A method and system for abnormal monitoring and intelligent control of the heat and moisture load of a room air conditioner

Abstract

The present invention discloses a method and system for abnormal monitoring and intelligent control of the heating and humidity load of a room air conditioner, belonging to the technical field of intelligent control of air conditioners. The method includes: collecting indoor and outdoor environmental parameters and air conditioner operation data through sensors to construct a labeled data set; performing cloud training based on an improved lightweight Informer model, and combining a transfer learning adaptation module to quickly transfer the model to the target room; deploying the trained model to the air conditioner controller to real-time identify abnormal working conditions such as window opening, multiple people gathering, and high heat sources, and execute control strategies such as energy-saving mode and fresh air adjustment. The system includes a multi-source environmental parameter sensor module, a data processing module, an abnormal identification module, a control execution module, and a communication feedback module. The present invention solves the problem that traditional air conditioning systems cannot dynamically perceive user behavior through time series modeling and transfer learning, and improves energy consumption efficiency and comfort.

Description

Technical Field

[0001] The present invention relates to the technical fields of artificial intelligence and room air conditioner control, and particularly to a method and system for abnormal monitoring and intelligent control of the heat and moisture load of a room air conditioner. Background Art

[0002] With the development of smart home technology, as the main source of building energy consumption, the operation control mode of room air conditioners is gradually evolving from traditional fixed setpoint control to intelligent and data-driven prediction and adjustment methods. During the use of room air conditioners, due to users' frequent behaviors such as opening windows, opening doors, gathering activities, enabling additional heat sources or moisture sources, etc., sudden changes in indoor heat and moisture loads will occur, leading to increased system energy consumption, decreased equipment operation efficiency, and even adverse effects on indoor air quality and comfort. Traditional air conditioner control systems are mostly based on the PID control logic with fixed setpoints (such as Chinese patent document CN119713504A), and adjust the compressor frequency or damper opening through the deviation between the return air temperature and the setpoint. However, such methods have significant defects: firstly, they cannot dynamically perceive load disturbances caused by user behaviors (such as opening windows, multiple people gathering) or environmental mutations (such as high heat sources, high moisture sources), resulting in energy waste and decreased comfort; secondly, they rely on manual experience to adjust the setpoint and lack data-driven adaptive capabilities.

[0003] In addition, traditional air conditioner control systems mainly rely on simple temperature and humidity closed-loop control logic and cannot accurately identify and respond to the above abnormal working conditions. Especially in unattended or energy-saving scenarios, it is difficult to effectively perceive user behaviors and perform adaptive adjustments. Some studies have begun to attempt to use machine learning models for working condition identification, but most rely on static features or short-term behavior patterns and are difficult to capture the temporal evolution characteristics of user behaviors.

[0004] In recent years, long time series modeling methods, especially the Informer (Information Aggregator for Long-Term Series Forecasting) model based on the Transformer architecture, have shown excellent performance in fields such as power load forecasting and meteorological change modeling. However, the original Informer model has a large number of structural parameters and relies on large-scale data training, which is not suitable for resource-constrained edge control devices. At the same time, there are structural differences and usage behavior differences in different rooms, resulting in difficulty for a single model to be generalized for multi-scenario deployment.

[0005] Therefore, how to construct a lightweight, migratable, and multi-room personalized modeling abnormal recognition model and work in coordination with the on-site air conditioner operation system has become a key problem to be solved urgently. Summary of the Invention

[0006] The object of the present invention is to overcome the deficiencies existing in the prior art and propose a method and system for abnormal monitoring and intelligent control of the heating and humidity loads of a room air conditioner.

[0007] The object of the present invention is achieved by at least one of the following technical solutions.

[0008] A method for abnormal monitoring and intelligent control of the heating and humidity loads of a room air conditioner, comprising the following steps:

[0009] (1) Collect indoor and outdoor environmental parameters of a typical room and air conditioner operation data, construct a tagged abnormal condition data set, and upload it to a cloud server;

[0010] (2) Based on the abnormal condition data set, use an improved Informer model for cloud modeling training to obtain a lightweight condition classification model;

[0011] (3) Based on the condition classification model, through a transfer learning adaptation module, obtain a personalized model and deploy it to the air conditioner intelligent controller of the target room to predict the type of abnormal condition in real time;

[0012] (4) Execute the corresponding air conditioner operation mode control strategy according to the prediction result, and the cloud server iteratively optimizes the adaptation module based on user feedback data to achieve model optimization.

[0013] Further, in step (1), the environmental parameters include indoor temperature (return air temperature) and relative humidity (return air relative humidity), CO2 concentration , outdoor temperature and relative humidity and human activity infrared information ; the air conditioner operation data includes air conditioner power , air conditioner set temperature , air conditioner set wind speed and time stamp ; the abnormal conditions include window opening, door opening, high heat source, high humidity source, and multiple people gathering.

[0014] Further, in step (2), the encoder of the improved Informer model is a single-layer structure, and the number of attention heads is set to 2-4 (preferably 2); and multiple layers of decoders are removed, and only the prediction label of the current moment or the next moment is output.

[0015] Further, in step (2), the improved Informer model is trained using a weighted cross-entropy loss function to solve the problem of unbalanced condition category samples, and the loss function is:

[0016] ;

[0017] Among them, is the total number of samples for batch training, is the total number of categories ( , a total of 7 categories: 0 to 6), is the weight of the th category, , is the true label of the th sample, is the probability that the model predicts the th sample as category .

[0018] Furthermore, in step (3), the adaptation module is used to implement room-level model migration and personalized recognition based on the shared Informer model backbone, and adapt the trained working condition classification model to the target room; the structure of the adaptation module:

[0019] ;

[0020] Among them, is the lower projection weight matrix, is the upper projection weight matrix, and the activation function is ReLU, is the input, is the adaptation module; and are randomly initialized using a Gaussian distribution, ; ; is the standard deviation of the Gaussian distribution.

[0021] Furthermore, in step (3), obtaining the personalized model and deploying it to the air conditioner intelligent controller in the target room specifically includes:

[0022] For the working condition classification model obtained in step (2), set the core parameters of the model encoder as a shared frozen layer, and insert an initialized adaptation module at the output layer of the encoder (Encoder) to obtain a basic model;

[0023] Deploy the basic model to the air conditioner intelligent controller in the target room, only activate the parameters of the adaptation module, and freeze all other layers; collect the indoor and outdoor environmental parameters and air conditioner operation data in the target room to construct a small-scale training set;

[0024] Train the adaptation module through transfer initialization and parameter fine-tuning on the constructed small-scale training set;

[0025] After training, merge the adaptation module with the basic model into a personalized model, and deploy the personalized model to the air conditioner intelligent controller.

[0026] Further, in step (4), the control strategy includes:

[0027] When it is predicted that the window is opened or the door is opened, if the continuous detection duration exceeds the set threshold, switch to the energy-saving mode;

[0028] When it is predicted that there is a gathering of multiple people, increase the fresh air volume or raise the air conditioner wind speed;

[0029] When it is predicted that there is a high heat source or a high humidity source, adjust the set temperature or turn on the dehumidification function respectively.

[0030] Further, in step (4), the feedback data includes the recognition results marked by the user and the control log, and the model optimization is realized by regularly updating the parameters of the adaptation module through the cloud server.

[0031] The present invention also provides a room air-conditioning heat and humidity load abnormal monitoring and intelligent control system, which includes:

[0032] A multi-source environmental parameter sensor module for collecting indoor and outdoor environmental parameters and air conditioner operation data;

[0033] A data processing module for filtering, normalizing and feature construction of the data;

[0034] A cloud server for lightweight working condition classification model training, optimization of the adaptation module and data management;

[0035] An abnormal recognition module, embedded with an adaptation module and the personalized model, deployed through an edge computing device, supporting real-time inference and control decision-making;

[0036] A control execution module for adjusting the air conditioner operation mode according to the abnormal recognition result;

[0037] A communication and feedback module for data uploading, model updating and user interaction.

[0038] Compared with the prior art, the present invention has the following advantages and technical effects:

[0039] (1) Significantly reduce the computing resource requirements, by simplifying the multi-layer encoder of the original Informer model to a single layer and reducing the number of attention heads from 8 to 2, meeting the real-time requirements of edge devices (such as air conditioner controllers).

[0040] (2) Improve the recognition accuracy of abnormal working conditions, introduce a weighted cross-entropy loss function to solve the problem of sample imbalance.

[0041] (3) Optimize the input feature design (such as time difference, sliding window statistics), enhance the sensitivity of the model to mutation signals (such as the sudden drop in temperature caused by opening the window), and have a higher recognition accuracy.

[0042] (4) Achieve fast cross-scenario migration and adaptation. By inserting a lightweight adaptation module, only fine-tuning the parameters of the adaptation module in the target room requires less training time, and both the efficiency and accuracy of cross-room deployment can be improved.

[0043] (5) Closed-loop control and energy efficiency optimization. After real-time identification of abnormal operating conditions, dynamically execute adaptive control strategies (such as switching to an energy-saving mode when the window is opened and increasing the fresh air volume when many people gather), reducing the comprehensive energy consumption while maintaining the comfort of indoor temperature and humidity.

[0044] (6) Enhance system robustness and scalability. Remove the decoder module from the model, simplify the model structure, reduce redundant calculations, and improve system stability;

[0045] (7) Support user feedback and cloud iterative updates. By continuously optimizing the parameters of the adaptation module, adapt to long-term environmental changes and the evolution of user behavior. Description of the Drawings

[0046] Figure 1 is a flowchart of a method for abnormal monitoring and intelligent control of the heating and humidity load of a room air conditioner in an embodiment of the present invention.

[0047] Figure 2 is a flowchart of cloud modeling in the embodiment.

[0048] Figure 3 is a schematic diagram of the composition of a system for abnormal monitoring and intelligent control of the heating and humidity load of a room air conditioner in the embodiment. Detailed Embodiments

[0049] In order to make the objectives, technical solutions and advantages of the present invention clearer and more concise, the present invention will be further described in detail below with reference to the accompanying drawings and an embodiment. In the following description, specific details such as specific system structures and technologies are proposed for the purpose of illustration rather than limitation, so as to thoroughly understand the embodiments of the present application. However, those skilled in the art should understand that the present application can also be implemented in other embodiments without these specific details. In other cases, the detailed descriptions of well-known systems, devices, circuits and methods are omitted to avoid unnecessary details from interfering with the description of the present application.

[0050] It should be understood that when used in the specification of the present application and the appended claims, the term "comprising" indicates the presence of the described features, wholes, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or their combinations.

[0051] References to "one embodiment" or "some embodiments" or the like described in the specification of this application mean that a specific feature, structure, or characteristic described in connection with that embodiment is included in one or more embodiments of this application. Thus, statements such as "in one embodiment", "in some embodiments", "in other some embodiments", "in still other embodiments", etc. that appear in different places in this specification do not necessarily all refer to the same embodiment, but mean "one or more but not all embodiments", unless otherwise specifically emphasized in another way. The terms "comprising", "including", "having" and their variants all mean "including but not limited to", unless otherwise specifically emphasized in another way.

[0052] As Figure 1 , the following is the flow of a method for abnormal monitoring and intelligent control of the thermal and humidity load of a room air conditioner based on transfer learning and an improved lightweight Informer model proposed in this embodiment.

[0053] I. Collection of experimental data for a typical room

[0054] The collection includes human activity infrared information (infrared sensor or human monitoring device), indoor temperature (return air temperature) and relative humidity (return air relative humidity), CO2 concentration , outdoor temperature and relative humidity , air conditioner power , air conditioner set temperature and air conditioner set wind speed , timestamp and other data, and construct a data set containing abnormal condition labels.

[0055] As an embodiment, an air conditioner, a multi-source environmental parameter sensor module, and an air conditioner intelligent controller can be deployed in a representative typical room. By simulating behavior conditions such as opening windows, opening doors, high heat sources, high humidity sources, and gathering of multiple people, the environmental and air conditioner operation data inside and outside the room are collected, and a time series data set containing labels is constructed. The specific steps are as follows:

[0056] (1) Experimental environment setting: Select a representative air conditioner application environment (such as north-south orientation, different floors, etc.), and under different indoor and outdoor temperature and humidity backgrounds, simulate different user behaviors by manually setting the experimental scenarios.

[0057] (2) The data collection period is minutes, and the collected parameters include human activity infrared information (infrared sensor or human monitoring device), indoor temperature (return air temperature) and relative humidity (Return air relative humidity), CO2 concentration , outdoor temperature and relative humidity , air conditioner power , air conditioner set temperature and air conditioner set wind speed , timestamp etc.

[0058] As an abnormal condition simulation, the following experimental scenarios are carried out to enrich the data labels.

[0059] Window / door opening: Simulate different opening angles and observe the thermal and humidity disturbances;

[0060] High heat source / high humidity source: Start high heat or high humidity equipment such as electric heaters or humidifiers;

[0061] Multiple people gathering: Simulate the entry of people to increase the heat load and CO2 concentration;

[0062] Normal operation: Continuously operate under the condition of no interference;

[0063] Through the above experiments, a working condition data set with labels is constructed. The abnormal working conditions include: window opening (working condition 1), door opening (working condition 2), high heat source (working condition 3), high humidity source (working condition 4), multiple people gathering (working condition 5), equipment abnormality (working condition 6), and normal operation is recorded as (working condition 0).

[0064] II. Data preprocessing and cloud modeling

[0065] Upload the collected data to the cloud server for data cleaning, label sorting and feature construction (feature engineering).

[0066] (1) Denoise the data and smooth the data using the S-G filtering method. The filter form is:

[0067] ;

[0068] In the formula: is the data at the moment after filtering; is the window size; is the filtering coefficient; is the data at the past moment,

[0069] ;

[0070] Normalize the denoised data. Use the Z-Score method to normalize each dimension feature. The formula is as follows:

[0071] ;

[0072] In the formula: is instantaneous data; is the mean value of the data; is the standard deviation of the data. is the normalized value.

[0073] (2) Then perform feature engineering. The following features are constructed in this embodiment.

[0074] 1) Time difference feature (Δ), the following are the indoor temperature (return air temperature) and relative humidity (return air relative humidity), CO2 concentration , air conditioner power corresponding time difference features:

[0075] ;

[0076] ;

[0077] ;

[0078] ;

[0079] Wherein, , are respectively at the instant of and the indoor temperature at the instant of , are respectively at the instant of and the relative humidity at the instant of , are respectively at the instant of and the CO2 concentration at the instant of , are respectively at the instant of and the air conditioner power at the instant of .

[0080] 2) Cross-parameter difference feature

[0081] Indoor-outdoor temperature difference: ; Relative humidity difference: ; Wherein is the outdoor temperature at the instant of , and is the relative humidity at the instant of .

[0082] 3) Sliding window statistical feature

[0083] Through a time window with a length of (which can be 3 minutes in this embodiment), the change trend is statistically analyzed to improve the perception ability of slow signals.

[0084] Average indoor temperature: ;

[0085] Standard deviation of indoor temperature: ;

[0086] 5) One-hot time period feature

[0087] Encode the time features of hours (0 - 23) and day of the week (0 - 6).

[0088] (3) Using the processed data, train an improved lightweight Informer model in the cloud to output a working condition classification model. As Figure 2 , it is the cloud modeling process, which specifically includes:

[0089] Step 1: Construct input samples. With a sliding window length of Construct sequence input: , ;

[0090] Step2: Improve the lightweight Informer model structure, and the structure has the following key features:

[0091] a) The encoder is a single-layer structure, and the number of attention heads is set to 2;

[0092] b) Remove multiple layers of Decoder and only output the prediction label of the current moment or the next moment;

[0093] Step3: Introduce class weights to optimize the loss function. Use a weighted multi-class cross-entropy loss function to solve the sample imbalance:

[0094] ;

[0095] Among them, is the total number of samples in batch training, is the total number of classes ( , a total of 7 classes: corresponding to working conditions 0 to 6), is the weight of the th class, , is the true label of the th sample, is the probability that the model predicts the th sample as the th class.

[0096] Step4: Model evaluation. The accuracy and F1-Score are used to evaluate the model performance, and higher recall rate for high-energy consumption anomalies is ensured preferentially.

[0097] Accuracy

[0098] 1 is the indicator function, which is 1 if the prediction is correct and 0 otherwise;

[0099] The F1-Score is calculated using weighted average:

[0100] ；

[0101] is the class 's F1-score; : the number of true samples of class ;

[0102] ;

[0103] Among them, , indicating that among the samples predicted by the model as class , the proportion that truly belongs to ; , indicating that among the samples that truly belong to class , the proportion correctly predicted by the model. (True positive): the number of samples where the truth is and the prediction is ; (False positive): the number of samples where the truth is not but the prediction is ; (False negative): the number of samples where the truth is but the prediction is not ;

[0104] III. Transfer learning adaptation module

[0105] Deploy the base model in the target room (new room) and insert the adaptation module, and only train the adaptation module to achieve rapid model transfer.

[0106] The fine-tuning steps for transfer learning modeling in the target room are as follows:

[0107] Step1. Base model construction. Build a complete Informer model trained on typical room data, set the core parameters of the Encoder as a shared frozen layer, and insert an initialized adaptation module at the output layer of the Encoder to obtain the base model;

[0108] Step2. Initialize the target room model. Deploy the base model to the intelligent controller of the target room air conditioner, only activate the parameters of the adaptation module, and freeze all other layers; collect a small amount of new room data to construct a small-scale training set.

[0109] Step3. Local training and fast convergence. Train the adaptation module on the target room data. Since the total number of parameters is small, the training time is short. If there is not enough data, transfer initialization + parameter fine-tuning can be used.

[0110] Step4. Model fusion and deployment. After training, merge the new room adaptation module with the base model into a personalized model.

[0111] To solve the problem of insufficient initial data in the new room, this embodiment designs a transfer learning strategy, uses parameter sharing and an adaptation module for transfer fine-tuning in the new room, and quickly adapts the trained typical room model to the target room.

[0112] The adaptation module proposed in this embodiment is used to achieve room-level model transfer and personalized recognition on the basis of sharing the backbone of the Informer model. The adaptation module uses a bottleneck structure (a two-layer fully connected neural network) to achieve low-dimensional mapping and residual connection.

[0113] Structure of the adaptation module: ,

[0114] Among them, is the lower projection weight matrix, is the upper projection weight matrix, and the activation function is ReLU, , is the input, is the adaptation module.

[0115] IV. Model Deployment and Prediction

[0116] Deploy the trained room personalized model to the intelligent controller of the air conditioner, collect real-time data, input it into the Informer model, and output the prediction result of the current working condition; this embodiment supports a continuous update and incremental training mechanism; if the system detects a decrease in the recognition rate or user feedback on misjudgment, the adaptation module can be retrained.

[0117] V. Abnormality Recognition and Intelligent Control

[0118] Judge the type of working condition according to the recognition result, including window opening, door opening, high heat source, high humidity source, multiple people gathering, etc., and execute the corresponding air conditioner operation mode adjustment strategy.

[0119] As an embodiment, the operation mode adjustment strategy includes:

[0120] When it is predicted that the window or door is opened, if the continuous detection duration exceeds the set threshold, switch to the energy-saving mode;

[0121] When it is predicted that there are multiple people gathering, increase the fresh air volume or raise the air conditioner wind speed;

[0122] When it is predicted that there is a high heat source or a high humidity source, adjust the set temperature or turn on the dehumidification function respectively.

[0123] VI. Control Feedback and Remote Synchronization

[0124] Record the control decision log and upload it to the cloud server, supporting user feedback correction and model update.

[0125] As an embodiment, such as Figure 3 , a system for abnormal monitoring and intelligent control method of room air conditioner heat and humidity load mainly includes:

[0126] A cloud server, used for lightweight working condition classification model training, optimization of the adaptation module and data management;

[0127] An air conditioner intelligent control module, which may include: a data processing module, used to realize on-site data collection, data processing and operation control, and data processing includes filtering, normalization and feature construction; an abnormal identification module, embedded with an adaptation module and the personalized model, deployed through an edge computing device, supporting real-time inference and control decision-making; a control execution module, which adjusts the air conditioner operation mode according to the abnormal identification result; a communication and feedback module, used for data uploading, model update and user interaction.

[0128] A user APP, used to receive the identification result, provide feedback and remotely control the air conditioner.

[0129] In this embodiment, it also includes a multi-source environmental parameter sensor module, used to collect indoor and outdoor environmental parameters and air conditioner operation data; in this embodiment, the multi-source environmental parameter sensor module may include an indoor temperature sensor, an indoor humidity sensor, an outdoor temperature sensor, a human body infrared temperature sensor, a carbon dioxide sensor, a power sensor, etc.

[0130] As an embodiment, the air conditioner intelligent control module is deployed on the air conditioner intelligent controller. The air conditioner intelligent controller in this embodiment refers to an air conditioner controller with core computing resources, storage and communication functions, and can adopt an existing air conditioner controller or upgrade the chip on the existing air conditioner controller to meet the operation of the relevant modules of the present invention; its specific working steps are as follows:

[0131] Step1: Real-time multi-source data collection. Through various sensors deployed in the room, periodically collect the aforementioned environmental and equipment parameters. The data collection period is recommended to be set to 1 minute, and all raw data enters the next-stage processing module;

[0132] Step 2: Data preprocessing and feature construction. Filter and denoise the collected multi-dimensional data (Savitzky-Golay filter) and perform normalization (Z-score normalization); construct differential features and cross features; use a sliding window to build the input sequence;

[0133] Step 3: Abnormal condition identification and reasoning. Input the processed input feature sequence into the Informer lightweight model deployed in the air conditioner intelligent controller. The model performs reasoning and outputs the predicted result of the operating condition at the current moment:

[0134] Step 4: Intelligent control of the operating mode. According to the identification result, the system matches the preset control strategy library and executes the corresponding adjustment of the air conditioner operating mode, such as:

[0135] a) Opening a window or a door (operating conditions 1 and 2): If it persists for ≥6 minutes (2 sampling periods), automatically switch to the energy-saving operating mode;

[0136] b) Multiple people gathering (operating condition 3): Increase the fresh air volume or wind speed and control the CO2 concentration;

[0137] c) High heat source (operating condition 4): Increase the set temperature by 1°C - 2°C and switch to the energy-saving mode;

[0138] d) High humidity source (operating condition 5): Turn on the dehumidification mode and appropriately reduce the humidity setting;

[0139] e) Equipment abnormality (operating condition 6): Pause operation and give an alarm on the cloud;

[0140] f) Normal (operating condition 0): Restore the operating state set by the user.

[0141] Step 5: Control feedback and remote synchronization. The controller stores the identification result and control action in the form of a log and uploads it regularly; if the user notification function is enabled, the system pushes a message to the user APP through the cloud; the user can confirm, adjust, or annotate the identification result through the feedback button for subsequent continuous optimization of the model.

[0142] Step 6: Model and policy update mechanism. The air conditioner intelligent controller regularly evaluates the model identification accuracy and the consistency of user feedback; if the model performance deteriorates or the environment changes significantly, the cloud update process is triggered. In this embodiment, after training the model with typical room data, it is migrated to the target room (new room) through the adaptation module. The air conditioner intelligent controller identifies abnormal conditions based on real-time data and executes control strategies. The user can feedback and correct the model parameters through the APP.

[0143] It should be understood that the present invention is not limited to the above specific example embodiments. Any person skilled in the art within the technical scope disclosed by the present invention, any modifications, substitutions, combinations, simplifications, etc. made under the spirit and principle of the present invention are all equivalent replacement methods and should be included within the protection scope of the present invention.

Claims

1. A method for abnormal monitoring and intelligent control of heat and humidity load of room air conditioner, characterized in that: The following steps are involved: (1) Collect indoor and outdoor environmental parameters of typical rooms and air conditioning operation data, build a labeled abnormal operating condition data set, and upload it to the cloud server; (2) Based on the abnormal working condition data set, the Informer model is used to perform cloud modeling training to obtain a lightweight working condition classification model; the Informer model is an improved Informer model; the encoder of the improved Informer model is a single-layer structure, the number of attention heads is 2-4, and the multi-layer decoder is removed, and only the predicted label of the current moment or the next moment is output; (3) Based on the working condition classification model, a personalized model is obtained through transfer learning adaptation module and deployed to the air conditioning intelligent controller of the target room to predict the abnormal working condition type in real time; (4) Execute the corresponding air-conditioning operation mode control strategy according to the prediction results. The cloud server iteratively optimizes the adaptation module based on the user feedback data, and then optimizes the personalized model.

2. A room air conditioning heat and humidity load abnormal monitoring and intelligent control method according to claim 1, characterized in that: In step (1), the environmental parameters include indoor temperature and relative humidity, CO2 concentration, outdoor temperature and relative humidity, and infrared information of human activities; the air conditioning operation data includes air conditioner power, air conditioner set temperature, air conditioner set wind speed and timestamp; the abnormal operating conditions include open windows, open doors, high heat sources, high humidity sources, and large crowds.

3. A room air conditioning heat and humidity load abnormal monitoring and intelligent control method according to claim 1, characterized in that: In step (2), the improved Informer model is trained using a weighted cross entropy loss function to solve the problem of unbalanced samples of working condition categories. The loss function is: ； in, is the total number of samples for batch training, is the total number of categories, For the The weight of the class, , For the The true labels of samples, For the model Samples are predicted as categories probability.

4. A room air conditioning heat and humidity load abnormal monitoring and intelligent control method according to claim 1, characterized in that: In step (3), the adaptation module is used to realize room-level model migration and personalized recognition based on the shared Informer model backbone, and adapt the trained working condition classification model to the target room; the adaptation module structure is: ； in, is the down-projection weight matrix, is the up-projection weight matrix, the activation function is ReLU, For input, For the adapter module; and The random initialization of adopts Gaussian distribution.

5. A room air conditioning heat and humidity load abnormal monitoring and intelligent control method according to claim 1, characterized in that: In step (3), the step of obtaining the personalized model and deploying it to the air conditioning intelligent controller of the target room specifically includes: For the working condition classification model obtained in step (2), the core parameters of the model encoder are set to the shared frozen layer, and the initialized adaptation module is inserted into the output layer of the encoder to obtain the basic model; Deploy the basic model to the target room air conditioner intelligent controller, activate only the adaptation module parameters, and freeze all other layers; collect the indoor and outdoor environmental parameters of the target room and the air conditioner operation data to build a small-scale training set; The adaptation module is trained by transferring initialization and fine-tuning parameters on the constructed small-scale training set; After training is completed, the adaptation module and the basic model are merged into a personalized model, and the personalized model is deployed to the air-conditioning intelligent controller.

6. A room air conditioning heat and humidity load abnormal monitoring and intelligent control method according to claim 1, characterized in that: In step (4), the control strategy includes: When the prediction is that the window or door is open, if the continuous detection time exceeds the set threshold, it will switch to energy-saving mode; When a large gathering of people is predicted, increase the fresh air volume or the air conditioning speed; When a high heat source or a high humidity source is predicted, the set temperature is adjusted or the dehumidification function is turned on respectively.

7. A room air conditioning heat and humidity load abnormal monitoring and intelligent control method according to claim 1, characterized in that: In step (4), the feedback data includes the recognition results and control logs annotated by the user, and the cloud server regularly updates the adaptation module parameters.

8. A system for implementing the method for abnormal monitoring and intelligent control of heat and humidity load of a room air conditioner according to any one of claims 1 to 7, characterized in that: include: Multi-source environmental parameter sensor module, used to collect indoor and outdoor environmental parameters and air conditioning operation data; Data processing module, used for filtering, normalizing and feature construction of data; Cloud server, used for lightweight working condition classification model training, adaptation module optimization and data management; An anomaly recognition module, which is used to embed the adaptation module and the personalized model, and is deployed through edge computing devices to support real-time reasoning and control decisions; A control execution module, used for adjusting the air-conditioning operation mode according to the abnormality recognition result; Communication and feedback module, used for data uploading, model updating and user interaction.

9. The system according to claim 8, characterized in that The multi-source environmental parameter sensor module includes an indoor temperature sensor, an indoor humidity sensor, an outdoor temperature sensor, a human infrared temperature sensor, a carbon dioxide sensor and a power sensor.

Images