Construction method of large vertical control model in field of air conditioners
By constructing a large-scale vertical control model for the air conditioning field, the problems of understanding multimodal commands and matching user needs in air conditioning control systems have been solved, achieving efficient and personalized air conditioning control and improving system intelligence and user satisfaction.
Patent Information
- Application Number
- CN202511667660.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-14
- Publication Date
- 2026-02-13
AI Technical Summary
Existing air conditioning control systems suffer from problems such as high response delay, weak multimodal command understanding, inaccurate user intent parsing, difficulty in balancing energy consumption and comfort, poor robustness, and lack of safety boundaries. This leads to a disconnect between control decisions and user needs, and a lack of dynamic adjustment capabilities.
We construct a large-scale control model for the air conditioning industry. Through multimodal data fusion, physical constraint coding, user feedback mechanisms, and safety boundary definition, combined with incremental pre-training and noise adversarial training, we design dynamic instruction fine-tuning and preference-aware gating to achieve personalized strategy adjustment and risk avoidance.
It significantly improves the intelligence level and user experience of the air conditioning system, enhances the accuracy and adaptability of command parsing, reduces the control error rate, and strengthens personalized service capabilities and equipment security.
Smart Images
Figure CN121525862A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of control model construction technology, specifically referring to a method for constructing a vertical control model in the field of air conditioning. Background Technology
[0002] In current smart home systems, air conditioning control still primarily relies on rule engines or shallow machine learning models, resulting in issues such as high response latency, weak understanding of multimodal commands, inaccurate interpretation of user intent, and difficulty in balancing energy consumption and comfort. While general-purpose large language models have made progress in natural language processing, their direct application to air conditioning control faces three major bottlenecks: first, a lack of embedded modeling of the physical laws governing air conditioning; second, an inability to effectively integrate multimodal interactive data such as voice, images, and sensors; and third, a lack of user-environment-device collaborative feedback mechanisms, making it difficult to achieve dynamic adjustments to personalized strategies. Furthermore, existing solutions exhibit poor robustness under industrial noise interference and lack defined safety boundaries, easily leading to device malfunctions or even hardware damage.
[0003] However, existing methods for constructing large-scale control models for vertical categories in the air conditioning field still have certain shortcomings. Existing technologies lack targeted training and evaluation mechanisms, leading to a significant increase in command parsing error rates. They rely on single-step command matching and cannot handle multi-intent conflicts. The models lack dynamic command fine-tuning capabilities and cannot automatically adjust strategies based on user preferences, resulting in a disconnect between control decisions and user needs. They lack preference-aware gating mechanisms, failing to prioritize high-weight strategies in conflict scenarios, leading to a decline in user experience. Furthermore, they lack physical constraint encoding, failing to transform core rules into executable constraints, making the equipment prone to overload or damage in extreme scenarios. Therefore, a method for constructing large-scale control models for vertical categories in the air conditioning field is proposed. Summary of the Invention
[0004] The purpose of this invention is to provide a method for constructing a large-scale control model for vertical categories in the field of air conditioning, so as to solve the problems mentioned in the background art.
[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for constructing a large-scale control model for vertical categories in the field of air conditioning, comprising the following steps: S1. Adapt the system to the characteristics of the air conditioning field by selecting and optimizing the base model architecture; S2. Based on the optimized base model, structured fusion of domain knowledge is carried out through multimodal data collection and knowledge graph construction. S3. Based on the generated hybrid knowledge base, incremental pre-training and industrial noise adversarial training are performed. S4. Combining pre-trained models, dynamic instruction fine-tuning and complex intent parsing are performed through a dedicated instruction set and thought chain technology; S5. Based on intent parsing capabilities, it performs preference alignment and dynamic strategy adjustment through user, environment, and device feedback. S6. Based on the strategy, define the security boundary and deploy the risk avoidance mechanism through physical constraint coding and simulation testing; S7. Through edge data collection and federated learning, the model is continuously optimized and the knowledge flow is dynamically updated.
[0006] Preferably, in step S1, based on the special requirements of real-time performance and low latency in the air conditioning field, different base models are tested through rapid access capability testing to evaluate their inference speed, parameter quantity, and energy consumption ratio in air conditioning control tasks. A model architecture with high time-series processing capability and low latency characteristics is selected as the basis. For the air conditioning control scenario, the selected base model is architecturally adjusted, including embedding a time-series awareness module into the model, optimizing the attention mechanism to adapt to the dynamic changes in air conditioning operating parameters, trimming redundant computing paths to meet the resource limitations of embedded devices, and integrating air conditioning field professional terminology, working principles, and common instruction patterns into the model through knowledge distillation technology to construct specific input / output interfaces for the air conditioning field, enabling the base model to accurately parse air conditioning control instructions and generate responses that conform to physical laws.
[0007] Preferably, in step S2, structured text is extracted from air conditioning product technical documents, user manuals, maintenance records, customer service dialogue logs, and cross-language instructions, covering thermodynamic principles, control logic, fault codes, and user command patterns. Real-time operating data is collected through the built-in sensors of the air conditioning equipment, and a time-series dataset is generated by combining environmental sensors. Voice commands from user interaction with the air conditioner, customer service call recordings, and equipment prompts are recorded, and the correspondence between voice and control commands is marked. Photos of the air conditioner's exterior, screenshots of the control panel interface, images of the fault alarm interface, and user operation videos are collected, and key visual elements are marked.
[0008] Preferably, in step S2, the control panel image and text description are cross-modal aligned using the CLIP model to generate image-text association vectors, and the contextual relationships of operation actions in the image are labeled. The air conditioner model, functional modules, control commands, and user behaviors are used as core entities to establish multi-level relationships between entities, including physical relationships, functional relationships, and user relationships. Attribute mapping is performed, including text attributes, numerical attributes, and spatiotemporal attributes. The feature vectors of text, images, and speech are mapped to a unified semantic space through multimodal embedding, cross-modal similarity is calculated, entity, relationship, and attribute data are stored through an RDF triple structure, and a graph database is constructed using Apache Jena.
[0009] Preferably, in step S3, the air conditioner operating parameters in the knowledge graph are synchronized with the sensor time-series data to generate a timestamped context description, and the voice commands are concatenated with the corresponding control panel images to form a text / image and voice triplet, and the intermodal relationships are labeled. Let be the prediction error of the i-th sample under noise interference. For the prediction error of the same sample under noise-free conditions, calculate the relative impact of noise on the prediction error. Noise robustness assessment index Implementation as follows: , In the formula, N represents the total number of noise test samples.
[0010] Preferably, in step S4, the air conditioning control commands are divided into basic operations, mode switching, advanced control, and anomaly handling. Based on the generated knowledge graph, physical parameters, user intent, and cross-modal context are associated with each command. Industrial noise is injected into the command samples, and the noise type and intensity are labeled to form a command response mapping pair with interference. The typical reasoning chain structure, parameter derivation, and conflict resolution of air conditioning control are preset, and a multimodal evidence chain is embedded in the reasoning chain.
[0011] Preferably, in step S4, a learnable instruction adaptation layer is added based on the pre-trained model. Supervised training is performed using clearly labeled single-step instruction data to optimize the model's understanding of standard instructions. During fine-tuning, adversarial examples are mixed in, allowing the model to still output a complete inference chain even in noisy environments. The instruction parsing quality assessment index is achieved by weighting and integrating the instruction accuracy assessment with the actual assessment: , In the formula, This indicates the performance evaluation metrics for instruction parsing. This represents the weighting coefficient, and M represents the total number of test samples for the instruction. This represents the predicted output of the j-th instruction. This indicates the target output of the j-th instruction.
[0012] Preferably, in step S5, user ratings, voice evaluations, and behavior logs are collected through a pop-up window of the air conditioner APP. The fuzzy instructions are converted into structured feedback in real time through a trained intent parsing model, and the intensity of user emotions is labeled. Indoor sensors and external APIs are connected to generate dynamic environmental feature vectors. The compressor operating frequency, energy consumption data, and fault codes are collected in real time. Equipment health indicators are calculated and associated with the triggering scenarios of user feedback. Based on intent parsing capabilities, user feedback is accurately mapped to control parameters.
[0013] Preferably, in step S5, preference weights are extracted from the user's historical interactions, and the current preference vector is calculated statistically using a sliding window. When a user command conflicts with the environment or device status, the conflict detection module is triggered. A preference-aware gating is embedded in the model inference layer. Based on the current preference vector and environmental features, the control strategy weights are dynamically adjusted. In conflict scenarios, high-weight strategies are executed first, and the decision basis is recorded. The ternary feedback is combined with the intent parsing output to construct fine-tuning samples. Only the command adaptation layer parameters are updated, and incremental training is performed using the fine-tuning dataset.
[0014] Preferably, in step S7, the core physical rules of air conditioning are extracted from the constructed knowledge graph, the rules are transformed into executable constraints, high-frequency and high-risk scenarios are extracted based on conflict detection logs, a dynamic environment is constructed in the simulation platform, the strategy is run in the simulation environment, the statistical boundary failure probability is simulated using Monte Carlo simulation, and the safety boundary is dynamically adjusted using adaptive thresholds; a three-level avoidance mechanism is pre-designed. Level 1 Real-time Monitoring: Embed a constraint checking module in the inference layer to intercept instructions that violate security boundaries in real time; Secondary policy downgrade: When a conflict occurs, automatically switch to the preset security policy; Level 3 safety rollback: In extreme scenarios, it forces entry into safe mode; Deploy circumvention mechanisms to the inference engine of edge devices.
[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. In order to address the shortcomings of traditional air conditioning control systems in terms of energy consumption management, response efficiency, personalized adaptation, and safety boundaries, this invention constructs a large vertical control model to achieve full-process innovation from basic model optimization to continuous evolution. This method improves the accuracy of instruction parsing through multimodal data fusion and forms a closed-loop optimization by combining physical constraints and user feedback, significantly improving the intelligence level of the air conditioning system and the user experience. 2. This invention injects noise samples such as the sound of home appliances running and sensor drift, and designs noise robustness evaluation indexes to enable the model to maintain the accuracy of instruction parsing under interference conditions. Combined with joint training of time series data and cross-modal features, it improves the model's adaptability to dynamic environmental parameters and significantly reduces the control error rate caused by noise. 3. By designing a step-by-step inference chain, the present invention can generate a complete control path containing multimodal evidence. Through dynamic instruction fine-tuning, the model can adapt to different user preferences, significantly improving the flexibility of instruction parsing, realizing intelligent decision-making in multi-intent scenarios, and improving the balance between user satisfaction and equipment energy efficiency. 4. This invention utilizes ternary feedback data acquisition and preference alignment technology, combined with user APP ratings, environmental sensor data, and equipment health indicators. The model can adjust strategy weights in real time and automatically generate fine-tuning samples for model iteration. Through a preference-aware gating mechanism, the model can prioritize the execution of high-weight strategies in conflict scenarios, while recording the decision basis for subsequent optimization. This significantly improves the personalized service capabilities of the air conditioning system, ensuring that the control strategy is highly aligned with user needs and reducing the frequency of manual adjustments by users. Attached Figure Description
[0016] Figure 1 The present invention provides the operational flow of a method for constructing a large-scale vertical control model in the field of air conditioning. Figure 1 ; Figure 2 The present invention provides the operational flow of a method for constructing a large-scale vertical control model in the field of air conditioning. Figure 2 ; Figure 3 The present invention provides the operational flow of a method for constructing a large-scale vertical control model in the field of air conditioning. Figure 3 ; Figure 4 The present invention provides the operational flow of a method for constructing a large-scale vertical control model in the field of air conditioning. Figure 4 . Detailed Implementation
[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Example
[0018] Please see Figures 1-4 As shown, the present invention provides a technical solution comprising the following steps: S1. Adapt the system to the characteristics of the air conditioning field by selecting and optimizing the base model architecture; S2. Based on the optimized base model, structured fusion of domain knowledge is carried out through multimodal data collection and knowledge graph construction. S3. Based on the generated hybrid knowledge base, incremental pre-training and industrial noise adversarial training are performed. S4. Combining pre-trained models, dynamic instruction fine-tuning and complex intent parsing are performed through a dedicated instruction set and thought chain technology; S5. Based on intent parsing capabilities, it performs preference alignment and dynamic strategy adjustment through user, environment, and device feedback. S6. Based on the strategy, define the security boundary and deploy the risk avoidance mechanism through physical constraint coding and simulation testing; S7. Through edge data collection and federated learning, the model is continuously optimized and the knowledge flow is dynamically updated.
[0019] In this embodiment, S1, based on the special requirements of real-time performance and low latency in the air conditioning field, tests different base models, such as Pi0, GROOT, ZERITH-VO, etc., through rapid access capability testing, evaluates their inference speed, number of parameters and energy consumption ratio in air conditioning control tasks, and selects a model architecture with high time processing capability and low latency characteristics as the basis.
[0020] For air conditioning control scenarios, the architecture of the selected base model is adjusted, including embedding a timing awareness module into the model, optimizing the attention mechanism to adapt to the dynamic changes in air conditioning operating parameters, and trimming redundant computing paths to meet the resource constraints of embedded devices.
[0021] By incorporating air conditioning terminology, working principles, and common command patterns into the model using knowledge distillation technology, specific input / output interfaces for the air conditioning field are constructed, enabling the base model to accurately parse air conditioning control commands and generate responses that conform to physical laws.
[0022] In this embodiment, step S2 extracts structured text from air conditioning product technical documents, user manuals, maintenance records, customer service dialogue logs, and cross-language instructions, covering thermodynamic principles, control logic, fault codes, and user instruction patterns.
[0023] Real-time operating data is collected by the built-in sensors of the air conditioning equipment, and time-series datasets are generated by combining them with environmental sensors. Voice commands from user interactions with the air conditioner, customer service call recordings, and equipment prompts are recorded. The correspondence between voice commands and control commands is marked. Photos of the air conditioner's exterior, screenshots of the control panel interface, images of the fault alarm interface, and videos of user operations are collected, and key visual elements are marked.
[0024] In this embodiment, step S2 uses the CLIP model to perform cross-modal alignment between the control panel image and text description, generating an image-text association vector and annotating the contextual relationship of the operation actions in the image.
[0025] The system uses air conditioner models, functional modules, control commands, and user behaviors as core entities, establishes multi-level relationships between entities, including physical relationships, functional relationships, and user relationships, and performs attribute mapping, including text attributes, numerical attributes, and spatiotemporal attributes.
[0026] Specifically, feature vectors of text, images, and speech are mapped to a unified semantic space through multimodal embedding, cross-modal similarity is calculated, entity, relation, and attribute data are stored through RDF triple structure, and a graph database is built using Apache Jena.
[0027] In step S3, the air conditioning operating parameters in the knowledge graph are synchronized with the sensor time-series data to generate a context description with a timestamp, and the voice commands are spliced with the corresponding control panel images to form a text-image and voice triplet, and the intermodal relationships are labeled.
[0028] Specifically, let's set Let be the prediction error of the i-th sample under noise interference. For the prediction error of the same sample under noise-free conditions, calculate the relative impact of noise on the prediction error. The relative impact value represents the proportion of the error change caused by noise to the original error. The larger the value, the more significant the noise impact. This is a noise robustness assessment index. Implementation as follows: , In the formula, N represents the total number of noise test samples.
[0029] In this embodiment, step S4 divides the air conditioning control commands into basic operations, mode switching, advanced control, and exception handling. Based on the generated knowledge graph, each command is associated with physical parameters, user intent, and cross-modal context. Industrial noise is injected into the command samples, and the noise type and intensity are labeled to form a command response mapping pair with interference.
[0030] The typical reasoning chain structure for pre-defined air conditioning control is as follows: Intent recognition: extracting core needs from user input; Parameter derivation: matching physical rules with environmental data; Conflict resolution: when multiple intents conflict, generating a compromise solution based on the weight relationships in the knowledge graph, and embedding a multimodal evidence chain in the reasoning chain.
[0031] In this embodiment, step S4 adds a learnable instruction adaptation layer based on the pre-trained model. Supervised training is performed using clearly labeled single-step instruction data to optimize the model's understanding of standard instructions. During fine-tuning, adversarial examples are mixed in, allowing the model to still output a complete inference chain even in noisy environments. The instruction parsing quality assessment index is achieved by weighting and integrating the instruction accuracy assessment with the actual assessment: , In the formula, This indicates the performance evaluation metrics for instruction parsing. This represents the weighting coefficient, and M represents the total number of test samples for the instruction. This represents the predicted output of the j-th instruction. This indicates the target output of the j-th instruction.
[0032] In this embodiment, step S5 collects user ratings, voice evaluations, and behavior logs through an air conditioning APP pop-up window, and converts ambiguous commands into structured feedback in real time through a trained intent parsing model, and labels the user's emotional intensity.
[0033] By connecting indoor sensors and external APIs, dynamic environmental feature vectors are generated. The compressor's operating frequency, energy consumption data, and fault codes are collected in real time. Equipment health indicators are calculated, such as EHI=100-10×fault frequency-5×energy consumption anomaly rate. These indicators are then linked to the triggering scenarios of user feedback. Based on intent parsing capabilities, user feedback is accurately mapped to control parameters.
[0034] In this embodiment, step S5 extracts preference weights from the user's historical interactions and calculates the current preference vector using a sliding window. When a user command conflicts with the environment or device status, the conflict detection module is triggered.
[0035] A preference-aware gating system is embedded in the model inference layer. Based on the current preference vector and environmental features, the weights of the control strategy are dynamically adjusted. In conflict scenarios, high-weight strategies are executed first, and the decision basis is recorded. The three-factor feedback, user command, environmental state, and device response are combined with the intent parsing output to construct fine-tuning samples. Only the parameters of the command adaptation layer are updated, and incremental training is performed using the fine-tuning dataset.
[0036] In this embodiment, step S7 extracts the core physical rules of air conditioning from the constructed knowledge graph, transforms the rules into executable constraints, extracts high-frequency and high-risk scenarios based on conflict detection logs, constructs a dynamic environment in the simulation platform, runs the strategy in the simulation environment, simulates the statistical boundary failure probability through Monte Carlo simulation, and dynamically adjusts the safety boundary through adaptive thresholds; a three-level avoidance mechanism is pre-designed. Level 1 Real-time Monitoring: Embed a constraint checking module in the inference layer to intercept instructions that violate security boundaries in real time; Secondary policy downgrade: When a conflict occurs, automatically switch to the preset security policy; Level 3 safety rollback: In extreme scenarios, it forces entry into safe mode; Deploy circumvention mechanisms to the inference engine of edge devices.
[0037] Working principle: By evaluating and selecting a suitable base model architecture for air conditioning control tasks, customized adjustments are made to meet the real-time and low-latency requirements of the air conditioning field. Specifically, this includes: embedding a timing awareness module into the model to adapt to the dynamic changes in air conditioning operating parameters; optimizing the attention mechanism to improve the ability to capture key control features; and trimming redundant computation paths to adapt to the resource limitations of embedded devices. Through knowledge distillation technology, air conditioning terminology and control logic are injected into the model to build specific input / output interfaces, enabling the model to accurately parse air conditioning control commands and generate responses that conform to physical laws. By extracting structured text from technical documents, user manuals, maintenance records, and cross-language instructions in the air conditioning field, covering core knowledge such as thermodynamic principles, control logic, and fault codes; collecting real-time operating data and environmental data of air conditioning equipment through sensors, and annotating it with multimodal data such as user interaction voice commands and control panel images; using the CLIP model to perform cross-modal alignment of control panel images and text descriptions to generate image-text association vectors and annotate the contextual relationships of operation actions; and storing entities such as air conditioning models, functional modules, and user behaviors, as well as their physical, functional, and user relationships, through RDF triple structures, constructing a unified semantic space for multimodal embedding, and forming a queryable domain knowledge graph. Based on the constructed hybrid knowledge base, the robustness of the model is enhanced through incremental pre-training and adversarial training against industrial noise. Air conditioning operating parameters in the knowledge graph are synchronized with sensor time-series data to generate timestamped contextual descriptions. Voice commands and control panel images are concatenated into text / image and voice triples, and cross-modal relationships are labeled. During pre-training, adversarial examples are generated by injecting industrial noise, enabling the model to maintain high-precision command parsing capabilities even under noise interference. The training strategy is dynamically adjusted using noise robustness evaluation metrics. Based on the pre-trained model, a dynamic command fine-tuning process is designed, combining an air conditioning-specific command set and thought chain technology. The model's understanding of standard commands is optimized using clearly labeled single-step command data, and fine-tuning is performed accordingly. The model employs mixed industrial noise adversarial examples to ensure it can still output a complete inference chain even in noisy environments. For complex intentions, the model generates compromise solutions through weight relationships in a knowledge graph and embeds multimodal evidence chains into the inference chain. It collects user ratings, voice evaluations, and behavior logs from an air conditioning app, combining this data with environmental sensor and equipment status data to construct a three-dimensional feedback system involving users, environment, and equipment. An intent parsing model is used to convert ambiguous user commands into structured feedback in real time, and emotional intensity is labeled. User preference weights are extracted based on historical interaction data, and dynamic preference vectors are generated statistically through a sliding window. A preference-aware gating system is embedded in the model's inference layer, dynamically adjusting control strategy weights based on the current preference vector and environmental characteristics, prioritizing the execution of certain strategies. A high-weight strategy is implemented, and the decision-making basis is recorded through a conflict detection module. Based on the strategy, the core physical rules of the air conditioning field are transformed into executable constraints, and high-frequency, high-risk scenarios are constructed through a simulation platform. The failure probability of the safety boundary is statistically analyzed through Monte Carlo simulation, and the safety threshold is dynamically adjusted to adapt to the real risk level. A three-level risk avoidance mechanism is designed: Level 1 is real-time monitoring to intercept illegal commands, Level 2 is automatic switching to safe mode through strategy downgrading, and Level 3 is safety fallback to force entry into standby state in extreme scenarios. An inference engine is deployed on edge devices. User operation logs, environmental parameters, and device status data are collected in real time through edge devices, and multi-device data is aggregated using a federated learning framework for model fine-tuning to avoid centralized leakage of privacy data.The optimized strategy weights and knowledge graph updates are injected into the knowledge base built by S2, forming a closed-loop ecosystem. For example, the cooling strategy is automatically optimized in high-temperature summer scenarios, or the energy-saving mode is dynamically adjusted according to user preferences. Through continuous edge data collection and model iteration, the large-scale air conditioning control model possesses long-term adaptability and self-optimization capabilities.
[0038] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their likenesses.
[0039] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.
Claims
1. A method for constructing a large-scale control model for vertical categories in the air conditioning field, characterized in that, Includes the following steps: S1. Adapt the system to the characteristics of the air conditioning field by selecting and optimizing the base model architecture; S2. Based on the optimized base model, structured fusion of domain knowledge is carried out through multimodal data collection and knowledge graph construction. S3. Based on the generated hybrid knowledge base, incremental pre-training and industrial noise adversarial training are performed. S4. Combining pre-trained models, dynamic instruction fine-tuning and complex intent parsing are performed through a dedicated instruction set and thought chain technology; S5. Based on intent parsing capabilities, it performs preference alignment and dynamic strategy adjustment through user, environment, and device feedback. S6. Based on the strategy, define the security boundary and deploy the risk avoidance mechanism through physical constraint coding and simulation testing; S7. Through edge data collection and federated learning, the model is continuously optimized and the knowledge flow is dynamically updated.
2. The method for constructing a large-scale vertical control model in the air conditioning field according to claim 1, characterized in that: S1, based on the special requirements of real-time performance and low latency in the air conditioning field, tests different base models through rapid access capabilities to evaluate their inference speed, parameter quantity, and energy consumption ratio in air conditioning control tasks. A model architecture with high time-series processing capabilities and low latency characteristics is selected as the foundation. For the air conditioning control scenario, the selected base model is architecturally adjusted, including embedding a time-series awareness module into the model, optimizing the attention mechanism to adapt to the dynamic changes in air conditioning operating parameters, trimming redundant computing paths to meet the resource constraints of embedded devices, and integrating air conditioning field terminology, working principles, and common instruction patterns into the model through knowledge distillation technology to construct specific input / output interfaces for the air conditioning field, enabling the base model to parse air conditioning control instructions and generate responses that conform to physical laws.
3. The method for constructing a large-scale vertical control model in the air conditioning field according to claim 1, characterized in that: S2 extracts structured text from air conditioning product technical documents, user manuals, maintenance records, customer service dialogue logs, and cross-language instructions, covering thermodynamic principles, control logic, fault codes, and user command patterns. It collects real-time operating data through the air conditioning equipment's built-in sensors, combines it with environmental sensors to generate a time-series dataset, records user voice commands interacting with the air conditioner, customer service phone recordings, and equipment prompts, labels the correspondence between voice and control commands, collects photos of the air conditioner's exterior, screenshots of the control panel interface, images of the fault alarm interface, and user operation videos, and labels key visual elements.
4. The method for constructing a large-scale vertical control model in the air conditioning field according to claim 3, characterized in that: S2 uses the CLIP model to perform cross-modal alignment between the control panel image and text description, generating image-text association vectors, annotating the contextual relationships of operation actions in the image, and establishing multi-level relationships between entities, including physical relationships, functional relationships, and user relationships, by using air conditioner model, functional modules, control commands, and user behaviors as core entities. Attribute mapping is also performed, including textual attributes, numerical attributes, and spatiotemporal attributes. Multimodal embedding maps the feature vectors of text, images, and speech to a unified semantic space, calculates cross-modal similarity, stores entity, relationship, and attribute data using an RDF triple structure, and constructs a graph database using Apache Jena.
5. The method for constructing a large-scale vertical control model in the air conditioning field according to claim 1, characterized in that: In step S3, the air conditioning operating parameters in the knowledge graph are synchronized with the sensor time-series data to generate a timestamped context description. Voice commands are then concatenated with the corresponding control panel images to form a text / image / voice triplet, and the intermodal relationships are labeled. Let be the prediction error of the i-th sample under noise interference. For the prediction error of the same sample under noise-free conditions, calculate the relative impact of noise on the prediction error. Noise robustness assessment index Implementation as follows: ， In the formula, N represents the total number of noise test samples.
6. The method for constructing a large-scale vertical control model in the air conditioning field according to claim 1, characterized in that: S4 divides air conditioning control commands into basic operations, mode switching, advanced control, and anomaly handling. Based on the generated knowledge graph, it associates physical parameters, user intent, and cross-modal context for each command. It injects industrial noise into the command samples, labels the noise type and intensity, and forms a command response mapping pair with interference. It presets a typical reasoning chain structure for air conditioning control: intent recognition, parameter derivation, and conflict resolution, and embeds a multimodal evidence chain in the reasoning chain.
7. The method for constructing a large-scale vertical control model in the air conditioning field according to claim 6, characterized in that: The S4 layer, based on the pre-trained model, adds a learnable instruction adaptation layer. Supervised training is performed using clearly labeled single-step instruction data to optimize the model's understanding of standard instructions. During fine-tuning, adversarial examples are mixed in, allowing the model to still output a complete inference chain even in noisy environments. The instruction parsing quality assessment index is achieved by weighting and integrating the instruction accuracy assessment with the actual assessment: ， In the formula, This indicates the performance evaluation metrics for instruction parsing. This represents the weighting coefficient, and M represents the total number of test samples for the instruction. This represents the predicted output of the j-th instruction. This indicates the target output of the j-th instruction.
8. The method for constructing a large-scale vertical control model in the air conditioning field according to claim 1, characterized in that: The S5 process collects user ratings, voice evaluations, and behavior logs through a pop-up window in the air conditioning app. It converts ambiguous commands into structured feedback in real time using a trained intent parsing model, labels the intensity of user emotions, connects to indoor sensors and external APIs, generates dynamic environmental feature vectors, collects compressor operating frequency, energy consumption data, and fault codes in real time, calculates equipment health indicators, and associates them with the triggering scenarios of user feedback. Based on intent parsing capabilities, it accurately maps user feedback to control parameters.
9. The method for constructing a large-scale vertical control model in the air conditioning field according to claim 8, characterized in that: In step S5, preference weights are extracted from the user's historical interactions, and the current preference vector is calculated using a sliding window. When a user command conflicts with the environment or device status, the conflict detection module is triggered. A preference-aware gating is embedded in the model inference layer. Based on the current preference vector and environmental features, the control strategy weights are dynamically adjusted. In conflict scenarios, high-weight strategies are executed first, and the decision basis is recorded. The ternary feedback is combined with the intent parsing output to construct fine-tuning samples. Only the command adaptation layer parameters are updated, and incremental training is performed using the fine-tuning dataset.
10. The method for constructing a large-scale vertical control model in the air conditioning field according to claim 1, characterized in that: S7 extracts the core physical rules of air conditioning from the constructed knowledge graph, transforms the rules into executable constraints, extracts high-frequency and high-risk scenarios based on conflict detection logs, constructs a dynamic environment in the simulation platform, runs the strategy in the simulation environment, simulates the statistical boundary failure probability through Monte Carlo simulation, and dynamically adjusts the safety boundary through adaptive threshold. Pre-designed three-level avoidance mechanism: Level 1 Real-time Monitoring: Embed a constraint checking module in the inference layer to intercept instructions that violate security boundaries in real time; Secondary policy downgrade: When a conflict occurs, automatically switch to the preset security policy; Level 3 safety rollback: In extreme scenarios, it forces entry into safe mode; Deploy circumvention mechanisms to the inference engine of edge devices.