation on fine but non-executable semantics. This strategy of allocating computing resources on demand effectively avoids unnecessary complex calculations in scenarios where the expected physical effects cannot be achieved, thereby reducing overall latency and energy consumption. This makes it easier to deploy and run on edge devices with limited computing power, such as smart home hubs, achieving computational efficiency optimization on resource-constrained edge devices.

[0012] 3. By making the model's analytical strategy strongly correlated with environmental conditions, it ensures that the fragrance control instructions generated under any environment, including their precise spatial and temporal control requirements, match the physical diffusion capacity represented by the current environmental disturbance index. Whether it's simple atmosphere creation or complex spatial sequence control, the output instructions are executable under the current physical conditions, thus guaranteeing the reliability and consistency of the conversion process from user instructions to the final olfactory experience. Attached Figure Description

[0013] Figure 1 A schematic diagram illustrating the steps of the multi-channel intelligent fragrance control method based on a large language model provided in this application embodiment;

[0014] Figure 2 A schematic diagram of the logic flow of the multi-channel intelligent fragrance control method based on a large language model provided in the embodiments of this application;

[0015] Figure 3 A schematic diagram of the structure of a multi-channel intelligent fragrance control system based on a large language model provided in an embodiment of this application. Detailed Implementation

[0016] This application's embodiments address the technical problem in the prior art where, under dynamically changing environmental interference, the robustness of intelligently balancing the depth and computational efficiency of instruction parsing based on a large language model is insufficient. This is achieved through a multi-channel intelligent fragrance control method and system based on a large language model.

[0017] The fundamental contradiction in existing fragrance control solutions based on large language models lies in the disconnect between the model's static computational configuration and the dynamic physical environment. Environmental disturbances directly determine the physical feasibility of any fine-grained spatial control signal, and the depth of instruction parsing by the large language model is directly proportional to the computational resources invested. A model with a fixed configuration cannot perceive the environment, and therefore, in a disturbed environment, it will still consume resources to deeply parse the spatial details in the instructions, generating control signals that cannot be executed, resulting in wasted resources and a broken experience; in a stable environment, it may fail to fully utilize ideal conditions to achieve the best experience due to insufficient parsing.

[0018] Therefore, the core logic of this scheme is to establish a closed-loop link from physical environment perception directly to model computational resource allocation. First, the environment needs to be quantitatively perceived, converting continuous airflow velocity into an index representing the intensity of disturbance, which serves as the primary basis for judging the quality of physical execution conditions. Once this scheme determines that the environment is stable and has the foundation for fine-grained control, the next step is to evaluate the value density of the user commands themselves. This is achieved by analyzing the commonality of keywords in the commands and calculating a semantic dispersion index. If the command describes common, concentrated concepts, excessively complex parsing is unnecessary; the attention pruning ratio should be increased to complete the parsing with higher computational efficiency. If the command involves rare, dispersed concepts, it means that deeper semantic association and disambiguation are needed; in this case, the pruning ratio should be reduced, and more computational resources should be invested to ensure depth of understanding. This path ensures that, when conditions permit, the allocation of computational resources matches the complexity of the commands.

[0019] Conversely, when this scheme detects severe environmental disturbances and the lack of a physical basis for fine-grained control, the logic shifts to avoid ineffective computation. At this point, it needs to identify components in the instruction that have explicit requirements for space, direction, or timing, and calculate the spatial correlation of the instruction. A higher correlation indicates that the instruction contains more fine-grained spatial requirements that are destined to fail under the current disturbance environment. Based on this, this scheme dynamically reduces the attention head pruning ratio, essentially proactively lowering the model's analytical granularity, guiding the model to ignore secondary spatial modifiers, and focusing on extracting the core odor components and overall intensity of the instruction, thereby generating a simplified but executable control intent. This effectively eliminates the waste of resources on ineffective details.

[0020] This fundamental logic can be further extended to continuous interaction and multi-user scenarios. If the environment stabilizes after a disturbance, this solution can progressively restore the model's analytical capabilities based on historical data of disturbance decay. When the same user issues new instructions within a short period, comparing the semantic vector distance between subsequent instructions can determine whether it's a correction or a new intent. If it's a correction, the pruning ratio is appropriately increased to more finely distinguish subtle differences. When different user instructions conflict, increasing the pruning ratio aims to improve the model's ability to distinguish between two independent intents, and decoupling planning is combined with user spatial location to achieve personalized experiences for different zones. When user instructions conflict with entertainment content scenarios, the weights of instructions and scenario tags in the fusion prompts are dynamically adjusted based on their semantic similarity to generate a coordinated control intent.

[0021] In summary, this solution dynamically adjusts the core computational parameter of attention head pruning ratio in a large language model through a series of quantifiable features such as environmental perturbation index, semantic dispersion, and instruction space correlation. Its innovation lies in the real-time, quantitative coupling of external physical environment constraints and internal computational resource allocation. This allows the solution's "thinking" depth to adapt to varying "execution" conditions, ultimately intelligently balancing parsing accuracy, computational efficiency, and the physical feasibility of control signals in dynamically changing personal usage scenarios, achieving a stable and reliable personalized olfactory experience.

[0022] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0023] Figure 1 This is a schematic diagram illustrating the steps of a multi-channel intelligent fragrance control method based on a large language model provided in this application embodiment. The method includes the following steps: acquiring the first user's original speech; processing the speech using a pre-trained first large language model and a first attention head pruning ratio to obtain a first language instruction; acquiring the first user's continuous airflow velocity data and calculating its root mean square value using a sliding time window as an environmental disturbance index; if the environmental disturbance index is less than or equal to a preset threshold, calculating a semantic dispersion index based on the first language instruction, increasing the first attention head pruning ratio accordingly to obtain a second attention head pruning ratio, and re-obtaining the first language instruction as the second language instruction, and outputting the first fragrance accordingly. The system executes control commands. If the environmental disturbance index is greater than a preset threshold, the system calculates the command space correlation degree based on the first language command and reduces the first attention head pruning ratio accordingly to obtain the third attention head pruning ratio. After the first time window, if the environmental disturbance index is still greater than the preset threshold, the system re-obtains the first language command based on the third attention head pruning ratio, records it as the third language command, and outputs and executes the second fragrance control command accordingly. If the environmental disturbance index is less than or equal to the preset threshold, the system obtains the range of the environmental disturbance index within the first time window, increases the third attention head pruning ratio based on it and the preset threshold, obtains the fourth attention head pruning ratio, and re-obtains the first language command based on it, records it as the fourth language command, and outputs and executes the third fragrance control command accordingly.

[0024] Figure 2 A schematic diagram of the logic flow of the multi-channel intelligent fragrance control method based on a large language model provided in the embodiments of this application.

[0025] Large language models are natural language processing models based on deep learning. They are pre-trained on massive amounts of text data and possess the ability to understand, generate, and process human language. Large language models are used to parse users' speech input and convert it into executable language commands.

[0026] Attention head pruning ratio refers to a parameter used to dynamically adjust the computational complexity of the attention mechanism in large language models. By adjusting this ratio, the model's parsing depth of input instructions and computational resource consumption can be controlled, thereby achieving a balance between model computational efficiency and instruction parsing accuracy.

[0027] Language commands are structured text information extracted from the user's original speech after processing by a large language model, used to guide the control operation of the fragrance device. These commands convey the user's olfactory intentions and control requirements.

[0028] Airflow velocity data refers to the speed information of airflow in a user's environment, collected in real time by sensors. This data reflects the dynamic changes in the environment and is an important basis for assessing the degree of environmental disturbance.

[0029] The environmental disturbance index is a quantitative indicator that characterizes the stability of the user's environment, calculated based on continuous airflow velocity data. This index is used to determine whether the environment is in a stable state, thereby guiding the large language model to adjust its instruction parsing strategy.

[0030] The semantic dispersion index is a metric that measures the degree to which semantic information is concentrated or dispersed in language instructions. When the environment is stable, this index is used to assess the granularity of instructions and adjust the attention pruning ratio accordingly to meet the user's need for fine-grained control.

[0031] Instruction space associativity is an indicator that measures the degree to which language instructions contain spatial, directional, or temporal relational information. When environmental perturbations are significant, this index is used to assess the intensity of spatial accuracy requirements in instructions and adjust the attention pruning ratio accordingly to avoid over-analysis that prevents fine-grained control.

[0032] Fragrance control commands are instructions used to drive multi-channel fragrance devices to perform specific fragrance release operations. These commands include parameters such as fragrance type, concentration, release channel, and release sequence.

[0033] A sliding time window is a data processing technique that defines a fixed-length window on a continuous data stream and slides it over time to perform real-time analysis and calculations on the data within the window. In this method, it is used to process continuous airflow velocity data in real time to calculate the environmental disturbance index.

[0034] The root mean square (RMS) value is the square root of the average of the squares of a set of values. In this method, the RMS value of airflow velocity data is used to quantify the intensity of environmental disturbances, providing a stable indicator of environmental disturbance.

[0035] The preset threshold refers to a pre-defined reference value. In this method, the preset threshold is used to compare with the environmental disturbance index to determine whether the environment is in a stable or disturbed state, and to trigger the corresponding processing logic.

[0036] The first time window refers to a preset time period within a specific processing flow for observing and evaluating the duration of environmental conditions. In this method, it is used to further determine whether the fragrance control strategy needs to be adjusted when environmental disturbances persist.

[0037] The range refers to the difference between the maximum and minimum values ​​in a set of values. In this method, the range of the environmental disturbance index within the first time window is used to assess the fluctuation range of the environmental disturbance to guide the restorative adjustment of the attention pruning ratio.

[0038] A sliding time window can be used to process continuous airflow velocity data. For example, the arithmetic mean of all airflow velocity values ​​within the window can be calculated, or the maximum value within the window can be taken as the representative value. Subsequently, the root mean square (RMS) value of the airflow velocity data within the window is calculated, and this RMS value is used as an environmental disturbance index. This index reflects the degree of dynamic change in the user's environment.

[0039] This embodiment proposes a multi-channel intelligent fragrance control method based on a large language model. By introducing environmental perturbation index, semantic dispersion index, and instruction space correlation, it achieves dynamic adaptive adjustment of the attention head pruning ratio of the large language model. This method can intelligently balance the depth of instruction parsing, computational efficiency, and the executability of control signals according to the real-time dynamic changes in the user's environment. Therefore, it effectively avoids the problems of wasted computational resources and ineffective implementation of control signals caused by excessive instruction parsing under environmental perturbations. This allows fragrance control to provide a stable and accurate personalized olfactory experience in real and varied usage scenarios, while maintaining the consistency and credibility of the user experience.

[0040] In practical applications, users may issue new voice commands within a short period of time. These new commands may have semantic connections to or deviate from the previous commands. If these subsequent commands are processed independently, it may lead to frequent changes in the fragrance control strategy or an inability to accurately capture subtle changes in user intent, thereby affecting the consistency and accuracy of the user experience.

[0041] Furthermore, after receiving a second, third, or fourth language instruction, if within the second time window, the first user's second original speech is received and a fifth language instruction is obtained accordingly, then the following steps are performed: If a second language instruction is received, the current attention head pruning ratio is the second attention head pruning ratio; if a third language instruction is received, the current attention head pruning ratio is the third attention head pruning ratio; if a fourth language instruction is received, the current attention head pruning ratio is the fourth attention head pruning ratio; the current first language instruction and the fifth language instruction are input into the pre-trained second large language model to obtain the first olfactory intent. The system calculates the first and second olfactory intention vectors; it then calculates the Euclidean distance between the first and second olfactory intention vectors, denoted as the semantic deviation between the first instructions; if the semantic deviation between the first instructions is less than or equal to the first deviation threshold, it outputs and executes the first fragrance control instruction; if the semantic deviation between the first instructions is greater than the first deviation threshold, it increases the current attention head pruning ratio based on the difference between the semantic deviation between the first instructions and the first deviation threshold, obtaining the fifth attention head pruning ratio; based on the pre-trained first large language model, it reprocesses the fifth language instruction through the fifth attention head pruning ratio to output and execute the fourth fragrance control instruction.

[0042] In this embodiment, the currently effective attention head pruning ratio is recorded to ensure that when processing new instructions, the model parameter adjustments made based on the previous understanding of the environment and user intent can be inherited and utilized.

[0043] If the semantic deviation between the first instructions is less than or equal to the first deviation threshold, it is determined that the user's new instruction is a confirmation or fine-tuning of the current state, rather than a fundamental change. In this case, the first fragrance control instruction is output and executed to maintain the current fragrance release state, avoid unnecessary adjustments, and maintain the consistency of the user experience.

[0044] If the semantic deviation between the first instructions exceeds a first deviation threshold, it indicates a significant change in user intent. To more accurately capture this change, the current attention head pruning ratio is increased based on the difference between the semantic deviation between the first instructions and the first deviation threshold, resulting in a fifth attention head pruning ratio. Increasing the pruning ratio means that the pre-trained primary language model will focus more on core semantic information when processing language instructions, reducing attention to secondary or noisy information, thereby improving sensitivity and accuracy to new user intent. Subsequently, the pre-trained primary language model uses this fifth attention head pruning ratio to reprocess the fifth language instruction, generating a more accurate language instruction representation, and outputting and executing the fourth fragrance control instruction. The fourth fragrance control instruction will reflect the user's latest, refined olfactory intent and be executed, thereby adjusting the fragrance release.

[0045] Through the above technical solution, this application can effectively address scenarios where users issue continuous commands within a short period of time. When the first user issues the second original speech and generates a fifth language command within the second time window, the new command is no longer simply processed independently, but rather its semantic deviation from the currently effective first language command is first assessed.

[0046] By inputting new and old instructions into a pre-trained second language model, an olfactory intent vector is obtained and its Euclidean distance is calculated, which can quantitatively determine the continuity or change of user intent.

[0047] If the semantic deviation is small, it indicates that the user's intention remains consistent or only slightly adjusted. In this case, maintaining the original first fragrance control command avoids unnecessary fragrance switching and ensures a smooth and consistent user experience. This avoids frequent and meaningless adjustments to fragrance control caused by minor repetitions or confirmation commands from the user.

[0048] Conversely, if the semantic deviation is large, it indicates a significant change in the user's intent. In this case, the attention head pruning ratio is dynamically increased based on the degree of deviation, resulting in the fifth attention head pruning ratio. This adjustment allows the pre-trained first language model to focus more on the core semantics of the new user instruction when reprocessing the fifth language command, improving the accuracy of capturing the new intent. In this way, it can respond to changes in user intent in a timely and accurate manner, outputting the fourth fragrance control command, thereby achieving more precise fragrance adjustment that better meets the user's current needs.

[0049] This data processing method not only improves the sensitivity to changes in user intent, but also optimizes the response speed and accuracy of fragrance control, significantly enhancing the intelligent and personalized user interaction experience.

[0050] In multi-user scenarios, when multiple users issue fragrance commands simultaneously, if there are semantic conflicts between these commands, simply executing one command or simply stacking them may lead to a chaotic fragrance experience, failing to meet the personalized needs of different users, or even causing discomfort.

[0051] Furthermore, after obtaining the second, third, or fourth language instruction, if the third original voice of the second user is received within the third time window and a sixth language instruction is obtained accordingly, the following steps are performed: Calculate the semantic deviation between the first and sixth language instructions; if the semantic deviation between the second instructions is less than or equal to a preset multi-user conflict threshold, output and execute the first fragrance control instruction; if the semantic deviation between the second instructions is greater than the preset multi-user conflict threshold, perform multi-user spatial decoupling processing; the multi-user spatial decoupling processing is as follows: Obtain the distance from the preset location to the first user's location and the distance from the preset location to the second user's location, compare the distances of the two locations and obtain the larger value, which is recorded as the distance between the two locations; based on the distance between the two locations... The distance between the first user and the second user's position is used to determine the semantic deviation between the two instructions. The current attention head pruning ratio is increased to obtain the sixth attention head pruning ratio. Based on this ratio, the first and sixth language instructions are re-obtained and designated as the seventh and eighth language instructions, respectively. According to the distance between the first and second user's positions and the preset fragrance release spatial location topology rules, a first preset release channel group is assigned to the first user, and a second preset release channel group is assigned to the second user. Based on the seventh language instruction, a first sub-control strategy is generated to control the first preset release channel group. Based on the eighth language instruction, a second sub-control strategy is generated to control the second preset release channel group. The first and second sub-control strategies are merged to generate the fifth fragrance control instruction, which is then executed.

[0052] In this embodiment, the semantic deviation between the first language instruction and the sixth language instruction is calculated. This step quantifies the degree of semantic difference between different user instructions. Specifically, the method for calculating the semantic deviation between the first instructions can be referenced. The first language instruction and the sixth language instruction are input into a pre-trained second language model to obtain the corresponding olfactory intent vectors. Then, the Euclidean distance or other similarity measure between these vectors is calculated to reflect the degree of deviation between the two instructions in terms of fragrance intent. This deviation is a key indicator for determining whether multi-user instructions conflict.

[0053] If the semantic deviation between the two instructions is less than or equal to the preset multi-user conflict threshold, the first fragrance control instruction is output and executed. When the semantic deviation between the fragrance instructions of two users is low, that is, the instruction intentions are similar or compatible, it is determined to be a non-conflicting state. At this time, the currently executing fragrance control instruction can be used, or a compatible instruction can be generated according to the fusion intention of the two instructions to maintain the stability and consistency of the fragrance environment and avoid unnecessary fragrance switching or superposition.

[0054] If the semantic deviation between the two instructions exceeds a preset multi-user conflict threshold, multi-user spatial decoupling processing is executed. When the semantic deviation between the fragrance instructions of two users is high, exceeding the preset multi-user conflict threshold, a conflict is determined to exist. To resolve this conflict and meet the personalized needs of different users, multi-user spatial decoupling processing needs to be triggered. This means that instead of simply executing a single instruction or merging instructions, an attempt will be made to separate and independently control the fragrance needs of different users in physical space.

[0055] The distances between the first user's location and the second user's location are obtained, compared, and the larger value is recorded as the distance between the two locations. This step is fundamental to multi-user spatial decoupling processing. Precise location information of the first and second users within the fragrance control area is obtained using positioning technologies such as UWB, Bluetooth, Wi-Fi positioning, and visual recognition. Then, the distances from these two locations to a reference point, such as the center point of the fragrance control device or a preset benchmark, are calculated, and the larger distance is selected as the distance between the two locations. This distance information will be used for subsequent attention pruning adjustments to accommodate the spatial distribution of different users.

[0056] To better understand and differentiate the intentions of different users in multi-user conflict scenarios, the attention head pruning ratio of the large language model is dynamically adjusted based on the semantic deviation between second instructions representing the severity of the conflict and the distance between the spatially distributed positions representing users. Increasing the attention head pruning ratio helps the model focus more on key information and reduce attention to irrelevant or conflicting information, thereby more clearly parsing the independent intentions of each user and generating more accurate seventh and eighth language instructions.

[0057] Based on the distance between the first user's location and the second user's location, and according to preset fragrance release spatial topology rules, a first preset release channel group is assigned to the first user, and a second preset release channel group is assigned to the second user. This step is crucial for achieving spatial decoupling. Based on the acquired user location information and combined with preset fragrance release spatial topology rules, for example, dividing the fragrance area into multiple sub-areas, each sub-area corresponds to one or a group of fragrance release channels, and assigning a dedicated fragrance release channel group to each user. For example, if the first user is located in area A and the second user is located in area B, the first user will be assigned a channel group controlling area A, and the second user will be assigned a channel group controlling area B. This ensures that fragrance commands from different users can be released independently through different physical channels.

[0058] Through the above technical solutions, this application effectively solves the problem of fragrance command conflicts among multiple users. By introducing semantic deviation judgment and a decoupling mechanism based on user spatial location, it can intelligently identify and handle potential fragrance demand conflicts between different users. Increasing the attention head pruning ratio helps the large language model to more accurately understand and distinguish each user's personalized fragrance intent in complex multi-user contexts, avoiding control deviations caused by ambiguous or confused commands. Simultaneously, by assigning independent fragrance release channel groups to different users and achieving independent fragrance release in physical space, it ensures that each user receives a fragrance experience that matches their personalized commands, significantly improving the accuracy, personalization, and user satisfaction of intelligent fragrance control in multi-user scenarios, and avoiding mutual interference in fragrance experiences.

[0059] Furthermore, the specific process for obtaining the sixth attention head pruning ratio is as follows: obtaining the current attention head pruning ratio, the semantic deviation between the first instructions, the multi-user conflict threshold, and the preset gain coefficient; the specific calculation formula for the sixth attention head pruning ratio is as follows: ,in, This indicates the pruning ratio of the sixth attention point. This indicates the current attention head pruning ratio. Indicates the semantic deviation between the second instructions. For multi-user conflict threshold, This indicates the preset gain coefficient.

[0060] In this embodiment, the semantic deviation between the second instruction measures the degree of semantic difference between the first language instruction and the sixth language instruction. The larger the value, the more significant the semantic conflict between the two.

[0061] The multi-user conflict threshold is a preset limit. When the semantic deviation between the second instructions exceeds this threshold, it is considered that there is a conflict that needs to be resolved.

[0062] The preset gain coefficient is a configurable parameter used to adjust the sensitivity of the pruning ratio to changes in semantic deviation, and can be obtained from linear regression analysis of historical data.

[0063] Through the above technical solution, this application provides a method for quantitatively and adaptively determining the pruning ratio of the sixth attention head. It can dynamically adjust the attention head pruning ratio based on the actual degree of semantic conflict between the first and second users. This precise adjustment based on the degree of conflict avoids excessive or insufficient pruning, ensuring that the re-obtained seventh and eighth language instructions in multi-user space decoupling processing can more accurately reflect the true intentions of their respective users, thereby effectively resolving semantic conflicts. Simultaneously, the introduction of a gain coefficient allows for a flexible balance between response speed and stability. Ultimately, this helps generate more accurate first and second sub-control strategies, achieving efficient and interference-free multi-channel intelligent fragrance control, significantly improving the user experience in multi-user scenarios.

[0064] Users' voice commands are often issued in specific digital content interaction scenarios. If fragrance control is based solely on the voice command itself, while ignoring the contextual information contained in the digital content that the user is currently interacting with, the fragrance experience may be inconsistent with the user's overall immersive experience, or even conflict with it, thereby affecting user satisfaction.

[0065] Furthermore, after obtaining the first language instruction, the process also includes: synchronously acquiring the digital content metadata of the first user's current interaction, which at least includes the content scene tag and language scene tag of the current interaction; inputting the first language instruction and the digital content metadata into the pre-trained third language model to obtain the corresponding instruction feature vector and scene feature vector, and calculating the cosine similarity between the two as the instruction scene synergy; if the instruction scene synergy is greater than or equal to the synergy threshold, it is determined to be a synergy state, and the first fragrance control instruction is executed; if the instruction scene synergy is less than the synergy threshold, it is determined to be a conflict state, triggering content linkage fusion processing to obtain the sixth fragrance control instruction and execute it.

[0066] In this embodiment, when acquiring the first user's initial voice, the digital content metadata of the first user's current interaction is simultaneously acquired to ensure that the voice command is closely related to the digital content context information in which the user is situated. This can be achieved by deploying appropriate software development kits, browser plugins, or utilizing API interfaces provided by the operation on the user's device, such as a smartphone, tablet, smart TV, or computer. When the user issues a voice command, the running applications, playing media content, or browsing webpage information on the user's device can be captured in real time, and the relevant metadata can be extracted.

[0067] Digital content metadata includes at least content context tags and language context tags for the current interaction. These tags are highly abstract and generalized descriptions of the digital content and are key to understanding its context. Content context tags can be generated through automated content analysis, keyword extraction, and topic modeling of digital content, such as using models like Latent Dirichlet Allocation (LDA) or BERT, or by labeling based on a pre-defined classification system.

[0068] For example, a movie can be tagged as science fiction, action, or romance; an article can be tagged as news, technology, or finance. Language scene tags can be generated by analyzing the language style, sentiment, and subject matter used in digital content. This typically involves natural language processing techniques such as sentiment analysis and text classification. For instance, a serious academic paper might correspond to formal, professional language scene tags; a lighthearted comedy might correspond to humorous, casual language scene tags.

[0069] The first language instruction and digital content metadata are input into a pre-trained third language model to obtain corresponding instruction feature vectors and scene feature vectors. The purpose of this step is to map the heterogeneous speech instructions (processed into text form) and the scene labels of the digital content metadata into the same high-dimensional vector space. The pre-trained third language model acts as a semantic encoder, capable of capturing deep semantic information of the instructions and scenes. This model can be the same as or different from the first language model that processed the original speech, but it needs to undergo specialized pre-training or fine-tuning to understand and encode the semantics of fragrance-related instructions and various content scenes. For example, a Transformer-based model can be used, trained through contrastive learning or multi-task learning, to encode text instructions and scene labels into semantically meaningful vectors.

[0070] The cosine similarity between the two is calculated as the instruction-scene coherence degree. Cosine similarity is a commonly used metric to measure the cosine of the angle between two non-zero vectors. Here, it is used to quantify the semantic relevance between the instruction feature vector and the scene feature vector, i.e., the degree of coherence between the instruction and the current digital content scene. The calculation result ranges from -1 to 1, with a larger value indicating higher similarity and stronger coherence.

[0071] The aforementioned technical solution enables the simultaneous acquisition of the user's voice commands and the digital content metadata of their current interaction. A pre-trained third-party language model is then used to semantically encode both, calculating the degree of synergy between the command and the scene. This mechanism not only responds to the user's explicit voice requests but also intelligently considers the implicit context in which the user is situated. When the command and scene are highly synergistic, the initial fragrance command is executed directly, ensuring a consistent user experience. Conversely, when a conflict exists between the command and scene, timely identification and content fusion processing are triggered to generate a new fragrance control command, thus avoiding inconsistencies between the fragrance experience and the user's overall immersive experience. For example, if a user is watching a tense and exciting movie and their voice command accidentally triggers an overly relaxing fragrance, the system will recognize this conflict and adjust the fragrance output to better match the movie's atmosphere. This significantly improves the intelligence level of the smart fragrance control and user satisfaction, achieving a more personalized and contextualized fragrance experience.

[0072] Furthermore, the specific acquisition process of the sixth fragrance control instruction is as follows: the first language instruction and the content scene label are combined according to the difference between the synergy threshold and the instruction scene synergy and the preset format splicing rules to generate an initial fused text; the initial fused text is then used by the pre-trained first language model to generate a ninth language instruction based on the current attention head pruning ratio, and the sixth fragrance control instruction is output accordingly.

[0073] In this embodiment, the solution combines the first language instruction with content scene tags and the quantified difference value of the conflict to generate an initial fused text through a preset formatted splicing rule. This fused text organically combines the user's original intent with the current scene information, providing a more comprehensive and contextualized input for subsequent large language model processing. Subsequently, the pre-trained first large language model performs deep understanding and processing on the fused text under the current attention head pruning ratio to generate a ninth language instruction. This ninth language instruction not only retains the user's core intent, but more importantly, it has achieved semantic synergy with the current content scene, effectively eliminating the inconsistency between the instruction and the scene. Finally, based on this synergistic ninth language instruction, a sixth fragrance control instruction is output, ensuring that the release of fragrance can accurately match the user's actual needs and the digital content environment, thereby significantly improving the coherence and immersion of the user experience and avoiding fragrance experience deviations caused by inconsistencies between the instruction and the scene.

[0074] Furthermore, the specific calculation of the semantic dispersion index includes: segmenting and filtering the first language instructions using a preset language filtering tool to obtain a valid word sequence; querying a preset odor semantic database to obtain the inverse document frequency value of each valid word in the valid word sequence; calculating the standard deviation of the inverse document frequency value of each valid word in the valid word sequence, and normalizing the standard deviation to obtain the semantic dispersion index.

[0075] In this embodiment, this application provides a refined and quantitative method for calculating the semantic dispersion index. By segmenting and filtering stop words in the first language instruction, redundant information can be effectively removed, focusing on core semantic words. By combining a pre-set odor semantic database to obtain the inverse document frequency value of each effective word, the information weight of the word in the odor domain can be objectively evaluated. Based on this, the standard deviation of these inverse document frequencies is calculated and normalized to accurately quantify the semantic concentration or dispersion of the first language instruction. This calculation method avoids subjective judgment, making the acquisition of the semantic dispersion index more scientific and reliable. When the environmental disturbance index is low, the pruning ratio of the first attention head can be adjusted more reasonably based on this accurate semantic dispersion index. For example, if the semantic dispersion index is high, it indicates that the semantics of the user instruction are relatively ambiguous or contain multiple odor intentions. In this case, increasing the pruning ratio of the attention head helps the model capture potential semantics more broadly and avoids premature convergence to a single interpretation; conversely, if the semantic dispersion index is low, it indicates that the semantics of the user instruction are clear, and the pruning ratio can be appropriately reduced to make the model more focused on accurate semantic understanding. This allows fragrance control commands to be generated more accurately in response to the user's underlying intentions, especially in stable environments, thus improving the adaptability and user experience of smart fragrance control.

[0076] Furthermore, the calculation of instruction space relevance includes the following steps: acquiring and calling a pre-trained neural dependency parsing model to perform dependency syntactic analysis on the first language instruction to obtain syntactic structure data; extracting all dependency relation types representing location, direction, or temporal relationships from the syntactic structure data; counting the total number of extracted dependency relation types as the number of spatial semantic relations; processing the first language instruction using a preset word segmentation tool to obtain the total number of instruction words; dividing the number of spatial semantic relations by the total number of words in the first language instruction after word segmentation, and obtaining the quotient as the instruction space relevance.

[0077] In this embodiment, by introducing a pre-trained neural dependency analysis model, it no longer relies solely on keyword matching but can deeply analyze the syntactic structure of the first language instruction, accurately identifying the dependency relationship types representing location, direction, or temporal relationships. This analysis method based on syntactic structure rather than surface vocabulary enables more robust and accurate capture of the user's specific requirements for the fragrance release area and timing. By normalizing the number of spatial semantic relationships with the total number of words in the instruction, the resulting instruction spatial correlation can objectively reflect the refinement of the instruction. When the environmental disturbance index is high, a high instruction spatial correlation indicates that the user's instruction has clear and precise requirements for the fragrance release location or timing. In this case, the first attention pruning ratio will be reduced, allowing the first language model to invest more computational resources and attention when processing the instruction, to understand the spatial and temporal details in the instruction more meticulously, and to avoid losing key information due to over-pruning. This ensures that even in complex environments, fragrance control can still be performed according to the user's precise intent, significantly improving the accuracy of fragrance delivery and user experience. At the same time, this mechanism of dynamically adjusting the attention head pruning ratio also avoids unnecessary complex calculations when the instructions do not contain fine spatial information, thus achieving optimized allocation of computing resources.

[0078] Furthermore, the specific process for obtaining the fourth attention pruning ratio is as follows: the recovery gain coefficient is calculated based on the difference between the range of the environmental disturbance index within the first time window and the preset threshold. The specific calculation formula for the recovery gain coefficient is as follows: ,in, Indicates the recovery gain coefficient. This represents the range of the environmental disturbance index within the first time window. This represents the preset threshold. The fourth attention head pruning ratio is calculated based on the recovery gain coefficient, the third attention head pruning ratio, and the preset attention head pruning ratio. The specific calculation formula for the fourth attention head pruning ratio is as follows: ,in, This indicates the proportion of the third attention point being pruned. This indicates the preset attention pruning ratio.

[0079] In this embodiment, when the environmental disturbance index recovers from a high-disturbance state to a stable state, the present application can adaptively calculate the recovery gain coefficient based on the range of the environmental disturbance index within a first time window and a preset threshold. This recovery gain coefficient can quantify the smoothness of the environmental disturbance recovery and guide the adjustment of the fourth attention focus pruning ratio accordingly.

[0080] Specifically, the calculation of the fourth attention head pruning ratio fully considers the current third attention head pruning ratio and the desired preset attention head pruning ratio, achieving a smooth transition by restoring the gain coefficient. This dynamic and fine-grained adjustment mechanism avoids abrupt changes in the pruning ratio, enabling the large language model to reprocess the first language instructions with a more reasonable and optimized pruning ratio after the environment stabilizes. This generates a more accurate and user-intended fourth language instruction, ultimately outputting a more appropriate third fragrance control instruction. This not only improves the adaptability and robustness of fragrance control to environmental changes but also ensures the continuity of fragrance control and the stability of the user experience during the recovery period after environmental fluctuations.

[0081] Figure 3 This is a schematic diagram of the structure of a multi-channel intelligent fragrance control system based on a large language model provided in this application embodiment. The multi-channel intelligent fragrance control system based on a large language model includes: a language command acquisition module: used to acquire the first original speech of a first user, and process it according to a pre-trained first large language model through a first attention head pruning ratio to obtain a first language command; an environmental disturbance acquisition module: used to acquire the continuous airflow velocity data of the first user, and calculate its root mean square value through a sliding time window as an environmental disturbance index; and a first fragrance control command module: used to calculate a semantic dispersion index based on the first language command if the environmental disturbance index is less than or equal to a preset threshold, and increase the first attention head pruning ratio accordingly to obtain a second attention head pruning ratio, and re-obtain the first language command, recording it as the second language command, and output the first fragrance accordingly. The system includes: a control instruction and execution module; a judgment module: if the environmental disturbance index is greater than a preset threshold, it calculates the instruction space correlation degree based on the first language instruction, and reduces the first attention head pruning ratio accordingly to obtain the third attention head pruning ratio; a second fragrance control instruction module: if, after the first time window, the environmental disturbance index is still greater than the preset threshold, it re-obtains the first language instruction based on the third attention head pruning ratio, records it as the third language instruction, and outputs and executes the second fragrance control instruction accordingly; a third fragrance control instruction: if the environmental disturbance index is less than or equal to a preset threshold, it obtains the range of the environmental disturbance index within the first time window, increases the third attention head pruning ratio based on this and the preset threshold, obtains the fourth attention head pruning ratio, re-obtains the first language instruction based on this, records it as the fourth language instruction, and outputs and executes the third fragrance control instruction accordingly.

[0082] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention 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.

[0083] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as 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.

[0084] 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.

[0085] These computer program instructions can 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.

[0086] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0087] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A multi-channel intelligent fragrance control method based on a large language model, characterized in that, Includes the following steps: Obtain the first original speech of the first user, and process it according to the first attention head pruning ratio of the pre-trained first large language model to obtain the first language instruction; Acquire the continuous airflow velocity data of the first user, and calculate its root mean square value through a sliding time window as the environmental disturbance index. If the environmental disturbance index is less than or equal to the preset threshold, the semantic dispersion index is calculated according to the first language instruction, and the first attention head pruning ratio is increased accordingly to obtain the second attention head pruning ratio. The first language instruction is then obtained again and recorded as the second language instruction. The first fragrance control instruction is then output and executed accordingly. If the environmental disturbance index is greater than the preset threshold, the instruction space correlation degree is calculated based on the first language instruction, and the first attention head pruning ratio is reduced accordingly to obtain the third attention head pruning ratio. If the environmental disturbance index is still greater than the preset threshold after the first time window, the first language instruction is obtained again according to the third attention pruning ratio, recorded as the third language instruction, and the second fragrance control instruction is output and executed accordingly. If the environmental disturbance index is less than or equal to the preset threshold, the range of the environmental disturbance index within the first time window is obtained, and the third attention head pruning ratio is increased accordingly with the preset threshold to obtain the fourth attention head pruning ratio. The first language instruction is then obtained again and recorded as the fourth language instruction. The third fragrance control instruction is then output and executed accordingly.

2. The multi-channel intelligent fragrance control method based on a large language model according to claim 1, characterized in that, After receiving a second, third, or fourth language instruction, if within the second time window, the first user's second original voice is received and a fifth language instruction is obtained accordingly, then the following steps are performed: If a second language instruction is received, the current attention head pruning ratio is the second attention head pruning ratio; If a third language instruction is received, the current attention head pruning ratio is the third attention head pruning ratio; If a fourth language instruction is received, the current attention head pruning ratio is the fourth attention head pruning ratio; The first and second olfactory intent vectors are obtained by inputting the current first language instruction and the fifth language instruction into the pre-trained second language model. Calculate the Euclidean distance between the first olfactory intent vector and the second olfactory intent vector, and denote it as the semantic deviation between the first instructions; If the semantic deviation between the first instructions is less than or equal to the first deviation threshold, then the first fragrance control instruction is output and executed; If the semantic deviation between the first instructions is greater than the first deviation threshold, the current attention head pruning ratio is increased according to the difference between the semantic deviation between the first instructions and the first deviation threshold to obtain the fifth attention head pruning ratio. The fifth language instruction is then obtained by reprocessing the pre-trained first language model through the fifth attention head pruning ratio, and the fourth fragrance control instruction is output and executed accordingly.

3. The multi-channel intelligent fragrance control method based on a large language model according to claim 2, characterized in that, After receiving a second, third, or fourth language instruction, if within a third time window, the second user's third original voice is received and a sixth language instruction is obtained accordingly, then the following steps are performed: Calculate the semantic deviation between the second instruction of the first language instruction and the second instruction of the sixth language instruction; If the semantic deviation between the second instructions is less than or equal to the preset multi-user conflict threshold, then the first fragrance control instruction is output and executed. If the semantic deviation between the second instructions is greater than the preset multi-user conflict threshold, then multi-user space decoupling processing is performed. The multi-user space decoupling process is as follows: Get the distance from the preset location to the location of the first user and the distance from the preset location to the location of the second user, compare the distances of the two locations and take the larger value, which is recorded as the distance between the locations; Based on the distance between the positions and the semantic deviation between the second instruction, the current attention head pruning ratio is increased to obtain the sixth attention head pruning ratio. The first language instruction and the sixth language instruction are obtained again according to the sixth attention head pruning ratio and are denoted as the seventh language instruction and the eighth language instruction, respectively. Based on the distance between the first user's location and the second user's location, and according to the preset fragrance release space location topology rules, a first preset release channel group is assigned to the first user, and a second preset release channel group is assigned to the second user. Based on the seventh language instructions, generate the first sub-control strategy to control the first preset release channel group; Based on the eighth language instructions, generate a second sub-control strategy to control the second preset release channel group; The first sub-control strategy is merged with the second sub-control strategy to generate the fifth fragrance control instruction and execute it.

4. The multi-channel intelligent fragrance control method based on a large language model according to claim 3, characterized in that, The specific process for obtaining the sixth attention pruning ratio is as follows: Obtain the current attention head pruning ratio, semantic deviation between first instructions, multi-user conflict threshold, and preset gain coefficient; The specific formula for calculating the sixth point of attention pruning ratio is as follows: ,in, This indicates the pruning ratio of the sixth attention point. This indicates the current attention head pruning ratio. Indicates the semantic deviation between the second instructions. For multi-user conflict threshold, This indicates the preset gain coefficient.

5. The multi-channel intelligent fragrance control method based on a large language model according to claim 2, characterized in that, After receiving the first language instruction, the method further includes: synchronously acquiring the digital content metadata of the first user's current interaction, wherein the digital content metadata includes at least the content scene tag and the language scene tag of the current interaction; The first language instruction and the digital content metadata are respectively input into the pre-trained third language model to obtain the corresponding instruction feature vector and scene feature vector, and the cosine similarity between the two is calculated as the instruction-scene synergy. If the degree of coordination of the instruction scenario is greater than or equal to the degree of coordination threshold, it is determined to be in a coordinated state, and the first fragrance control instruction is executed. If the coordination degree of the instruction scenario is less than the coordination degree threshold, it is determined to be a conflict state, triggering content linkage and fusion processing to obtain and execute the sixth fragrance control instruction.

6. The multi-channel intelligent fragrance control method based on a large language model according to claim 5, characterized in that, The specific process for obtaining the sixth fragrance control command is as follows: The first language instruction and the content scene tag are combined according to the difference between the synergy threshold and the instruction scene synergy, and a preset formatting splicing rule is used to generate an initial fused text. The initial fused text is processed by the first pre-trained language model to generate the ninth language instruction based on the current attention head pruning ratio, and the sixth fragrance control instruction is output accordingly.

7. The multi-channel intelligent fragrance control method based on a large language model according to claim 1, characterized in that, The specific calculation of the semantic dispersion index includes: The first language instructions are segmented and filtered for stop words using a preset language filtering tool to obtain a valid word sequence; Query the preset odor semantic database to obtain the inverse document frequency value of each effective word in the effective word sequence; The standard deviation of the inverse document frequency value of each effective word in the effective word sequence is calculated, and the standard deviation is normalized and used as the semantic dispersion index.

8. The multi-channel intelligent fragrance control method based on a large language model according to claim 1, characterized in that, The calculation of the instruction space associativity includes the following steps: The pre-trained neural dependency parsing model is acquired and invoked to perform dependency parsing on first language instructions to obtain syntactic structure data. Extract all dependency relation types representing location, direction, or temporal relationships from the syntactic structure data; The total number of dependency relationship types is extracted and used as the number of spatial semantic relations. The total number of words in the first language instruction is obtained by processing the instruction using a preset word segmentation tool. The quotient obtained by dividing the number of spatial semantic relations by the total number of words after word segmentation of the first language instruction is used as the instruction spatial correlation degree.

9. The multi-channel intelligent fragrance control method based on a large language model according to claim 1, characterized in that, The specific process for obtaining the fourth attention pruning ratio is as follows: The recovery gain coefficient is calculated based on the difference between the range of the environmental disturbance index within the first time window and the preset threshold. The specific calculation formula for the recovery gain coefficient is as follows: ,in, Indicates the recovery gain coefficient. This represents the range of the environmental disturbance index within the first time window. Indicates a preset threshold; The fourth attention head pruning ratio is calculated based on the recovery gain coefficient, the third attention head pruning ratio, and the preset attention head pruning ratio. The specific formula for calculating the fourth point of focus pruning ratio is as follows: ,in, This indicates the proportion of the third attention point being pruned. This indicates the preset attention pruning ratio.

10. A multi-channel intelligent fragrance control system based on a large language model, characterized in that, include: Language instruction acquisition module: used to acquire the first original speech of the first user, and to obtain the first language instruction by processing it according to the first attention head pruning ratio based on the first pre-trained first large language model; Environmental disturbance acquisition module: used to acquire continuous airflow velocity data of the first user, and calculate its root mean square value through a sliding time window as the environmental disturbance index; First Fragrance Control Instruction Module: If the environmental disturbance index is less than or equal to a preset threshold, the module calculates the semantic dispersion index based on the first language instruction, increases the first attention head pruning ratio accordingly, obtains the second attention head pruning ratio, and re-obtains the first language instruction, which is recorded as the second language instruction. The module then outputs and executes the first fragrance control instruction. Judgment module: If the environmental disturbance index is greater than a preset threshold, it calculates the instruction space correlation degree based on the first language instruction, and reduces the first attention head pruning ratio accordingly to obtain the third attention head pruning ratio. The second fragrance control instruction module is used to obtain the first language instruction again according to the third attention pruning ratio after the first time window if the environmental disturbance index is still greater than the preset threshold, record it as the third language instruction, and output and execute the second fragrance control instruction accordingly. The third fragrance control instruction is used to obtain the range of the environmental disturbance index within the first time window if the environmental disturbance index is less than or equal to the preset threshold. Based on this, the third attention head pruning ratio is increased with respect to the preset threshold to obtain the fourth attention head pruning ratio. Based on this, the first language instruction is obtained again, which is recorded as the fourth language instruction. Based on this, the third fragrance control instruction is output and executed.

Images