Automobile air conditioner intelligent regulation and control method based on multi-task learning neural network

Through the combination of multi-task learning neural network and feature isolation layer, the problems of rough control particle size and network dependence of automotive air conditioning systems are solved, real-time and personalized air conditioning parameter settings are realized, and user experience and vehicle performance are improved.

CN119058319APending Publication Date: 2024-12-03SHENZHEN LANYOU TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202310633215.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2023-05-31
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

The existing automotive air conditioning system relies on manual operation of users, has rough control particle size, and is heavily dependent on the network environment, resulting in poor response timeliness. Multi-model processing increases the computing and storage burden of the vehicle computer, affecting the user experience.

Method used

Using intelligent regulation method based on multi-task learning neural network, by obtaining real-time data, a multi-task learning model including air conditioning switches, temperature regulation and internal and external circulation modes is built, feature isolation layer is embedded, network structure is defined, model training is carried out, and multi-task learning model is converted to ONNX The format is deployed on the vehicle side for real-time inference.

Benefits of technology

It realizes more refined and personalized thermal comfort control of air conditioners, reduces vehicle resource consumption, shortens model reasoning time, ensures the timeliness of user experience, and reduces operational complexity.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119058319A_ABST
    Figure CN119058319A_ABST
Patent Text Reader

Abstract

The invention discloses a multi-task learning neural network-based intelligent regulation and control method for an automobile air conditioner, which comprises the following steps of: firstly, constructing a multi-task learning model of three tasks according to various requirements of an air conditioner thermal comfort system, and then, constructing a feature isolation layer according to the correlation between the three tasks and an automobile connection data feature field; the feature isolation layer is embedded into a multi-task learning model, the model weight can be continuously improved along with data accumulation, then multiple devices of the air conditioner thermal comfort system share one model, and adjusting parameters of an air conditioner switch, temperature setting and an internal and external circulation mode in the air conditioner thermal comfort system are predicted; and finally, packaging the model structure and the model weight after cloud training and verification into a PB-format file, converting the PB-format file into an ONNX-format file, deploying the ONNX-format file at a vehicle end, and performing real-time reasoning. According to the invention, the vehicle end provides the processes of air conditioner switching, temperature adjustment and internal and external circulation mode setting operation for a vehicle owner in different scenes according to the vehicle connection data collected in real time.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of automobile air conditioners, and in particular to an intelligent control method for automobile air conditioners based on a multi-task learning neural network. Background Art

[0002] The automotive air conditioning system is the core system to ensure the thermal comfort of the car. Under the background of people's growing needs for a better life, the automotive air conditioning system is also facing new challenges and needs to achieve more refined and personalized thermal comfort settings. At the same time, with the upgrading of the industry, the intelligentization of automotive air conditioning has become one of the research focuses in the automotive field.

[0003] The current automotive air-conditioning system has the following problems: 1. It relies on manual operation by the user and is not intelligent enough, or the PLC processes and makes logical judgments on the data collected by the vehicle, and based on this, realizes the conversion of the air-conditioning air circulation mode and the improvement of the air quality in the car through the drive control mechanism. It is heavily dependent on the experience of product personnel, and the control granularity is rough, which deviates from the actual usage habits of users; 2. The model is deployed in the cloud, and the vehicle collects data and uploads it to the cloud for model reasoning. The generated prediction results are sent to the vehicle, which is heavily dependent on the network environment, has poor response timeliness, and affects the user experience; 3. Various air-conditioning systems, such as internal and external circulation, air-conditioning temperature setting and other operations, require multiple models to process independently. The increase in the number of models will increase the computing and storage burden of the vehicle computer, and will cause the vehicle computer to run slowly. Summary of the invention

[0004] The embodiment of the present invention provides an automobile air-conditioning intelligent control method based on a multi-task learning neural network to solve the deficiencies of the prior art described in the background technology.

[0005] The present invention provides an automobile air conditioner intelligent control method based on a multi-task learning neural network, comprising the following steps:

[0006] Obtain model input data, splice the model input data collected in real time into a growth vector and extract the feature fields of the model input data to obtain a model training data set, and build a multi-task learning neural network model for three tasks based on the air-conditioning operation to be controlled and the three tasks of model output;

[0007] Based on expert experience, according to the correlation between the feature fields of the model input data and the three tasks, the feature fields of the model input data are divided into three feature data groups corresponding to the three tasks, and an index dictionary of the three tasks and the feature data groups is constructed, and a feature isolation layer is constructed according to the index dictionary, and the feature isolation layer is embedded between the input layer of the multi-task learning neural network model and the feature extraction layers of the three tasks to obtain a model calculation graph;

[0008] Define the network structure according to the model calculation graph, select the corresponding activation function, loss function and model evaluation index according to the types of the three tasks, perform model training and verification to obtain a PB format model file, which contains the model structure and the weight coefficients obtained by training;

[0009] The PB format model file is converted into an ONNX model format file and deployed on the vehicle side. After the real-time data collected by the vehicle side is preprocessed, it is input into the ONNX model to complete model reasoning, and finally the air conditioning switch, temperature adjustment and internal and external circulation parameter settings that the car owner likes to execute at the current moment are given.

[0010] Furthermore, the data types of the model input data include vehicle connection data, vehicle environment data, vehicle owner attributes and operation preference data.

[0011] Furthermore, the characteristic fields of the model input data include at least air conditioning cycle mode, air conditioning switch status, air conditioning temperature setting, air quality, PM2.5, indoor and outdoor temperature, indoor and outdoor humidity, vehicle cruising range, and window switch status.

[0012] Furthermore, the three tasks include a switch binary classification task, a temperature adjustment multi-classification task, and an internal and external circulation mode binary classification task; wherein: the switch binary classification task predicts whether the air conditioner is turned on or off; the temperature adjustment multi-classification task adjusts the temperature value with a granularity of 1°C and a range of 16°C-32°C; the internal and external circulation mode binary classification task predicts whether the air conditioner turns on the internal circulation or the external circulation.

[0013] Furthermore, the multi-task learning neural network model is a mapping of source data and target data, which is defined as follows:

[0014] The source data is defined as X, which represents a sample set in a certain field; the target data is defined as Y, which represents a target set in a certain field; the mapping is defined as f, f:X→Y, which represents the conversion relationship from source to target; wherein the mapping includes a shared parameter layer and an independent parameter layer corresponding to each task, and the output end of the shared parameter layer is connected to the independent parameter layer corresponding to each task.

[0015] Furthermore, the feature fields of the model input data are divided into three types of feature data groups corresponding to three types of tasks, specifically:

[0016] The switch binary classification task corresponds to the first feature data group, which includes feature fields: vehicle interior temperature, vehicle exterior temperature, remaining mileage, window status, owner's gender, age, and owner's permanent location;

[0017] The internal and external circulation mode binary classification task corresponds to the second feature data group, including feature fields: internal and external temperature, humidity, carbon monoxide concentration, carbon dioxide concentration, PM2.5 concentration, air conditioning cooling and heating status, and the owner's permanent residence;

[0018] The temperature adjustment multi-classification task corresponds to the third feature data group, which includes fields: current setting temperature of the air conditioner, temperature inside and outside the car, humidity, gender and age of the owner, and permanent residence of the owner.

[0019] Furthermore, after the feature isolation layer is embedded between the input layer of the multi-task learning neural network model and the feature extraction layers of the three tasks, after the feature isolation layer obtains the input feature vector of the input layer of the multi-task learning neural network model, it automatically groups all feature fields according to the index dictionary of the three tasks and feature data groups, and enters the feature extraction layer of the corresponding task.

[0020] Furthermore, the model computation graph defines the network structure and specifically uses the TensorFlow framework to define the network computation graph structure.

[0021] Furthermore, the model training and verification are performed to obtain a PB format model file, specifically: the model data set is divided into a training set and a verification set in a ratio of 7:3, model training and verification are performed, the model structure and weights after model training are obtained, and the model is saved as a PB format model file.

[0022] Furthermore, the converting of the PB format model file into the ONNX model format file is specifically: using the ONNX module of the TensorFlow framework to convert the PB format model file into the ONNX model format file.

[0023] At least one of the above technical solutions adopted in the embodiments of the present invention can achieve the following beneficial effects:

[0024] 1. The present invention constructs a multi-task learning model including 2 classifications of air-conditioning switches, 17 classifications of air-conditioning temperatures and 2 classifications of internal and external circulation according to various requirements of the air-conditioning thermal comfort system. The user's operating habits of the air-conditioning thermal comfort system are learned through data, and the model weight can be continuously improved as the data accumulates.

[0025] 2. The present invention constructs a multi-task learning neural network model with a single embedded feature isolation layer to enable multiple devices in the air-conditioning thermal comfort system to share one model, predict the adjustment parameters of multiple air-conditioning thermal comfort systems, speed up model reasoning, and reduce vehicle computer resource consumption.

[0026] 3. The present invention packages the model structure and model weights generated by cloud training, converts them into ONNX models, and deploys them on the vehicle side, so as to control the time interval from data collection to model reasoning and to the execution of actions by the vehicle computer within 100ms, thereby ensuring the timeliness of the user experience and realizing data-driven more refined, more automated, higher-performance, and more personalized automotive air-conditioning thermal comfort system algorithm control. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0028] Figure 1 It is a schematic diagram of the principle structure summary of the present invention.

[0029] Figure 2 Schematic diagram of the SIMO multi-task learning model architecture of the present invention.

[0030] Figure 3 Schematic diagram of the multi-task learning model framework of the present invention.

[0031] Figure 4 Schematic diagram of the multi-task learning model framework of the embedded feature isolation layer of the present invention. DETAILED DESCRIPTION

[0032] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the specific embodiments of the present invention and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0033] The technical solutions provided by various embodiments of the present invention are described in detail below in conjunction with the accompanying drawings.

[0034] like Figure 1-4 As shown, the present invention provides an intelligent control method for automobile air conditioning based on a multi-task learning neural network embedded with a feature isolation layer, and the specific steps are as follows:

[0035] Step 1. Construct a multi-task learning neural network model: obtain model input data, including: vehicle connection data, vehicle environment data, owner attributes and operation preference data, and control the air-conditioning operations as needed: air-conditioning switch, air-conditioning temperature adjustment, internal and external circulation mode, corresponding to the three tasks of the model output: switch binary classification task, air-conditioning temperature adjustment multi-classification task, internal and external circulation mode binary classification task, and construct a multi-task learning neural network model for the three tasks.

[0036] Step 2, embedding feature isolation layer: Based on expert experience and the correlation between the model input feature fields and the three air-conditioning operation tasks, the model input feature fields are divided into three types of feature data groups corresponding to the three tasks, and an index dictionary of tasks and feature data groups is constructed. The feature isolation layer is constructed based on the index dictionary and embedded between the input layer of the multi-task learning neural network model and the three task feature extraction layers in step 1. After the vector input model of the full feature field enters the feature isolation layer, the feature isolation layer will automatically group the feature fields according to the index dictionary and then input them into the feature extraction layer of the corresponding task. At this point, the construction of the multi-task learning neural network embedded in the feature isolation layer is completed.

[0037] Step 3, model conversion: After the operation in step 2, the model calculation graph is obtained, and the network structure is defined according to the calculation graph. The corresponding activation function, loss function and model evaluation index are selected according to the types of the three tasks. The model training and verification are performed to obtain the PB format model file, which contains the model structure and the weight coefficients obtained through training. Then, the PB format model file is converted into the ONNX model format file and deployed on the vehicle side. After the real-time data collected by the vehicle side is preprocessed, it is input into the ONNX model to complete the model reasoning, and finally the air-conditioning switch, temperature adjustment and internal and external circulation parameter settings that the car owner likes to execute at the current moment are given.

[0038] The specific steps of step 1 are as follows:

[0039] Step 1.1: First, the vehicle splices the real-time collected data into a long vector package and uploads it to the cloud through the form of buried points. At this time, the fields included in the data include: air conditioning cycle mode, air conditioning switch status, air conditioning temperature setting, air quality, PM2.5, vehicle inside and outside temperature, vehicle inside and outside humidity, vehicle cruising range, window switch status, etc.

[0040] Step 1.2, then, combine cloud statistics to calculate the preference data of each car owner for air conditioning internal and external circulation mode, temperature setting, and air conditioning switch in different scenarios, and build a multi-task model training data set. In the cold start phase, expert experience is required to formulate labeling rules based on the data collected in step 1.1 for model training. After the real data is collected later, the data set labels are adjusted and the model is updated;

[0041] Step 1.3, build a multi-task learning model. The model output tasks are divided into 3: the first is the air conditioner switch binary classification, which needs to predict whether the air conditioner is turned on or off based on the step 1.3 data set; the second is the air conditioner adjustment temperature value, with a granularity of 1°C and a range of 16°C-32°C, so the second task can be defined as temperature 17 classification; the third is to predict whether the air conditioner turns on the inner or outer circulation based on the step 1.3 data set. According to the model input and output tasks, a multi-task learning model is built, and the number of tasks is 3;

[0042] Among them, a task is defined as follows:

[0043] 1. Source data: X represents a sample set in a certain field

[0044] 2. Target data: Y represents the target set in a certain field

[0045] 3. Mapping: f:X→Y, indicating the conversion relationship from source to target

[0046] For a multi-task learning model (MTL), multiple of the above tasks can be modeled at the same time. Some parameters in the model are shared by all tasks, called "shared parameters"; the other part of the parameters are unique to each task, called "independent parameters". Ideally, we hope that the "shared parameters" are updated after all samples calculate the Loss, while the "independent parameters" are only updated after the corresponding task samples calculate the Loss, thus achieving the purpose of training together. With the above definition, the main form of MTL can be defined: given a source dataset (X), and multiple target datasets (Y 1 , Y 2 ..., Y N ), MTL maps X to each target dataset separately, such as Figure 2 This form of MTL is called Single Input MultiOutput (SIMO).

[0047] With the help of SIMO-based multi-task learning model, the structure of the multi-task learning model constructed by this method is as follows: Figure 3 shown.

[0048] Step 1.4: Post-process the multi-task model output to obtain the outputs of the three tasks in step 1.3 and integrate the outputs of the three tasks into an output vector, which is the model output.

[0049] The specific steps of step 2 above are as follows:

[0050] Step 2.1, after obtaining the model training data set through step 1, the model input vector is grouped into features according to the three tasks combined with expert experience. The air conditioning switch binary classification task corresponds to feature group 1, including feature fields: indoor temperature, outdoor temperature, remaining mileage, window status, owner's gender, age, owner's permanent residence, etc.; the air conditioning internal and external circulation binary classification task corresponds to feature group 2, including feature fields: indoor and external temperature, humidity, carbon monoxide concentration, carbon dioxide concentration, PM2.5 concentration, air conditioning cooling and heating status, owner's permanent residence, etc.; the air conditioning temperature 17 classification task corresponds to feature group 2, including fields: current air conditioning setting temperature, indoor and outdoor temperature, humidity, owner's gender, age, owner's permanent residence, etc.

[0051] Among them, since the model is a multi-task learning model, one model contains multiple tasks, but not all model input feature fields are related to each task of the multi-task model. Directly entering all samples into the model training will introduce a lot of noise and affect the model effect. Therefore, in the model data set, the features need to be grouped according to their relevance to the task, and the shared parameter layer in the model architecture must be cancelled.

[0052] like Figure 4 As shown, it is a schematic diagram of the multi-task learning neural network framework of the present invention embedded with a feature isolation layer. Feature isolation is adopted because different tasks depend on different input features. For example, the task of adjusting the internal and external circulation of the air conditioner depends on the characteristics of carbon monoxide concentration, carbon dioxide concentration, PM2.5 concentration inside and outside the car, while the relationship between adjusting the air conditioning temperature and the characteristics of carbon monoxide concentration, carbon dioxide concentration, and PM2.5 concentration inside and outside the car is not significant. If feature isolation is not performed, the carbon monoxide concentration, carbon dioxide concentration, and PM2.5 concentration inside and outside the car will also be used as features to determine whether to adjust the temperature, which will cause noise interference to the model training. Therefore, this method groups the features that the model depends on according to the type of multi-task. The specific method is: add a corresponding feature isolation layer to the model architecture, isolate the input features that different tasks depend on, and each enters the model independently for training, so as to achieve the effect of non-interference and improve the accuracy of the model. For example Figure 4 The air conditioning switch feature group, the air conditioning internal and external circulation feature group, and the air conditioning temperature feature group in the image are used to isolate the input features that different tasks depend on.

[0053] Step 2.2: Build a feature group index dictionary. According to the feature grouping in step 2.1, set the feature grouping dictionary corresponding to the task.

[0054] Among them, the feature group index dictionary is (due to space limitations, not listed one by one):

[0055] {

[0056] "Air Conditioning Temperature":[

[0057] "Current setting temperature of air conditioner",

[0058] "Indoor humidity",

[0059] "Owner's age",

[0060] "Owner's gender",

[0061] "" ...

[0063] ],

[0064] "Air conditioning switch":[

[0065] "Inside temperature",

[0066] "Outside temperature",

[0067] "Remaining mileage",

[0068] "Window Status",

[0069] "Remaining mileage",

[0070] "" ...

[0072] ],

[0073] "Internal and external circulation":[

[0074] "PM2.5 concentration inside the car",

[0075] "PM2.5 concentration outside the car",

[0076] "CO concentration inside the car",

[0077] "Outside CO concentration",

[0078] "" ... ]

[0081] }

[0082] Step 2.3: embed the feature isolation layer. The feature isolation layer is embedded in the model structure in step 1. After obtaining the input feature vector of the model input layer, the feature isolation layer will automatically group all feature fields according to the feature group index dictionary set in step 2.2 and enter the corresponding task feature extraction layer.

[0083] After the feature isolation layer obtains the model input vector containing all feature fields, it divides the feature fields into three groups according to the feature group index dictionary. The air conditioner switch feature group will enter the air conditioner switch task in the multi-task model for learning and training. Similarly, each task learns independently without interfering with each other, reducing the generation of noise. The multi-task learning model framework embedded with the feature isolation layer is shown in Figure 1. Figure 4 shown.

[0084] The specific steps of step 3 above are as follows:

[0085] Step 3.1: Define the network model structure according to step 2 Figure 4 The model framework diagram in , uses the TensorFlow framework to define the network calculation graph structure, and sets the activation function, loss function, and evaluation index of the three tasks in turn, as shown in Table 1:

[0086] Table 1 Multi-task activation function, loss function, and evaluation index table

[0087]

[0088] Step 3.2: Save the model in PB format. After obtaining the model structure of step 3.1, divide the model data set into a training set and a validation set in a ratio of 7:3, perform model training and validation, obtain the model structure and weight after model training, and save it in PB format file;

[0089] Among them, PB files are a serialized format that can be used across languages ​​and operating environments. Almost any mainstream programming language can parse them, and other deep learning frameworks can also read, continue training, and migrate TensorFlow models.

[0090] Step 3.3: Convert the model PB format to ONNX file format. After obtaining the PB model format file in step 3.2, you need to load and use the TensorFlow framework ONNX module to convert the PB model file to an ONNX file. The vehicle side obtains the ONNX model file for real-time reasoning.

[0091] Among them, ONNX is an open format proposed by Microsoft and Facebook to represent deep learning models. The so-called openness means that ONNX defines a set of standard formats that are independent of the environment and platform to enhance the interactivity of various AI models. ONNX files not only store the weights of the neural network model, but also store the structural information of the model, the input and output of each layer in the network, and some other auxiliary information.

[0092] The innovations of the present invention are as follows:

[0093] 1. Deep learning neural network model. With the upgrading of the industry, the intelligence of automobile air-conditioning has become one of the research focuses in the automobile field. The number of vehicles of various brands is increasing day by day, and the amount of data of users adjusting the thermal comfort system of automobile air-conditioning is also becoming larger and larger. The present invention aims at the defects of the thermal comfort system of traditional automobile air-conditioning, and designs and develops a new deep learning neural network model based on the user big data of intelligent networked vehicles. The algorithm can automatically set the air-conditioning system switch, air-conditioning temperature, and internal and external circulation modes for users based on the environmental conditions inside and outside the car and the personalized analysis of users, which greatly reduces the operation complexity of users using car air-conditioning and improves the user experience of car air-conditioning.

[0094] 2. Multi-task model. The traditional automobile air-conditioning thermal comfort system algorithm constructs single-task models for multiple subsystems separately. For example, a temperature adjustment prediction model is constructed for the air-conditioning temperature adjustment system, and an air-conditioning switch prediction model is constructed for the air-conditioning system switch. If there are several subsystems, several single-task learning models need to be constructed. The present invention constructs a multi-task learning model. Compared with single-task learning, multi-task learning has the following advantages: 1. Multiple tasks share one model, and the amount of memory occupied is reduced; 2. Multiple tasks are forward calculated at one time to obtain results, and the reasoning speed is increased; 3. Related tasks can enhance the effectiveness of each other by sharing information and complementing each other.

[0095] 3. Feature isolation layer. Different tasks of the multi-task model are not particularly closely related to the model input features. To address this problem, the present invention constructs a feature group index dictionary based on the correlation between different tasks and feature fields, and embeds a feature isolation layer in the multi-task model architecture. After the model inputs the full feature field vector, the feature isolation layer can automatically divide the features into groups according to the feature group index dictionary to complete independent learning and training, and update the weights, thereby solving the problem of poor model performance caused by noise caused by the lack of close relationship between some features and tasks.

[0096] 4. Model conversion. Traditional algorithm models that use neural networks to control air conditioning thermal comfort are mostly deployed in the cloud. The vehicle side collects data in real time and uploads it to the cloud. The cloud model performs forward reasoning and outputs the results, which are then returned to the vehicle side. This method is too dependent on the network environment of the vehicle computer and has the defect of slow response speed. The present invention converts the multi-task learning model into ONNX format, which can be directly deployed on the vehicle side. After the vehicle side collects the data, it undergoes the same data preprocessing operation as the cloud side and enters the ONNX model file to complete forward reasoning. Within 100ms, the user's favorite air conditioning switch, temperature setting, and internal and external circulation mode parameter settings can be given in the current environment.

[0097] In summary, this method of intelligent control of automobile air conditioning based on multi-task learning neural network embedded with feature isolation layer, first, builds a multi-task learning model including 2 classifications of air conditioning switch, 17 classifications of air conditioning temperature and 2 classifications of internal and external circulation according to the various needs of air conditioning thermal comfort system, which solves the problem of multi-model resource consumption and slow reasoning time; then, builds feature isolation layer and feature group index dictionary, and embeds them into the multi-task learning model architecture (between the overall model input layer and each task feature extraction layer), completes feature isolation, and solves the problem of poor overall model effect caused by the lack of close connection between multi-task and model input features; finally, converts the verified model file into ONNX format and deploys it on the vehicle side for forward reasoning, no longer relying on the vehicle network, and can complete the parameter settings of air conditioning switch, air conditioning temperature, internal and external circulation in real time. This method realizes the deployment of a single model on the vehicle side, and pushes accurate, real-time and personalized air conditioning thermal comfort system operation parameters to car owners in different environments by model reasoning on the collected data, improving the car owner's experience of using air conditioning.

[0098] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the presence of other identical elements in the process, method, commodity or device including the element.

[0099] The above description is only an embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent substitution, improvement, etc. made within the spirit and principle of the present invention should be included in the scope of the claims of the present invention.

Claims

1. An intelligent control method for automobile air conditioning based on multi-task learning neural network, characterized in that: The following steps are involved: Obtain model input data, splice the model input data collected in real time into a growth vector and extract the feature fields of the model input data to obtain a model training data set, and build a multi-task learning neural network model for three tasks based on the air-conditioning operation to be controlled and the three tasks of model output; Based on expert experience, according to the correlation between the feature fields of the model input data and the three tasks, the feature fields of the model input data are divided into three feature data groups corresponding to the three tasks, and an index dictionary of the three tasks and the feature data groups is constructed, and a feature isolation layer is constructed according to the index dictionary, and the feature isolation layer is embedded between the input layer of the multi-task learning neural network model and the feature extraction layers of the three tasks to obtain a model calculation graph; Define the network structure according to the model calculation graph, select the corresponding activation function, loss function and model evaluation index according to the types of the three tasks, perform model training and verification to obtain a PB format model file, which contains the model structure and the weight coefficients obtained by training; The PB format model file is converted into an ONNX model format file and deployed on the vehicle side. After the real-time data collected by the vehicle side is preprocessed, it is input into the ONNX model to complete model reasoning, and finally the air conditioning switch, temperature adjustment and internal and external circulation parameter settings that the car owner likes to execute at the current moment are given.

2. The method for intelligently controlling automobile air conditioner based on multi-task learning neural network according to claim 1, characterized in that: The data types of the model input data include vehicle connection data, vehicle environment data, vehicle owner attributes and operation preference data.

3. The method for intelligently controlling automobile air conditioner based on multi-task learning neural network according to claim 1, characterized in that: The characteristic fields of the model input data include at least air conditioning circulation mode, air conditioning switch status, air conditioning temperature setting, air quality, PM2.5, indoor and outdoor temperature, indoor and outdoor humidity, vehicle cruising range, and window switch status.

4. The method for intelligently controlling automobile air conditioner based on multi-task learning neural network according to claim 1, characterized in that: The three tasks include a switch binary classification task, a temperature adjustment multi-classification task, and an internal and external circulation mode binary classification task; among them: the switch binary classification task predicts whether the air conditioner is turned on or off; the temperature adjustment multi-classification task adjusts the temperature value with a granularity of 1°C and a range of 16°C-32°C; the internal and external circulation mode binary classification task predicts whether the air conditioner turns on the internal circulation or the external circulation.

5. The method for intelligent control of automobile air conditioner based on multi-task learning neural network according to claim 1, characterized in that: The multi-task learning neural network model is a mapping of source data and target data, which is defined as follows: The source data is defined as X, which represents a sample set in a certain field; the target data is defined as Y, which represents a target set in a certain field; the mapping is defined as f, f:X→Y, which represents the conversion relationship from source to target; wherein the mapping includes a shared parameter layer and an independent parameter layer corresponding to each task, and the output end of the shared parameter layer is connected to the independent parameter layer corresponding to each task.

6. The method for intelligently controlling automobile air conditioner based on multi-task learning neural network according to claim 1, characterized in that: The feature fields of the model input data are divided into three types of feature data groups corresponding to the three tasks, specifically: The switch binary classification task corresponds to the first feature data group, which includes feature fields: vehicle interior temperature, vehicle exterior temperature, remaining mileage, window status, owner's gender, age, and owner's permanent location; The internal and external circulation mode binary classification task corresponds to the second feature data group, including feature fields: internal and external temperature, humidity, carbon monoxide concentration, carbon dioxide concentration, PM2.5 concentration, air conditioning cooling and heating status, and the owner's permanent residence; The temperature adjustment multi-classification task corresponds to the third feature data group, which includes fields: current setting temperature of the air conditioner, temperature inside and outside the car, humidity, gender and age of the owner, and permanent residence of the owner.

7. The method for intelligently controlling automobile air conditioner based on multi-task learning neural network according to claim 1, characterized in that: After the feature isolation layer is embedded between the input layer of the multi-task learning neural network model and the feature extraction layers of the three tasks, the feature isolation layer automatically groups all feature fields according to the index dictionary of the three tasks and feature data groups after obtaining the input feature vector of the input layer of the multi-task learning neural network model, and enters the feature extraction layer of the corresponding task.

8. The method for intelligently controlling automobile air conditioner based on multi-task learning neural network according to claim 1, characterized in that: The model calculation graph defines the network structure and specifically uses the TensorFlow framework to define the network calculation graph structure.

9. The method for intelligently controlling automobile air conditioner based on multi-task learning neural network according to claim 1, characterized in that: The model training and verification are performed to obtain a PB format model file, specifically: the model data set is divided into a training set and a verification set in a ratio of 7:3, model training and verification are performed, the model structure and weight after model training are obtained, and the model is saved as a PB format model file.

10. The method for intelligently controlling automobile air conditioner based on multi-task learning neural network according to claim 1, characterized in that: The converting of the PB format model file into the ONNX model format file is specifically: using the ONNX module of the TensorFlow framework to convert the PB format model file into the ONNX model format file.