Vehicle air conditioning adjustment method, second model training method, and electronic device

Abstract

The application provides a vehicle air conditioner adjusting method, a second model training method and an electronic device. The vehicle air conditioner adjusting method comprises the following steps: obtaining local data and sending the local data to a cloud connected with a target vehicle after the air conditioner of the target vehicle is turned on for a set time; the local data comprises local air conditioner data and local environment data; training a first model by using the local data when the local data on a vehicle machine end of the target vehicle meets a first training threshold; and / or receiving a newly trained second model sent by the cloud; the newly trained second model is configured to be trained by using cloud data when the cloud data meets a second training threshold; the cloud data comprises local data of a plurality of vehicles connected with the cloud; and adjusting a vehicle air conditioner parameter based on the newly trained first model and / or the newly trained second model. The application improves the accuracy of the model and the accuracy of air conditioner adjustment by training the model according to a plurality of local data.

Description

Technical Field

[0001] This application relates to the field of vehicle automatic control, and more specifically, to a vehicle air conditioning adjustment method, a second model training method, and electronic equipment. Background Technology

[0002] With the increasing demand for intelligent and comfortable vehicles, the vehicle's air conditioning system has become a key factor affecting the driving and riding experience. However, existing air conditioning systems have a low level of intelligence and lack the ability to deeply learn from user behavior. For example, the system cannot identify user habits and preferences (such as default temperature and fan speed) through historical data, nor can it dynamically optimize parameters based on real-time conditions (such as window opening / closing and sunlight intensity). Although some high-end models have introduced voice or touch interaction, they still rely on manual adjustments, failing to achieve automated and proactive comfort management, which significantly impacts the user experience. Summary of the Invention

[0003] In view of this, the purpose of this application is to provide a vehicle air conditioning adjustment method, a second model training method, and an electronic device, which can improve the accuracy of air conditioning adjustment and thus improve the user experience.

[0004] In a first aspect, embodiments of this application provide a vehicle air conditioning adjustment method, comprising: after the air conditioning of a target vehicle has been turned on for a set time, acquiring local data and sending the local data to a cloud connected to the target vehicle; wherein the local data includes local air conditioning data and local environmental data; training a first model using the local data when the local data on the vehicle's in-vehicle system meets a first training threshold; and / or receiving a newly trained second model sent from the cloud; wherein the newly trained second model is configured to be trained using cloud data when the cloud data meets a second training threshold; the cloud data includes the local data of multiple vehicles connected to the cloud; and adjusting vehicle air conditioning parameters based on the newly trained first model and / or the newly trained second model.

[0005] In the aforementioned implementation process, the first model is trained using local air conditioning data and local environmental data after the target vehicle's air conditioning is turned on. This approach, employing multi-dimensional data fusion and machine learning algorithms, can gradually improve user profiles and demand predictions, enhancing the accuracy and flexibility of the first model, and consequently, the accuracy and flexibility of air conditioning control. Furthermore, by setting up coordinated adjustments between the first and second models, the first model can quickly provide users with suitable air conditioning settings in the initial stages of vehicle use, meeting the needs of actual driving scenarios. With continuous user usage, the accuracy of predictions can be gradually improved through continuous data accumulation and model optimization. Moreover, with the continuous updating and addition of cloud data, a second model incorporating more data can be obtained, enabling more precise air conditioning control for users, thus improving not only the accuracy of air conditioning adjustments but also the user experience.

[0006] In one embodiment, training a first model using the local data when the local data on the vehicle's infotainment system meets a first training threshold includes: training the first model using the local data as training data when the local data reaches a first predetermined quantity and the acquisition time span of the local data meets a first predetermined time length; and / or training the first model using the local data as training data when the local data at any predetermined location on the target vehicle reaches a second predetermined quantity.

[0007] In the above implementation process, by setting a first predetermined quantity of local data and ensuring that the acquisition time span of the local data meets a first predetermined time length, using the local data as training data to train the first model can provide sufficient information to establish basic predictive capabilities while avoiding underfitting or excessive slowdown of startup time due to insufficient local data, thus improving the rationality of the local data used for training the first model. Furthermore, by setting a second predetermined quantity of local data from any predetermined location of the target vehicle and using it as training data to train the first model, excessively long acquisition times and large data storage volumes can avoid putting pressure on the vehicle's infotainment system, reducing memory pressure and improving the training efficiency of the first model.

[0008] In one embodiment, after training a first model using the local data on the vehicle's infotainment system when the local data meets a first training threshold, the method further includes: optimizing the first model based on positive samples and online gradient descent when it is determined that the training dataset of the first model has not been updated within a first set time period; wherein the training dataset is a set of local data used to train the first model, and the positive samples are system prediction parameters that the user has not manually changed within a second set time period.

[0009] In the above implementation process, optimizing the first model using positive samples allows it to quickly adapt to changes in user preferences, improving its accuracy. Furthermore, the optimization process is performed on the vehicle's infotainment system, resulting in low memory usage and short computation time, which reduces the space and memory required for the first model optimization and improves the efficiency of the infotainment system.

[0010] In one embodiment, adjusting vehicle air conditioning parameters based on a newly trained first model and / or a newly trained second model includes: obtaining a first update time of the newly trained first model and obtaining a second update time of the newly trained second model; wherein the newly trained first model refers to the first model most recently updated with respect to the current time, and the newly trained second model refers to the second model most recently updated with respect to the current time; determining the time interval between the first update time, the second update time, and the current time; and adjusting the vehicle air conditioning parameters using the newly trained first model and / or the newly trained second model with the smaller time interval.

[0011] In the above implementation process, by determining the air conditioning adjustment model based on the update time of the first model and the second model, the model used to adjust the air conditioning can always be the latest model, thereby improving the accuracy of air conditioning adjustment.

[0012] In one embodiment, after adjusting the vehicle air conditioning parameters based on the newly trained first model and / or the newly trained second model, the method further includes: acquiring user behavior data; wherein the user behavior data includes user speech and sound playback; and adjusting the vehicle air conditioning parameters according to the user behavior data.

[0013] In the above implementation process, by further adjusting the vehicle air conditioning parameters based on user behavior data, the vehicle air conditioning parameters can be further adjusted on the basis of the first model and / or the second model, so that the air conditioning parameters are more in line with user needs, thereby improving the accuracy of air conditioning adjustment and user experience.

[0014] Secondly, embodiments of this application provide a second model training method applied to the cloud. The method includes: acquiring local data sent by various vehicles connected to the cloud to obtain cloud data; wherein the cloud data includes local data of multiple vehicles connected to the cloud; training a second model using the cloud data when the cloud data meets a second training threshold; and sending the trained second model to the target vehicle; wherein the second model is applied to the vehicle air conditioning adjustment method in the first aspect or any possible implementation of the first aspect.

[0015] In the above implementation process, by training a second model in the cloud, and the cloud data used to train the second model comes from multiple vehicles connected to the cloud, the second model can learn the air conditioning adjustment methods of different users and different weather conditions. It can capture complex behavioral models with high accuracy, which can improve the accuracy of the second model and thus improve the flexibility and accuracy of air conditioning adjustment.

[0016] In one embodiment, training the second model using the cloud data when the cloud data meets the second training threshold includes: using the cloud data as training data to train the second model when the cloud data reaches a third predetermined quantity and the acquisition time span of the cloud data meets the second predetermined time length; and / or using the cloud data as training data to train the second model when the local data at any predetermined location of the target vehicle reaches a fourth predetermined quantity and the acquisition time span of the local data meets the third predetermined time length.

[0017] In the above implementation process, by setting the cloud data to a third predetermined quantity and ensuring the cloud data acquisition time span meets the second predetermined time length, using the cloud data as training data to train the second model can provide sufficient information to establish basic predictive capabilities while avoiding underfitting or excessively slowing down startup time due to insufficient cloud data, thus improving the rationality of the local data used for training the second model. Furthermore, by setting the local data from any predetermined location of the target vehicle to a fourth predetermined quantity, using the cloud data as training data to train the second model can avoid excessively long collection times and large data storage volumes, thus reducing cloud memory pressure and improving the training efficiency of the second model.

[0018] In one embodiment, after training the second model using the cloud data when the cloud data meets the second training threshold, the method further includes: calculating the accuracy of the second model using a set algorithm; sending the trained second model to the target vehicle includes: sending the trained second model to the target vehicle when the accuracy reaches the set threshold.

[0019] In the above implementation process, by setting the second model to be sent to the target vehicle only when the accuracy of the second model reaches a set threshold, the accuracy of the second model used for air conditioning adjustment can always meet the user's needs, thereby improving the accuracy of the second model.

[0020] Thirdly, embodiments of this application also provide a vehicle air conditioning adjustment device, comprising: a first acquisition module, configured to acquire local data after a set time has elapsed since the air conditioning of a target vehicle has been turned on, and to send the local data to a cloud connected to the target vehicle; wherein the local data includes local air conditioning data and local environmental data; a first training module, configured to train a first model using the local data when the local data on the vehicle's infotainment system meets a first training threshold; and / or a receiving module, configured to receive a newly trained second model sent from the cloud; wherein the newly trained second model is configured to be trained using cloud data when the cloud data meets a second training threshold; the cloud data includes the local data of multiple vehicles connected to the cloud; and a first adjustment module, configured to adjust vehicle air conditioning parameters based on the newly trained first model and / or the newly trained second model.

[0021] Fourthly, this application also provides a second model training device applied in the cloud. The device includes: a second acquisition module for acquiring local data sent by various vehicles connected to the cloud to obtain cloud data; wherein the cloud data includes local data of multiple vehicles connected to the cloud; a second training module for training a second model using the cloud data when the cloud data meets a second training threshold; and a sending module for sending the trained second model to the target vehicle; wherein the second model is applied to the vehicle air conditioning adjustment method described in the first aspect above, or any possible implementation of the first aspect.

[0022] Fifthly, embodiments of this application also provide an electronic device, including: a processor and a memory, wherein the memory stores machine-readable instructions executable by the processor, and when the electronic device is running, the machine-readable instructions are executed by the processor to perform the steps of the method described in the first aspect, or any possible implementation of the first aspect, the second aspect, or any possible implementation of the second aspect.

[0023] In a sixth aspect, embodiments of this application also provide a computer-readable storage medium storing a computer program that, when executed by a processor, performs the steps of the method described in the first aspect, or any possible implementation of the first aspect, or any possible implementation of the second aspect.

[0024] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, specific embodiments are described below in conjunction with the accompanying drawings. Attached Figure Description

[0025] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 A schematic diagram illustrating the interaction between the cloud and the vehicle-mounted system as provided in an embodiment of this application;

[0027] Figure 2 A block diagram illustrating an electronic device provided in an embodiment of this application;

[0028] Figure 3 A flowchart of a vehicle air conditioning adjustment method provided in an embodiment of this application;

[0029] Figure 4 A flowchart of the second model training method provided in the embodiments of this application;

[0030] Figure 5 This is a schematic diagram of the functional modules of the vehicle air conditioning control device provided in the embodiments of this application;

[0031] Figure 6 A schematic diagram of the functional modules of the second model training device provided in the embodiments of this application. Detailed Implementation

[0032] The technical solutions in the embodiments of this application will now be described with reference to the accompanying drawings.

[0033] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0034] With the booming development of the automotive industry and increasingly stringent requirements for driving and riding experiences, the performance and functions of vehicle air conditioning systems, as key components for regulating the comfort of the in-vehicle environment, have attracted much attention.

[0035] However, through long-term research, the inventors of this application have discovered that many existing systems employ traditional monitoring and control methods, primarily relying on manual intervention and basic sensors. While these methods can provide basic environmental monitoring, they suffer from drawbacks such as slow response times, low data processing efficiency, and the inability to provide real-time early warnings. Furthermore, existing technologies lack sufficient capabilities in data integration and intelligent analysis, making it difficult to effectively respond to emergencies in complex environments. They also struggle to provide optimal comfort adjustments based on users' personalized needs and external environmental characteristics. Simultaneously, existing systems lack the ability to learn user habits and cannot dynamically optimize air conditioning performance to improve user satisfaction.

[0036] In view of this, this application proposes a vehicle air conditioning adjustment method. By setting a first predetermined quantity of local data and ensuring the acquisition time span of the local data meets a first predetermined time length, the local data is used as training data to train a first model. This provides sufficient information to establish basic predictive capabilities while avoiding underfitting or excessively slowing down startup time due to insufficient local data, thus improving the rationality of the local data used for training the first model. Furthermore, by setting a second predetermined quantity of local data from any predetermined location of the target vehicle and using this local data as training data to train the first model, excessively long acquisition times and large data storage volumes can avoid putting pressure on the vehicle's infotainment system, reducing memory pressure and improving the training efficiency of the first model.

[0037] Therefore, an intelligent vehicle air conditioning control system is proposed, which provides more precise air conditioning control services through user behavior and environmental data analysis to improve the driving experience.

[0038] To facilitate understanding of this embodiment, the operating environment of a vehicle air conditioning adjustment method disclosed in this application will be described in detail first.

[0039] like Figure 1 The diagram illustrates the interaction between the cloud and the vehicle-mounted system according to an embodiment of this application. The cloud connects to one or more vehicle-mounted systems via a network for data communication or interaction. The cloud can be a web server, database server, etc. The vehicle-mounted system can be an in-vehicle computer, in-vehicle navigation system, personal digital assistant, etc.

[0040] The cloud connects to one or more vehicle-mounted systems via a network, enabling it to acquire local data from multiple vehicle-mounted systems and generate cloud-based data.

[0041] The cloud platform can train a second model using cloud data. This second model can then be used to adjust the air conditioning in the vehicle's infotainment system connected to the cloud. Once the second model is trained, the cloud platform can send the trained model to the connected vehicle's infotainment system.

[0042] The vehicle's infotainment system can also use its own local data to train a first model, which can then be used to adjust the air conditioning in the vehicle's infotainment system.

[0043] It should be understood that after receiving the second model, the vehicle's infotainment system can determine the model to be used to control the air conditioning based on the update time of the first and second models.

[0044] Optionally, the air conditioner can be adjusted independently using the first model, independently using the second model, or by using the first and second models in combination.

[0045] The aforementioned vehicle-mounted infotainment system is equipped with a software system (such as Android or iOS) that supports DCN model inference. The CAN bus of this infotainment system provides a data transmission latency of <10ms.

[0046] The second model is deployed in the cloud using TensorFlow Lite, supporting real-time inference on the vehicle's infotainment system. Additionally, cloud training is based on TensorFlow / PyTorch, and the API supports model deployment and data synchronization.

[0047] To facilitate understanding of this embodiment, the electronic device executing the vehicle air conditioning adjustment method and / or the second model training method disclosed in this application embodiment will be described in detail below. The electronic device may include two devices, one located in the cloud and the other located in the vehicle's infotainment system. The electronic device located in the cloud is used to execute the second model training method disclosed in the application embodiment, and the electronic device located in the vehicle's infotainment system is used to execute the vehicle air conditioning adjustment method disclosed in the application embodiment.

[0048] Optionally, the electronic device set in the cloud and the electronic device set in the vehicle can be the same or different, and the settings of the electronic device can be selected according to the actual situation.

[0049] like Figure 2 The diagram shown is a block illustration of an electronic device. The electronic device 100 may include a memory 111 and a processor 113. Those skilled in the art will understand that... Figure 1 The structure shown is for illustrative purposes only and does not limit the structure of the electronic device 100. For example, the electronic device 100 may also include components that are more... Figure 1 The more or fewer components shown, or having the same Figure 1 The different configurations shown.

[0050] The aforementioned memory 111 and processor 113 are electrically connected to each other directly or indirectly to enable data transmission or interaction. For example, these components can be electrically connected to each other via one or more communication buses or signal lines. The aforementioned processor 113 is used to execute executable modules stored in the memory.

[0051] The memory 111 can be, but is not limited to, Random Access Memory (RAM), Read Only Memory (ROM), Programmable Read-Only Memory (PROM), Erasable Programmable Read-Only Memory (EPROM), Electrically Erasable Programmable Read-Only Memory (EEPROM), etc. The memory 111 stores programs, and the processor 113 executes these programs upon receiving execution instructions. The methods executed by the electronic device 100 as defined in any embodiment of this application can be applied to or implemented by the processor 113.

[0052] The aforementioned processor 113 may be an integrated circuit chip with signal processing capabilities. The processor 113 may be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it may also be a digital signal processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor may be a microprocessor or any conventional processor.

[0053] The electronic device 100 in this embodiment can be used to execute various steps in the various methods provided in the embodiments of this application. The implementation process of the vehicle air conditioning adjustment method is described in detail below through several embodiments.

[0054] Please see Figure 3 This is a flowchart of a vehicle air conditioning adjustment method provided in an embodiment of this application. The following will describe... Figure 3 The specific process shown will be explained in detail.

[0055] Step S201: After the target vehicle's air conditioning is turned on for a set time, acquire local data and send the local data to the cloud connected to the target vehicle.

[0056] Local data includes local air conditioning data and local environmental data. Local air conditioning data refers to air conditioning operating status data (e.g., temperature setting, fan speed, mode on / off, etc.), while local environmental data refers to data about the environment inside and outside the vehicle (e.g., temperature, humidity, weather, etc.). The specific types of local air conditioning data and local environmental data can be selected according to actual needs.

[0057] It should be understood that after the target vehicle's air conditioning is turned on for a set time, the system can determine whether the vehicle is in a "vehicle started" state and whether the air conditioning control function is available based on the air conditioning's status. Only when the vehicle is in a "vehicle started" state and the air conditioning control function is available will local data be acquired through the data acquisition device.

[0058] The local data here can be air conditioning usage data during vehicle use, such as temperature settings, fan speed adjustments, and on / off status. In other words, it's the air conditioning usage data continuously collected from the user while the vehicle is in use.

[0059] In one embodiment, when the air conditioning usage data is manually adjusted by the user, the adjusted air conditioning usage data is recorded, and the local air conditioning data is updated based on the air conditioning usage data.

[0060] Optionally, local data can be acquired in real time, at set time intervals, or when changes are detected in the local data. The method of acquiring local data can be selected according to the actual situation.

[0061] The aforementioned local data may also include vehicle characteristic data, user behavior data, and user profile data. Vehicle characteristic data may include navigation destination, current location, total trip mileage, total trip duration, and window open / closed status. User behavior data may include user speech and audio playback. User profile data may include user temperature preferences, frequency of air conditioning use, preferred fan speed, and preferred air volume. This local data can be adjusted according to actual conditions.

[0062] It should be understood that after obtaining local data on the vehicle's infotainment system, the local data is stored on the vehicle's infotainment system and simultaneously sent to the cloud.

[0063] This local data can be stored in isolation according to user accounts to ensure privacy and security. Additionally, this local data can be encrypted and stored on the vehicle's infotainment system before being sent to the cloud.

[0064] The local storage here can be set with a certain storage limit. When the storage limit is reached, a corresponding strategy can be adopted to replace the old local data with new local data.

[0065] In one embodiment, after step S201, the method further includes preprocessing the local data. For example, this may involve cleaning, normalization, and outlier removal. The preprocessing method can be selected based on the specific circumstances.

[0066] It should be understood that in the event of a network failure, such as a failure to upload local data due to reasons related to the vehicle's network, the local data can be temporarily stored on the vehicle's device until the network is restored, at which point the local data will be synchronized to the cloud.

[0067] Once the network is restored, local data is uploaded via HTTPS, which can trigger the second model optimization. If the upload fails, it can be retried at regular intervals, using MD5 verification to ensure data integrity. An automatic synchronization mechanism ensures uninterrupted service during network failures, improving the synchronization success rate.

[0068] After acquiring the aforementioned local data, it can be divided into training and testing sets according to a certain ratio. The training set uses 5-fold cross-validation to evaluate model performance. Furthermore, to address the class imbalance problem, oversampling is employed for categories such as air conditioner usage time periods to improve the model's predictive ability for smaller sample categories.

[0069] The 5-fold cross-validation method divides the training dataset into five folds. Four folds are used to train the model each time, and one fold is used as the validation set. This process is repeated five times, and the average performance metric is calculated. 5-fold cross-validation is used to evaluate the stability of the model across different data distributions. It is suitable for both the first and second models, ensuring that the predicted temperature error is <2℃ and the accuracy is ≥80%.

[0070] Five-fold cross-validation is used to evaluate model performance, ensure the model's generalization ability across different subsets of data, prevent overfitting, and provide stable performance metrics (such as accuracy and mean squared error). The specific implementation steps are illustrated in the following example:

[0071] The training datasets consist of cloud-based or local datasets. The first model has 100 data points (approximately 7 days), and the second model has 50,000 data points (300 users, 1 month). The cloud-based or local datasets have been preprocessed (e.g., Min-Max normalization, one-thermal / multi-thermal encoding) and contain features (e.g., temperature, driving distance, POI) and labels (air conditioning set temperature). The training datasets are then divided into a training set (80%) and a test set (20%) in an 8:2 ratio. For example, for a large model with 50,000 data points, the training set has 40,000 data points, and the test set has 10,000. The training set is further divided into 5 folds, each equally divided (e.g., 40,000 data points divided into 5 parts, each with 8,000 data points). This further splitting uses a random seed (fixed at 42) to ensure repeatability, and stratified sampling maintains a consistent label distribution (e.g., uniform temperature distribution). The training set is divided into 5 parts, numbered fold 1 to fold 5. In the first fold: fold 1 is the validation set, and folds 2–5 are the training set. The model is trained and evaluated on fold 1. Second iteration: fold 2 is the validation set, folds 1, 3–5 are the training set, and so on. A total of 5 training and validation iterations are performed.

[0072] For the first model: a two-layer fully connected network (64 and 32 neurons, ReLU activation, Sigmoid output), Adam optimizer, learning rate 0.001, batch size = 256, training for 30 epochs.

[0073] Second model: Deep & Cross Network (512, 256, 128 neurons, ReLU activation, Sigmoid output), Adam optimizer, learning rate 0.001, training for 50 epochs, early stopping strategy (verification showed no decrease in loss after 5 epochs).

[0074] Understandably, oversampling can also improve the prediction ability of small sample classes. Class imbalance refers to the fact that the sample size of some classes (such as air conditioner use at night) is much smaller than that of other classes (such as during the day), causing the model to favor the majority class and make inaccurate predictions of the minority class.

[0075] Oversampling increases the amount of minority class data and improves model performance by interpolating synthetic samples in the feature space of minority class samples. This oversampling enhances the prediction ability of small class samples to address class imbalance problems (e.g., insufficient minority class samples during air conditioner usage periods). By synthesizing minority class samples, it improves the model's prediction ability for small class samples. The specific implementation steps are as follows:

[0076] For categorical features (e.g., air conditioner usage time periods: {morning, daytime, evening, late night}), count the number of samples for each category. Example: In 40,000 training data points, the time period distribution is as follows: Morning: 12,000 (30%), Daytime: 20,000 (50%), Evening: 7,000 (17.5%), Late Night: 1,000 (2.5%). Set an imbalance threshold: if a category's percentage is <10% (e.g., late night), mark it as a minority class.

[0077] The oversampling algorithm works as follows: For each minority class sample, find its k nearest neighbors (default k=5, using Euclidean distance). Randomly select a neighbor sample and interpolate between this sample and its neighbors to generate a synthetic sample. The calculation formula for this oversampling can be:

[0078] x new =x i +λ(x n -x i );

[0079] Where, x new For the synthetic sample, x i For the i-th sample, x n Let λ be a neighbor sample of the i-th sample, and let λ be a random number, where λ ∈ [0, 1].

[0080] The oversampling implementation process can be as follows: Ensure that features are standardized (e.g., continuous features Min-Max normalization, categorical features one-hot encoding), apply oversampling only to the training set, and maintain the original distribution of the test set. For minority classes (e.g., late night), calculate the target number of samples so that its proportion is close to that of the majority class (e.g., daytime, 50%). For example, increase the number of late night samples from 1000 to 8000 (approximately 20%). The KNN algorithm can be used to calculate the nearest neighbors of minority class samples in the feature space (normalized feature vectors). The feature space dimension is approximately 20 (10 high-weight features + interaction features). For each late night sample, randomly select one neighbor sample to generate a synthetic sample.

[0081] The labels of the synthesized samples here are inherited from the original samples. For example, "late at night".

[0082] After identifying the synthetic samples, multiple synthetic late-night samples are added to the training set, increasing the total number of samples in the training set. This also helps to balance the class distribution within the training set.

[0083] The first model and / or the second model can be trained using the balanced training set.

[0084] In one embodiment, the local data distribution can be analyzed periodically, and if a new minority class emerges (e.g., a new POI category), oversampling is rerun. If synthetic samples lead to overfitting, the proportion of synthetic samples is reduced.

[0085] The above oversampling can be performed in the cloud (e.g., single processing time < 1 minute). The synthesized sample can be stored as a sparse matrix, occupying < 1MB / 1000 records. Additionally, if there are < 5 minority class samples, SMOTE is skipped, and rule-based filling (e.g., copying the most recent sample) is used.

[0086] Step S202: If the local data on the vehicle's infotainment system meets the first training threshold, train the first model using the local data; and / or receive a newly trained second model sent from the cloud.

[0087] The newly trained second model is configured to be trained using cloud data, provided that the cloud data meets the second training threshold; the cloud data includes local data from multiple vehicles connected to the cloud.

[0088] It should be understood that, provided the local data meets the cold start training threshold, machine learning algorithms (such as decision trees, regression models, etc.) can be used to train the local data to obtain the first model.

[0089] The first model here can be a cold start model, which mainly focuses on adapting to short-term user behavior. By learning from local data over a period of time, a preliminary air conditioning control model can be built.

[0090] The aforementioned first training threshold is a threshold used to represent the acquisition status of local data. For example, the first training threshold can refer to the local data reaching a set number of data entries, the local data meeting a set time span, or the local data reaching a set number of data entries and meeting a set time span, etc. The first training threshold can be set according to the actual situation.

[0091] The local data here can be updated in real time after new data is generated on the vehicle's infotainment system. This replaces the old local data with the new local data, ensuring that the local data on the vehicle's infotainment system always remains within a reasonable threshold range.

[0092] In one embodiment, prior to step S202, the method further includes: analyzing the features in the local data by setting an algorithm to determine the order of importance of each local data point for air conditioning temperature prediction.

[0093] Optionally, the algorithm can be Lasso regression, random forest, etc., and the algorithm can be selected according to the actual situation.

[0094] Lasso regression is a linear regression model that uses L1 regularization to select and reduce features in the objective function. The coefficients in Lasso regression represent the weight of each feature in the model. A larger absolute value of the coefficient indicates that the feature (i.e., local data) is more important to the target variable; a zero coefficient indicates that the feature is completely ignored by the model. By performing Lasso regression analysis on the air conditioning temperatures of the front and rear rows separately, and averaging the Lasso coefficients of the two models, a comprehensive feature importance map can be generated.

[0095] Random forest is an ensemble learning method based on decision trees. It improves the performance and stability of the model by constructing multiple decision trees and averaging their predictions. The feature importance in random forest represents the contribution of a feature (i.e., local data) to the reduction of impurity at the split points across all trees. A higher feature importance value indicates that the feature is more important to the model's predictions. By using a random forest model to analyze the feature importance of the front and back row air conditioner temperatures separately, a feature importance ranking chart can be generated.

[0096] As shown in Table 1, the table presents the absolute value of the average coefficient and the importance value for each local data point.

[0097] Table 1:

[0098]

[0099] It should be understood that after determining the importance order of each local data point, the first model can be trained based on the importance of each local data point. For example, appropriate weights can be assigned to local data points of different importance levels, and then the first model can be trained.

[0100] Similarly, cloud data can be analyzed by setting algorithms to determine the order of importance of each piece of cloud data for air conditioning temperature prediction, and the second model can be trained according to the importance of the cloud data.

[0101] Understandably, if a second model is trained in the cloud, the cloud can transmit the trained second model to the vehicle's infotainment system for storage. Simultaneously, if a second model already exists on the vehicle's infotainment system, the new second model can be used to update the old second model stored there.

[0102] Step S203: Adjust the vehicle air conditioning parameters based on the newly trained first model and / or the newly trained second model.

[0103] The model used to adjust the vehicle air conditioning parameters can be either a newly trained first model or a newly trained second model, or both can be adjusted simultaneously.

[0104] It should be understood that if one of the models is used to adjust the vehicle air conditioning parameters, the model with the closest update time to the current time between the newly trained first model and the newly trained second model can be selected for adjusting the air conditioning parameters.

[0105] In the aforementioned implementation process, the first model is trained using local air conditioning data and local environmental data after the target vehicle's air conditioning is turned on. This approach, employing multi-dimensional data fusion and machine learning algorithms, can gradually improve user profiles and demand predictions, enhancing the accuracy and flexibility of the first model, and consequently, the accuracy and flexibility of air conditioning control. Furthermore, by setting up coordinated adjustments between the first and second models, the first model can quickly provide users with suitable air conditioning settings in the initial stages of vehicle use, meeting the needs of actual driving scenarios. With continuous user usage, the accuracy of predictions can be gradually improved through continuous data accumulation and model optimization. Moreover, with the continuous updating and addition of cloud data, a second model incorporating more data can be obtained, enabling more precise air conditioning control for users, thus improving not only the accuracy of air conditioning adjustments but also the user experience.

[0106] In one possible implementation, if the local data on the vehicle's infotainment system meets a first training threshold, the first model is trained using the local data, including: training the first model with the local data as training data when the local data reaches a first set quantity and the acquisition time span of the local data meets a first set time length; and / or training the first model with the local data as training data when the local data at any set location of the target vehicle reaches a second set quantity.

[0107] The first set quantity here is used to constrain the number of local data entries. For example, 50, 100, 200, etc. This first set quantity can be selected according to the actual situation.

[0108] The aforementioned first set time length refers to the time span for continuously acquiring local data. For example, 3 days, 7 days, 10 days, 30 days, etc., and this first set time length can be selected according to the actual situation.

[0109] For example, if the first training threshold is 100 pieces of local data accumulated, meeting the 7-day time span condition, then the system automatically starts training the first model when it determines that the local data has accumulated to 100 pieces and meets the 7-day time span condition. That is, the first model is trained using these 100 accumulated pieces of local data that meet the 7-day time span condition.

[0110] The second set quantity here is used to constrain the number of local data entries at a given location. For example, 50, 100, or 200 entries can be selected based on the actual situation.

[0111] Optionally, the first set quantity can be the same as or different from the second set quantity, and the setting method of the first set quantity and the second set quantity can be selected according to the actual situation.

[0112] Understandably, multiple locations within a car (e.g., driver's seat, front passenger seat, rear left seat, rear right seat) may generate corresponding local data. Therefore, during the training of the first model, training can begin when the data from any of these multiple locations accumulates to a second training threshold.

[0113] The first and second training thresholds need to be set to provide enough information to establish basic prediction capabilities, while not being too low to cause underfitting or too high to slow down startup time.

[0114] Since the first model needs to be started as soon as possible, the amount of data corresponding to the first training threshold (e.g., 100 data points) can be collected in a short time to ensure a fast model deployment. This also provides a reference for subsequent dynamic data updates, gradually improving the model's performance.

[0115] Similarly, the amount of data corresponding to the second training threshold can be collected in a short time, ensuring a fast model deployment speed. This also provides a reference for subsequent dynamic data updates, gradually improving the model's performance.

[0116] In one embodiment, when new local data is generated on the vehicle's infotainment system, the first model can be updated in real time based on the new local data, so that the first model remains up-to-date.

[0117] For example, the real-time update of the first model is achieved through the following mechanism: Every 5 minutes, air conditioning usage data (e.g., temperature setting, airflow, mode), environmental data (e.g., in-vehicle and out-of-vehicle temperature, humidity, geographical location), and vehicle status (e.g., vehicle speed, battery status) are collected, and these data undergo preprocessing such as cleaning, normalization, and feature extraction. An online gradient descent algorithm is used to incrementally train the first model in real time on a local lightweight neural network (e.g., two fully connected layers, 64 and 32 neurons, with <100,000 parameters), updating the model parameters to allow the first model to adapt to the latest user behavior. The initial learning rate can be 0.001, combined with L2 regularization and Dropout to prevent overfitting.

[0118] When the network fails, local data can be encrypted with AES-256 and stored in local flash memory, managed using a priority queue (to retain the latest data). Once the network is restored, it will be automatically uploaded to the cloud via HTTPS protocol, and a verification mechanism can be used to ensure data integrity.

[0119] In the above implementation process, by setting a first predetermined quantity of local data and ensuring that the acquisition time span of the local data meets a first predetermined time length, using the local data as training data to train the first model can provide sufficient information to establish basic predictive capabilities while avoiding underfitting or excessive slowdown of startup time due to insufficient local data, thus improving the rationality of the local data used for training the first model. Furthermore, by setting a second predetermined quantity of local data from any predetermined location of the target vehicle and using it as training data to train the first model, excessively long acquisition times and large data storage volumes can avoid putting pressure on the vehicle's infotainment system, reducing memory pressure and improving the training efficiency of the first model.

[0120] In one possible implementation, after training a first model using local data provided that the local data on the vehicle's infotainment system meets a first training threshold, the method further includes: optimizing the first model based on positive samples and online gradient descent if the training dataset of the first model has not been updated within a first set time period.

[0121] The training dataset is a collection of local data used to train the first model.

[0122] It should be understood that when the training dataset of the first model has not been updated for a long period of time, the first model can be adjusted and optimized using positive samples, thereby improving the accuracy of the first model.

[0123] For example, when the training data set of the first model has not been updated for more than 3 months, causing the reference value of the old training dataset to decrease, it can be determined that the user's usage preferences may have changed. At this time, the first model can collect positive sample data from users and adjust and optimize the first model based on these positive samples to improve the accuracy of the first model.

[0124] Optimization when the training dataset is outdated: If the training dataset for the first model has not been updated for a set period of time, causing predictions based on the old training dataset to fail to reflect current user preferences, an optimization process can be initiated to ensure the accuracy and adaptability of the first model. The specific implementation steps are as follows:

[0125] Lag detection mechanism: The last update time of the local dataset is recorded by timestamp. Each time local data is collected, the latest timestamp of the local data is checked to see if it is more than a first set time period away from the current time.

[0126] If the first set time period is exceeded, the training dataset is marked as "lagging," triggering the optimization process. The lag determination can include the following conditions: the time span of local data in the training dataset (the time difference between the earliest and latest local data) exceeds the first set time period; the user's air conditioning usage frequency is below average within the most recent third set time period (e.g., fewer than 5 trips, indicating a possible change in user behavior), etc. These lag determination conditions can be selected based on the actual situation.

[0127] The determination of the training dataset can be achieved as follows: A metadata table is maintained on the vehicle's infotainment system to record the update time of the training dataset and local data statistics (e.g., the quantity of local data, time span, etc.). An embedded database (e.g., SQLite) is then used to store the metadata, reducing query time complexity and ensuring real-time performance.

[0128] The positive samples mentioned above are system prediction parameters that the user did not manually change within the second set time period.

[0129] The second time period can be set according to the user's actual situation. For example, the second time period can be 3 minutes, 5 minutes, 10 minutes, 12 minutes, etc., and can be selected according to the actual situation.

[0130] It should be understood that if the user does not manually change the system's predicted temperature setting within the second set time period, it indicates that the predicted temperature meets the user's needs. Such samples can then be labeled as "true positives" (i.e., positive samples) and used to optimize the first model.

[0131] In one embodiment, positive samples can be obtained by starting a timer after each air conditioning adjustment to monitor whether the user manually adjusts the temperature within a second set time period. If no adjustment is made, the current sample (e.g., air conditioning settings, environmental data, vehicle status, etc.) is marked as a positive sample and stored in a local database.

[0132] The system also allows setting the positive sample collection period and the target number of samples. If the number of positive samples collected within the collection period is less than the target number, the system can automatically extend the collection period or reduce the target number of samples. The collection period and target number can be selected based on the actual situation.

[0133] For example, the positive sample collection period is 7 days, and the target number of positive samples is at least 20. If fewer than 20 positive samples are collected within 7 days, the collection period can be extended to 14 days, or the target number of samples can be reduced to 10.

[0134] The positive samples mentioned above can be stored in JSON format. This can include feature vectors (such as temperature, airflow, and interior / exterior vehicle temperatures) and labels (predicted temperature).

[0135] Optionally, a local storage limit can be set. If the number of positive samples collected exceeds the local storage limit, an in-first-out (FIFO) strategy will be used to replace the old samples to prevent storage overflow.

[0136] For example, if the local storage limit is 50 positive samples, then if more than 50 positive samples are collected, the first positive sample stored will be removed from the local storage and the latest positive sample will be stored.

[0137] The first model optimization described above can also be triggered by certain conditions. For example, the first model optimization can be initiated when the number of positive samples reaches a sufficient level or when the lag time exceeds a certain period.

[0138] For example, the first model optimization can be initiated when 20 positive samples are collected or the lag time exceeds 90 days.

[0139] Optionally, positive samples can be preprocessed before optimizing the first model.

[0140] The preprocessing procedure for positive samples here can be to normalize continuous features to [0,1]. The normalization formula can be:

[0141] Where, x min To minimize the positive samples, x max To maximize the number of positive samples, x is the current positive sample.

[0142] When optimizing the first model, positive samples can be used first, supplemented by regular local data collected during the collection period.

[0143] In one embodiment, the first model can be optimized in the following way:

[0144] Based on positive samples, the first model was fine-tuned using online gradient descent. The acquisition period model was a lightweight fully connected neural network.

[0145] The loss function for the first model can be:

[0146] in, To predict temperature, yi is the actual temperature of the positive sample, and n is the number of positive samples.

[0147] Add L2 regularization to prevent overfitting:

[0148] Where λ=0.01, ω j These are the weights for the first model.

[0149] The optimization steps for the first model here can be as follows: Load the model parameters of the current first model, input positive samples into the first model, and use the Adam optimizer to update the weights based on the gradients of the positive samples. Then, perform multiple rounds of fine-tuning, using an early stopping strategy (stop if the validation loss does not decrease for several consecutive rounds (e.g., 3 rounds)). Finally, use the positive samples from the most recent rounds (e.g., 5 rounds) as the validation set to evaluate whether the loss function of the optimized model is lower than that before optimization.

[0150] The learning rate of the Adam optimizer can be set to 0.0005 to ensure stability.

[0151] The aforementioned fine-tuning can be performed on the edge computing unit at the vehicle's infotainment system. This fine-tuning requires less than 50MB of memory and has an inference time of less than 0.5 seconds.

[0152] If there are insufficient positive samples, pseudo-labels are generated by combining environmental data (e.g., temperature difference between inside and outside the vehicle), and training data is supplemented based on set rules (e.g., reduce by 1℃ when the temperature difference is >10℃).

[0153] Understandably, when optimizing the first model, its parameters can also be optimized. The optimization of the parameters of the first model can be achieved in the following ways:

[0154] The initial parameters of the first model are trained based on a predetermined number of data points (e.g., 100). These initial parameters may include a weight matrix and a bias vector. Before optimization, a copy of the current parameters can be saved as a rollback point to prevent optimization failure.

[0155] Then, positive samples are preprocessed, and sparse vectors are generated by one-hot encoding based on class features. Feature selection can be based on importance ranking using Lasso regression and / or random forest, prioritizing the retention of high-weight features.

[0156] Finally, the temperature is predicted by forward propagation calculation.

[0157] The formula for forward propagation here can be:

[0158]

[0159] The formula for calculating gradients and updating weights and biases using the Adam optimizer can be:

[0160]

[0161] The formula for weighting positive samples can be:

[0162]

[0163] Where W1 and W2 are the weight matrices of the first and second layers, respectively; b1 and b2 are the bias terms of the first and second layers, respectively; X is the input data; ReLU is the activation function (corrected linear unit); Sigmoid is the activation function (sigmoid function); W is the weight matrix; b is the bias vector; and η is the learning rate. Let W be the partial derivative of the loss function with respect to the weights W. Let be the partial derivative of the loss function with respect to the bias b, 2.0 be the loss weight for positive samples, and 1.0 be the loss weight for normal samples. The squared error between the predicted and actual values ​​for positive samples. This represents the squared error between the predicted and actual values ​​for a standard sample.

[0164] In one embodiment, after optimizing the first model, the optimized first model needs to be validated. The validation method for the optimized first model can be: calculating the mean squared error using a validation set and comparing the model performance before and after optimization. Specifically, if the mean squared error decreases by >10% after optimization (e.g., from 0.05 to 0.045), the parameters are saved; otherwise, the parameters are rolled back to those before optimization.

[0165] Additionally, if user behavior changes significantly (e.g., positive samples show a temperature preference changing from 22℃ to 26℃), the system increases the learning rate (e.g., from 0.0005 to 0.001) to accelerate adaptation. If the optimized model's prediction error exceeds 2℃, cloud intervention is triggered, positive samples are uploaded to the cloud, and the first model is retrained using data from multiple vehicles.

[0166] The aforementioned positive samples are stored together with regular data in a local database and are encrypted. Positive samples can be marked as "high priority" to ensure they are uploaded first and / or used to train the first model.

[0167] Understandably, since the first model is trained on the vehicle's infotainment system, the training time for the first model can be set when the vehicle is not in use, or it can be set to be trained on the back end, so as to avoid the training of the first model affecting the normal adjustment of the air conditioner.

[0168] In the above implementation process, optimizing the first model using positive samples allows it to quickly adapt to changes in user preferences, improving its accuracy. Furthermore, the optimization process is performed on the vehicle's infotainment system, resulting in low memory usage and short computation time, which reduces the space and memory required for the first model optimization and improves the efficiency of the infotainment system.

[0169] In one possible implementation, step S203 includes: obtaining the first update time of the newly trained first model and the second update time of the newly trained second model; determining the time interval between the first update time, the second update time and the current time; and adjusting the vehicle air conditioning parameters using the newly trained first model and / or the newly trained second model with smaller time intervals.

[0170] Here, the newly trained first model refers to the first model that was most recently updated at the current time, and the newly trained second model refers to the second model that was most recently updated at the current time.

[0171] The first update time here refers to the most recent update time of the newly trained first model, and the second update time refers to the most recent update time of the newly trained second model.

[0172] It should be understood that, in order to ensure the accuracy and real-time performance of air conditioning adjustments, the latest updated first model and / or second model is usually selected for adjustment. Therefore, when determining the model for adjusting the air conditioning, it can be based on the update time of the first and second models.

[0173] When a newly trained first model or a newly trained second model with a smaller time interval is determined, it is equivalent to determining the model most recently updated from the current time. Adjusting the air conditioner based on the most recently updated model can improve the accuracy of the air conditioner adjustment.

[0174] In one embodiment, step S203 may further include: after the newly trained second model has been trained and the accuracy of the newly trained second model has reached an accuracy threshold, switching to the newly trained second model and adjusting the air conditioner through the newly trained second model.

[0175] For example, once the second model has been trained and its accuracy is ≥80%, the cloud can send the model parameters of the newly trained second model or the compressed model of the newly trained second model to the vehicle's infotainment system via HTTPS protocol, so that the air conditioning on the vehicle's infotainment system can be adjusted by the newly trained second model.

[0176] Before switching between the newly trained first model and the newly trained second model, the vehicle-mounted system can run a switching test: simultaneously using the first and second models to predict multiple trip data points, comparing the mean squared error (MSE) with the user-manually adjusted frequency. If the MSE of the second model is greater than the set MSE value (e.g., 10%) and the adjustment frequency is less than the set adjustment frequency value (e.g., 20%), the switching can be confirmed.

[0177] In another embodiment, step S203 may further include: if the accuracy of the newly trained second model decreases, or if the newly trained second model does not have an effective prediction structure, switching to the newly trained first model and adjusting the air conditioner through the newly trained first model.

[0178] For example, if the prediction error of the second model is greater than 2°C or the user adjustment frequency is greater than 30% (multiple consecutive trips), then switch back to the newly trained first model and adjust the air conditioner through the newly trained first model to ensure stability.

[0179] In this process, if the accuracy of the second trained model decreases, or if the newly trained second model does not have an effective predictive structure, feedback can be sent to the cloud. After analyzing the reasons, the cloud can retrain the second model.

[0180] The switching between the first and second models described above is merely illustrative, and the switching between the first and second models can be selected according to the actual situation.

[0181] The cloud can use the latest local data (including positive samples) uploaded by the connected vehicles to form new cloud data, fine-tune the second model, and then re-issue the updated second model to the vehicle terminal.

[0182] The prediction results of the first and / or second models here can be multiple sets of temperature values ​​(e.g., driver's seat, passenger's seat, rear left, rear right) and airflow suggestions (e.g., 1–6 or weak / medium / strong).

[0183] Optionally, a prediction result can be generated at regular intervals. If the error between the predicted temperature and the previous result is within a set range (e.g., <0.5℃), the air conditioner is not adjusted; otherwise, the prediction is re-evaluated and the air conditioner is adjusted.

[0184] Understandably, this air conditioning adjustment method may include: after the air conditioning of the target vehicle is turned on, firstly acquiring local data. This local data includes, but is not limited to, local air conditioning data and local environmental data, which are collected and processed through the vehicle's infotainment system. After acquiring the local data, it is determined whether the local data meets a first training threshold. If the local data meets the first training threshold, then the first model is trained using the local data; otherwise, local data collection continues.

[0185] Meanwhile, the cloud continuously receives local data from various vehicles connected to it, forming cloud data. This cloud data also includes local air conditioning data and local environmental data from multiple vehicles. The cloud then determines whether the cloud data meets the second training threshold. If the cloud data meets the second training threshold, the second model is trained using the cloud data; otherwise, cloud data collection continues.

[0186] After the second model is trained in the cloud, its accuracy is calculated. When the accuracy of the second model reaches a set threshold, the cloud sends the trained second model to the target vehicle. Upon receiving the new second model, the target vehicle can adjust its air conditioning parameters based on the newly trained first model and / or the newly trained second model.

[0187] In the above implementation process, by determining the air conditioning adjustment model based on the update time of the first model and the second model, the model used to adjust the air conditioning can always be the latest model, thereby improving the accuracy of air conditioning adjustment.

[0188] In one possible implementation, after step S203, the method further includes: acquiring user behavior data; and adjusting vehicle air conditioning parameters based on the user behavior data.

[0189] The user behavior data includes user speech and sound playback. Of course, this user behavior data can also include actions such as wearing a hat, wearing a scarf, taking off a coat, and fanning oneself; the user behavior data can be selected according to the actual situation.

[0190] It should be understood that this user behavior data can be used to reflect whether the user is satisfied with the current air conditioning settings, and then the air conditioning parameters can be further adjusted based on this user behavior data.

[0191] For example, when it is detected that a user is playing music or video, the airflow is automatically adjusted to a low setting to avoid volume interference.

[0192] When a user is on a call, the fan speed is automatically adjusted to a low setting to avoid excessive airflow affecting call quality.

[0193] When the system detects that a user has triggered the active voice function, the airflow will be automatically adjusted to a lower setting to improve the accuracy of voice recognition.

[0194] When the system detects that the user has started wearing a hat, it automatically adjusts the airflow to the highest setting to prevent the interior temperature from getting too low and causing the user to feel cold.

[0195] When the system detects that the user has started to take off their coat, it automatically adjusts the airflow to the highest setting to prevent the interior temperature from becoming too high and causing the user to feel hot.

[0196] The above-described air conditioning adjustment method is merely an example; the adjustment method can be selected according to the actual situation.

[0197] In one embodiment, the prediction results can be updated periodically, and the decision to adjust the air conditioning system is determined based on the updated and previous prediction results. Specifically, if the updated and previous prediction results are the same, the air conditioning temperature remains unchanged. If the updated and previous prediction results are different, the air conditioning temperature is automatically adjusted to the new predicted temperature.

[0198] The first and second models mentioned above can continuously learn users' temperature preferences, wind speed preferences, and other characteristics, and are gradually optimizing the air conditioning control based on these characteristics to ensure that the air conditioning always meets the user's personalized needs.

[0199] In the above implementation process, by further adjusting the vehicle air conditioning parameters based on user behavior data, the vehicle air conditioning parameters can be further adjusted on the basis of the first model and / or the second model, so that the air conditioning parameters are more in line with user needs, thereby improving the accuracy of air conditioning adjustment and user experience.

[0200] Please see Figure 4 This is a flowchart of the second model training method provided in the embodiments of this application. The following will describe... Figure 4 The specific process shown will be explained in detail.

[0201] Step S301: Obtain local data sent by each vehicle connected to the cloud to obtain cloud data.

[0202] The cloud data includes local data from multiple vehicles connected to the cloud.

[0203] The cloud data here can include user behavior data, environmental data, vehicle status, time characteristics, and other data. The type of cloud data can be selected according to the actual situation.

[0204] Optionally, user behavior data may include air conditioning temperature settings, fan speed, mode, and adjustment frequency; environmental data may include in-vehicle and out-of-vehicle temperatures, humidity, weather, and geographical location; vehicle status may include vehicle speed, remaining battery power, window status, and sunroof status; and time characteristics may include trip date, time period, trip duration, and estimated arrival time. The specific data for each type of cloud data can be adjusted according to actual needs.

[0205] Optionally, the cloud data can be acquired at regular intervals, in real time, or triggered when local data is updated. The method of acquiring the cloud data can be selected according to the actual situation.

[0206] In one embodiment, prior to step S302, the method further includes preprocessing the cloud data. This includes, for example, handling missing values, data standardization, feature encoding, and optimizing sampling strategies. The preprocessing method for the cloud data can be selected based on the actual situation.

[0207] Specifically, missing value handling can be achieved as follows: For continuous features (e.g., temperature), mean imputation or median imputation strategies can be used to maintain the data distribution characteristics. For categorical features (e.g., air conditioning mode), mode imputation or interpolation methods can be used. Outliers (e.g., temperature > 40°C) can be removed or corrected by setting rules.

[0208] The specific methods for data standardization can be as follows: For continuous input features (such as temperature and driving distance), Min-Max normalization can be used to map the data to the [0, 1] interval, ensuring that the influence of features of different dimensions on the model is balanced. For categorical features (such as location and pattern), one-hot or multi-hot encoding can be used to avoid the interference of the order assumption between categories on model training.

[0209] One-hot and multi-hot encoding are used to convert categorical features (e.g., navigation location, air conditioning usage location) into numerical vectors, avoiding assumptions about the order of categories (e.g., assuming "driver's seat" is greater than "passenger's seat"), and ensuring that model training is not affected by partial order. The specific implementation steps are as follows:

[0210] One-hot encoding: Maps a single-class feature to a binary vector, with each class corresponding to one dimension. A value of 1 indicates presence, and 0 indicates absence. Suitable for single-valued class features, such as air conditioning modes (automatic / manual).

[0211] Multi-hot encoding: Maps multi-valued categorical features (e.g., multiple navigation locations) to binary vectors, allowing multiple dimensions to be 1. Suitable for multi-select features, such as air conditioning usage location (driver's seat, front passenger seat, rear left, rear right).

[0212] For example, the heat-only code (air conditioning mode = "Auto") is: [1, 0] (Auto = 1, Manual = 0).

[0213] Multi-hot encoding (using position = [driver's seat, rear right]): [1, 0, 0, 1] (driver's seat = 1, passenger's seat = 0, rear left = 0, rear right = 1).

[0214] Min-Max normalization here is used to map continuous input features (such as temperature and driving distance) to the [0,1] interval to eliminate the influence of features with different dimensions on model training and ensure that the contribution of features to the model is balanced. The implementation steps can be as follows: For each continuous feature, predefine or dynamically calculate the maximum and minimum values. Based on physical constraints or business rules (e.g., temperature range is fixed at 15℃–30℃, driving distance is 0–1000 miles), calculate the minimum and maximum values ​​from historical data, updating them once a day to prevent changes in data distribution from affecting the normalization.

[0215] The normalized values ​​are stored in the feature vector and used as model input. This normalization calculation can be optimized using C++ on the vehicle's infotainment system (single calculation time < 1ms, memory usage < 1KB).

[0216] For anomaly handling: If sensor data is abnormal, the median of the most recent period can be used for interpolation, and the anomaly log should be recorded. If range update fails, a defined default range can be used.

[0217] The data preprocessing here is relatively quick (e.g., less than 1ms), which can improve prediction accuracy. Furthermore, normalized data and sparse coding vectors can save storage space. For example, 100 samples occupy less than 100KB, thus improving storage efficiency.

[0218] Similar to local data, the cloud data mentioned above can also be ranked by importance using Lasso regression and / or random forest. Prioritize the cloud data with higher weights for training the second model.

[0219] In one embodiment, the cloud data can also be segmented according to time periods to generate small-batch cloud datasets.

[0220] In addition, interactive features (intersections of time periods and navigation locations) can be added to this cloud data to capture potential relationships between various cloud data.

[0221] Step S302: If the cloud data meets the second training threshold, train the second model using the cloud data.

[0222] It should be understood that, given that the cloud data meets the second training threshold, a more complex algorithm (e.g., deep learning) can be used to train the cloud data to obtain the second model.

[0223] The second model here can be a large model, which is mainly based on data with a longer time span than the first model, to learn the air conditioning adjustment methods of different users and in different weather conditions.

[0224] The aforementioned second training threshold is a threshold used to represent the acquisition status of cloud data. For example, the second training threshold can refer to the cloud data reaching a set number of data entries, the cloud data meeting a set time span, or the cloud data reaching a set number of data entries and meeting a set time span. This second training threshold can be set according to actual circumstances.

[0225] The aforementioned cloud data is continuously updated and incrementally added during the use of each vehicle, ensuring that the cloud data is constantly enriched. Once the cloud data meets the training conditions for the first model, training of the first model is initiated.

[0226] The second model is applied to the vehicle air conditioning adjustment method in the above embodiments.

[0227] The second model here can be obtained by using Deep & Cross Network (DCN), which combines deep neural networks and cross-feature learning for model training.

[0228] The second model can have the following input layers: an embedding layer for categorical features and a dense layer for continuous features. The feature interaction layer can be a DCN module that extracts cross-relationships between features (e.g., the interaction between temperature and humidity). The backbone network can be a three-layer fully connected network (e.g., 512, 256, and 128 neurons, ReLU activation function). The output layer can be a single-node Sigmoid layer that outputs a normalized temperature value.

[0229] The second model described above can also be regularized to prevent overfitting.

[0230] In one embodiment, the input features of the second model can be multi-dimensional features such as time, navigation information, air conditioning settings, vehicle status, and environmental status.

[0231] In this model, the one-hot / multi-hot encoded vectors are input into the second model. The high-dimensional vectors are then reduced in dimensionality through an embedding layer. The embedding matrix is ​​optimized during training and initialized with a random normal distribution. Normalized features are input into the second model and undergo a linear transformation through a dense layer, outputting a multi-dimensional vector with ReLU activation. Finally, the embeddings and the dense vector are concatenated.

[0232] Understandably, since the second model is a large model, it typically has a large number of model parameters. Therefore, the architecture of this second model can be optimized, for example, by removing the attention mechanism and DCN layer to reduce the number of model parameters.

[0233] The specific strategies for removing the attention mechanism and the DCN layer can be:

[0234] Attention mechanism removed: The attention mechanism is used in complex models to enhance feature weights, increasing parameters by approximately 20%. The second model removes this directly, replacing it with a simple fully connected layer that relies on Lasso to select 10 high-weight features.

[0235] Remove the DCN layer: The DCN layer (cross-feature extraction) captures feature interactions (e.g., temperature × humidity) in the first model, but has a large number of parameters (approximately 300,000). The first model is replaced with a linear layer, which takes 10-dimensional features as input and maps them directly to a 64-dimensional hidden layer.

[0236] After removing the attention mechanism and DCN layer, the model accuracy can be further tested. If the accuracy drops by a set percentage (e.g., 5%), then hidden layer neurons can be added.

[0237] Step S303: Send the trained second model to the target vehicle.

[0238] The second model can be sent directly to the target vehicle after each training session, or it can be sent to the target vehicle after the second model has been trained and its accuracy has reached a set threshold. The strategy for sending the second model to the target vehicle can be selected according to the actual situation.

[0239] The second model is used to adjust the air conditioning of the target vehicle.

[0240] In one embodiment, the second model can regulate the air conditioner by calculating the predicted temperature and outputting a normalized temperature value that is then denormalized to the actual temperature range.

[0241] The formula for calculating the predicted temperature can be:

[0242]

[0243] The actual temperature range can be expressed as:

[0244]

[0245] Where W3, W4, and W5 are the weights, and b3, b4, and b5 are the biases. To predict temperature, T pred This represents the actual temperature range.

[0246] The second model described above can predict the temperature of different areas inside the vehicle (e.g., driver's seat, front passenger seat, rear left seat, rear right seat) based on feature subsets of each region. The prediction for each region can be performed independently, outputting multiple sets of temperature values ​​to adjust the air conditioning in each area.

[0247] Understandably, once the second model in the cloud completes training and generates a new version, it can be pushed to the user's in-vehicle infotainment system. The in-vehicle system will then replace the old second model and use the new one for climate control.

[0248] In one embodiment, whenever new local data is generated on the vehicle's in-vehicle system, both the cloud-based data and the local data can be updated in real time. New cloud-based data replaces older cloud-based data, keeping the cloud dataset within a reasonable range. As new cloud-based data accumulates, the second model on the cloud is continuously updated and pushed to the vehicle's in-vehicle system, ensuring that the second model on the in-vehicle system is always up-to-date.

[0249] It should be understood that when making predictions using a newly trained first model or a newly trained second model, inverse normalization can be employed. That is, the output of the newly trained first model or the newly trained second model is normalized to a temperature value, which is then inversely normalized to the actual temperature. Inverse normalization is performed on the vehicle's infotainment system to ensure that the output conforms to the air conditioning hardware's control range.

[0250] In the above implementation process, by training a second model in the cloud, and the cloud data used to train the second model comes from multiple vehicles connected to the cloud, the second model can learn the air conditioning adjustment methods of different users and different weather conditions. It can capture complex behavioral models with high accuracy, which can improve the accuracy of the second model and thus improve the flexibility and accuracy of air conditioning adjustment.

[0251] In one possible implementation, step S302 includes: training the second model with cloud data as training data when the cloud data reaches a third predetermined quantity and the acquisition time span of the cloud data meets a second predetermined time length; and / or training the second model with cloud data as training data when the local data at any predetermined location of the target vehicle reaches a fourth predetermined quantity and the acquisition time span of the local data meets a third predetermined time length.

[0252] The third setting here is used to constrain the amount of data in the cloud. For example, 50, 100, 200, etc. This third setting can be selected according to the actual situation.

[0253] The second set time length mentioned above refers to the time span for continuously acquiring cloud data. For example, 1 month, 2 months, 3 months, 6 months, etc. This second set time length can be selected according to the actual situation.

[0254] For example, if the second training threshold is the accumulation of data from 300 vehicles in the cloud, the total amount of cloud data reaches 50,000 records, and the time span condition of one month is met, then the system automatically starts the training of the second model when it determines that the cloud data has accumulated to 50,000 records and meets the one-month time span condition. That is, the second model is trained using the accumulated 50,000 records of cloud data that meet the one-month time span condition.

[0255] The second set quantity here is used to constrain the number of local data entries at a given location. For example, 50, 100, or 200 entries can be selected based on the actual situation.

[0256] Understandably, when determining cloud data, the decision on whether to start training the second model can also be based on the amount of cloud data generated after local data from multiple locations in the car (e.g., driver's seat, passenger seat, rear left, rear right) is uploaded to the cloud.

[0257] For example, when the accumulated air conditioning data for the four positions inside the car (driver's seat, front passenger seat, rear left seat, and rear right seat) reaches 500 records and the time range reaches one month, the training of the second model can be started.

[0258] In one embodiment, when new local data is generated on the vehicle's infotainment system, the system can send the new local data to the cloud, and the cloud can update its data based on the new local data.

[0259] In the above implementation process, by setting the cloud data to a third predetermined quantity and ensuring the cloud data acquisition time span meets the second predetermined time length, using the cloud data as training data to train the second model can provide sufficient information to establish basic predictive capabilities while avoiding underfitting or excessively slowing down startup time due to insufficient cloud data, thus improving the rationality of the local data used for training the second model. Furthermore, by setting the local data from any predetermined location of the target vehicle to a fourth predetermined quantity, using the cloud data as training data to train the second model can avoid excessively long collection times and large data storage volumes, thus reducing cloud memory pressure and improving the training efficiency of the second model.

[0260] In one possible implementation, after step S302, the method further includes: calculating the accuracy of the second model by setting an algorithm.

[0261] Accuracy refers to the predictive performance of the second model on the validation set, specifically the proportion of samples where the predicted temperature differs from the actual user-set temperature by within ±2℃. The formula for calculating this accuracy is:

[0262]

[0263] Among them, the correctly predicted samples satisfy To predict temperature, y q This is the actual temperature.

[0264] The aforementioned validation set consists of 10% of the training data randomly selected from the cloud, and is therefore representative.

[0265] The accuracy here can be calculated using 5-fold cross-validation after the second model has been trained in the cloud.

[0266] It should be understood that the vehicle's infotainment system calculates the local accuracy rate (based on whether the user adjusts the temperature) after each prediction. If the accuracy rate is below 75%, a cloud optimization request can be triggered.

[0267] Step S303 includes: sending the trained second model to the target vehicle when the accuracy reaches a set threshold.

[0268] The threshold set here is the minimum accuracy that the second model can be used for air conditioning regulation.

[0269] Understandably, if the accuracy of the second model reaches the set threshold, it means that the accuracy of the second model in adjusting the air conditioning basically meets the user's needs. In this case, the second model can be sent to the target vehicle, and the air conditioning of the target vehicle can be adjusted through the second model.

[0270] In one embodiment, the set threshold can be 80%. That is, when the accuracy reaches 80%, the trained second model is sent to the target vehicle.

[0271] In another embodiment, the evaluation metric of the second model may further include mean squared error. The formula for calculating the mean squared error can be:

[0272] in, To predict temperature, y q is the actual temperature, and m is the amount of cloud data used in the training.

[0273] Given the mean squared error, the relationship between the mean squared error and the error threshold can be used to determine whether to send the trained second model to the target vehicle.

[0274] The mean square error mentioned above is used to measure the deviation between the predicted temperature and the actual temperature.

[0275] Step S303 here may also include: sending the trained second model to the target vehicle when the accuracy reaches a set threshold or the mean square error reaches an error threshold.

[0276] For example, if the threshold is set to 80% and the error threshold is 0.04, if the accuracy is determined to be <80% or MSE > 0.04, then training or optimizing the hyperparameters (such as learning rate, Dropout rate) of the second model continues until the conditions are met. If the accuracy is determined to be ≥80% or MSE ≤ 0.04, then the trained second model is sent to the target vehicle.

[0277] Understandably, when the training data set of the second model has not been updated for a long time (e.g., more than 3 months), causing the old training dataset to become less relevant, the accuracy of the second model can be optimized by analyzing the behavioral characteristics of multiple users in the current time period, extracting patterns, and applying them to the current users.

[0278] That is, based on the data extraction model, the usage behavior of multiple users is recorded and analyzed to extract patterns applicable to the current user, and these patterns are fed back into the second model to ensure the accuracy and real-time performance of the second model.

[0279] In the above implementation process, by setting the second model to be sent to the target vehicle only when the accuracy of the second model reaches a set threshold, the accuracy of the second model used for air conditioning adjustment can always meet the user's needs, thereby improving the accuracy of the second model.

[0280] Based on the same application concept, this application also provides a vehicle air conditioning adjustment device corresponding to the vehicle air conditioning adjustment method. Since the principle of the device in this application is similar to that of the aforementioned vehicle air conditioning adjustment method, the implementation of the device in this application can refer to the description in the above-mentioned method embodiments, and the repeated parts will not be described again.

[0281] Please see Figure 5 This is a functional module diagram of the vehicle air conditioning adjustment device provided in this application embodiment. Each module in the vehicle air conditioning adjustment device in this embodiment is used to execute the steps in the above method embodiments. The vehicle air conditioning adjustment device includes a first acquisition module 301, a first training module 302, a receiving module 303, and an adjustment module 304; wherein,

[0282] The first acquisition module 301 is used to acquire local data after the air conditioning of the target vehicle has been turned on for a set time, and send the local data to the cloud connected to the target vehicle; wherein, the local data includes local air conditioning data and local environmental data.

[0283] The first training module 302 is used to train a first model using the local data on the vehicle's infotainment system when the local data meets a first training threshold.

[0284] The receiving module 303 is used to receive a newly trained second model sent from the cloud; wherein the newly trained second model is configured to be trained using the cloud data when the cloud data meets a second training threshold; the cloud data includes the local data of multiple vehicles connected to the cloud.

[0285] The first adjustment module 304 is used to adjust the vehicle air conditioning parameters based on the newly trained first model and / or the newly trained second model.

[0286] In one possible implementation, the first training module 302 is further configured to: train the first model using the local data as training data when the local data reaches a first predetermined quantity and the time span of acquiring the local data satisfies a first predetermined time length; and / or train the first model using the local data as training data when the local data at any predetermined location of the target vehicle reaches a second predetermined quantity.

[0287] In one possible implementation, the vehicle air conditioning control device further includes an optimization module for optimizing the first model based on positive samples and online gradient descent when it is determined that the training dataset of the first model has not been updated within a first set time period; wherein the training dataset is a collection of local data used to train the first model, and the positive samples are system prediction parameters that the user has not manually changed within a second set time period.

[0288] In one possible implementation, the first adjustment module 304 is further configured to: obtain a first update time of the newly trained first model and a second update time of the newly trained second model; wherein the newly trained first model refers to the first model most recently updated from the current time, and the newly trained second model refers to the second model most recently updated from the current time; determine the time interval between the first update time, the second update time and the current time; and adjust the vehicle air conditioning parameters using the newly trained first model and / or the newly trained second model with the smaller time interval.

[0289] In one possible implementation, the vehicle air conditioning adjustment device further includes a second adjustment module 304 for acquiring user behavior data; wherein the user behavior data includes user speech and sound playback; and the vehicle air conditioning parameters are adjusted according to the user behavior data.

[0290] Based on the same application concept, this application also provides a second model training device corresponding to the second model training method. Since the principle of the device in this application is similar to that of the aforementioned second model training method embodiment, the implementation of the device in this embodiment can refer to the description in the above method embodiment, and the repeated parts will not be repeated.

[0291] Please see Figure 6 This is a functional module diagram of the second model training device provided in this application embodiment. Each module in the second model training device in this embodiment is used to execute the steps in the above method embodiments. The second model training device includes a second acquisition module 401, a second training module 402, and a sending module 403; wherein,

[0292] The second acquisition module 401 is used to acquire local data sent by each vehicle connected to the cloud, and obtain cloud data; wherein, the cloud data includes local data of multiple vehicles connected to the cloud.

[0293] The second training module 402 is used to train a second model using the cloud data when the cloud data meets the second training threshold.

[0294] The sending module 403 is used to send the trained second model to the target vehicle; wherein the second model is applied to the vehicle air conditioning adjustment method in the above embodiment.

[0295] In one possible implementation, the second training module 402 is further configured to: use the cloud data as training data to train the second model when the cloud data reaches a third predetermined quantity and the acquisition time span of the cloud data meets a second predetermined time length; and / or use the cloud data as training data to train the second model when the local data at any predetermined location of the target vehicle reaches a fourth predetermined quantity and the acquisition time span of the local data meets a third predetermined time length.

[0296] In one possible implementation, the second model training device further includes a calculation module for calculating the accuracy of the second model by setting an algorithm.

[0297] In one possible implementation, the sending module 403 is specifically used to: send the trained second model to the target vehicle when the accuracy reaches a set threshold.

[0298] Furthermore, embodiments of this application also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the steps of the vehicle air conditioning adjustment and / or the second model training method described in the above method embodiments.

[0299] The computer program product of the vehicle air conditioning adjustment and / or second model training method provided in this application embodiment includes a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the steps of the vehicle air conditioning adjustment and / or second model training method described in the above method embodiment. For details, please refer to the above method embodiment, which will not be repeated here.

[0300] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can also be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0301] In addition, the functional modules in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0302] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks. It should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0303] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application. It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0304] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for adjusting vehicle air conditioning, characterized in that, include: After the target vehicle's air conditioning is turned on for a set time, local data is acquired and sent to the cloud connected to the target vehicle; wherein, the local data includes local air conditioning data and local environmental data; If the local data on the in-vehicle infotainment system of the target vehicle meets a first training threshold, a first model is trained using the local data; and / or The system receives a newly trained second model sent from the cloud; wherein the newly trained second model is configured to be trained using the cloud data when the cloud data meets a second training threshold; the cloud data includes the local data of multiple vehicles connected to the cloud. Adjust vehicle air conditioning parameters based on the newly trained first model and / or the newly trained second model; The adjustment of vehicle air conditioning parameters based on the newly trained first model and / or the newly trained second model includes: Obtain the first update time of the newly trained first model and the second update time of the newly trained second model; wherein, the newly trained first model refers to the first model most recently updated with the current time, and the newly trained second model refers to the second model most recently updated with the current time; Determine the time interval between the first update time, the second update time, and the current time; The vehicle air conditioning parameters are adjusted using the newly trained first model and / or the newly trained second model with smaller time intervals.

2. The method according to claim 1, characterized in that, The step of training a first model using the local data on the in-vehicle infotainment system of the target vehicle, when the local data meets a first training threshold, includes: When the local data reaches a first predetermined quantity and the acquisition time span of the local data meets a first predetermined time length, the local data is used as training data to train the first model; and / or If the local data at any set location of the target vehicle reaches a second set quantity, the local data is used as training data to train the first model.

3. The method according to claim 1, characterized in that, After training the first model using the local data when the local data on the vehicle's infotainment system meets the first training threshold, the method further includes: If it is determined that the training dataset of the first model has not been updated within a first set time period, the first model is optimized based on positive samples and online gradient descent; wherein, the training dataset is a collection of local data used to train the first model, and the positive samples are system prediction parameters that the user has not manually changed within a second set time period.

4. The method according to any one of claims 1-3, characterized in that, After adjusting the vehicle air conditioning parameters based on the newly trained first model and / or the newly trained second model, the method further includes: Acquire user behavior data; wherein, the user behavior data includes user speech and played sound; Adjust vehicle air conditioning parameters based on the user behavior data.

5. A second model training method, characterized in that, Applied to the cloud, the method includes: Local data sent by each vehicle connected to the cloud is obtained to obtain cloud data; wherein, the cloud data includes local data of multiple vehicles connected to the cloud; If the cloud data meets the second training threshold, a second model is trained using the cloud data. Send the trained second model to the target vehicle; The second model is applied to the vehicle air conditioning adjustment method according to any one of claims 1-4.

6. The method according to claim 5, characterized in that, The step of training a second model using the cloud data when the cloud data meets the second training threshold includes: When the amount of cloud data reaches a third predetermined quantity, and the acquisition time span of the cloud data meets a second predetermined time length, the cloud data is used as training data to train the second model; and / or If the local data at any set location of the target vehicle reaches a fourth set quantity, and the acquisition time span of the local data meets a third set time length, the cloud data is used as training data to train the second model.

7. The method according to claim 5, characterized in that, After training the second model using the cloud data when the cloud data meets the second training threshold, the method further includes: The accuracy of the second model is calculated by setting an algorithm; Sending the trained second model to the target vehicle includes: If the accuracy reaches a set threshold, the trained second model is sent to the target vehicle.

8. An electronic device, characterized in that, include: The processor and memory, wherein the memory stores machine-readable instructions executable by the processor, wherein when the electronic device is running, the machine-readable instructions are executed by the processor to perform the steps of the method as described in any one of claims 1 to 4, or the method as described in any one of claims 5 to 7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, performs the steps of the method as described in any one of claims 1 to 5, or the method as described in any one of claims 5 to 7.

Images