Intelligent atomizer power prediction system and method, atomizer system and atomizer

Abstract

The invention discloses an atomizer power intelligent prediction system and method, an atomizer system and an atomizer. Multi-source heterogeneous data of one or more users is obtained; the multi-source heterogeneous data comprises one or more of user behavior data of the atomizer, equipment characteristic data of the atomizer, environment parameter data of the atomizer and time characteristic data of the atomizer; according to the multi-source heterogeneous data of each user, obtaining a feature vector corresponding to the user, and inputting the feature vector into a multi-model fusion framework to obtain a power prediction result corresponding to the user; wherein the multi-model fusion framework comprises a plurality of basic models. By comprehensively considering multi-dimensional information such as user use habits, equipment characteristics, environmental factors and time characteristics and combining hyper-parameter optimization and a three-level model degradation mechanism based on a multi-model fusion framework, on the premise of ensuring prediction precision, quick response is realized, the stability of power control service is ensured, and the prediction efficiency is improved. Therefore, high-precision intelligent prediction and self-adaptive control of the power of the atomizer are realized.

Description

Technical Field

[0001] This invention relates to the field of atomizers, and more specifically to an intelligent power prediction system, method, atomizer system, and atomizer. Background Technology

[0002] The power settings of existing atomizer products are mainly based on empirical values ​​or simple power levels. However, since only a few fixed power levels are usually provided, the accuracy of power control is limited and cannot meet the needs of fine control. Furthermore, the power control process is easily affected by factors such as battery power, making it difficult to guarantee the stability of power output.

[0003] In addition, traditional atomizer power control systems with fixed power levels cannot be personalized based on different users' preferences, physiological characteristics, and usage environments, nor can they predict the optimal power based on users' historical usage data. Therefore, a system is needed to accurately predict and intelligently control atomizer power, thereby providing a personalized and adaptive user experience. Summary of the Invention

[0004] The main technical problem solved by this invention is how to achieve a system for accurate prediction and intelligent control of atomizer power, thereby providing a personalized and adaptive user experience.

[0005] According to the first aspect, one embodiment provides an intelligent atomizer power prediction system based on multi-model fusion, comprising: The data acquisition module is used to acquire multi-source heterogeneous data from one or more users; the multi-source heterogeneous data includes one or more of the following: user behavior data of the atomizer, device characteristic data of the atomizer, environmental parameter data of the atomizer, and time characteristic data of the atomizer. The data preprocessing module is used to preprocess the multi-source heterogeneous data of each user to obtain the feature vector corresponding to each user; The power prediction module is used to input the feature vector corresponding to each user into the multi-model fusion framework to obtain the power prediction result corresponding to that user; wherein, the multi-model fusion framework includes multiple basic models; the model parameters and reference weights of each basic model in the multi-model fusion framework are obtained by searching in the preset search space of each basic model through the optimization engine in the hyperparameter optimization module.

[0006] According to the second aspect, one embodiment provides an intelligent prediction method for atomizer power based on multi-model fusion, including: Acquire multi-source heterogeneous data from one or more users; the multi-source heterogeneous data includes one or more of the following: user behavior data of the atomizer, device characteristic data of the atomizer, environmental parameter data of the atomizer, and time characteristic data of the atomizer. The multi-source heterogeneous data for each user is preprocessed to obtain the feature vector corresponding to each user. The feature vector corresponding to each user is input into the multi-model fusion framework to obtain the power prediction result corresponding to that user; wherein, the multi-model fusion framework includes multiple basic models; the model parameters and reference weights of each basic model in the multi-model fusion framework are obtained by the optimization engine searching in the preset search space of each basic model.

[0007] According to a third aspect, one embodiment provides an atomizer system, including: one or more atomizers and a cloud, wherein each atomizer and the cloud are communicatively connected. The cloud is used to receive multi-source heterogeneous data from each atomizer, obtain power prediction results for each atomizer based on the multi-source heterogeneous data of each atomizer and the aforementioned atomizer power intelligent prediction system based on multi-model fusion, and send the power prediction results of each atomizer to the corresponding atomizer. Each atomizer adjusts its power based on the power prediction result received from that atomizer.

[0008] According to the fourth aspect, one embodiment provides an atomizer, comprising: Memory, used to store computer programs; A processor is configured to implement the aforementioned intelligent prediction method for atomizer power based on multi-model fusion by executing the computer program stored in the memory.

[0009] The atomizer power intelligent prediction system, method, atomizer system, and atomizer according to the above embodiments collect multi-source heterogeneous data from one or more users to comprehensively consider multi-dimensional information such as user habits, device characteristics, environmental factors, and time characteristics. Then, through a multi-model fusion framework, different basic models are intelligently combined to construct an integrated learning architecture based on multi-model fusion, thereby establishing a power prediction model that is both general and reflects individual differences. By combining the characteristics of different basic models and hyperparameter optimization technology, high-precision intelligent prediction and adaptive control of atomizer power are achieved. Attached Figure Description

[0010] Figure 1 This is a system module diagram of an intelligent prediction system for atomizer power based on multi-model fusion. Figure 2The flowchart shows the method for intelligent prediction of atomizer power based on multi-model fusion. Figure 3 This is a system block diagram of an atomizer system; Figure 4 This is a schematic diagram of the structure of an atomizer; Figure 5 A comparison chart showing the power prediction results output by the multi-model fusion framework and the actual power. Figure 6 This is a residual distribution diagram. Detailed Implementation

[0011] The present invention will now be described in further detail with reference to specific embodiments and accompanying drawings. Similar elements in different embodiments are referred to by associated similar element reference numerals. In the following embodiments, many details are described to facilitate a better understanding of this application. However, those skilled in the art will readily recognize that some features may be omitted in different situations, or may be replaced by other elements, materials, or methods. In some cases, certain operations related to this application are not shown or described in the specification. This is to avoid obscuring the core parts of this application with excessive description. For those skilled in the art, detailed description of these related operations is not necessary; they can fully understand the related operations based on the description in the specification and general technical knowledge in the art.

[0012] Furthermore, the features, operations, or characteristics described in the specification can be combined in any suitable manner to form various embodiments. At the same time, the steps or actions in the method description can be rearranged or adjusted in a manner obvious to those skilled in the art. Therefore, the various orders in the specification and drawings are only for the clear description of a particular embodiment and do not imply a necessary order, unless otherwise stated that a particular order must be followed.

[0013] The serial numbers assigned to components in this document, such as "first" and "second," are used only to distinguish the described objects and have no sequential or technical meaning. The terms "connection" and "linkage" used in this application, unless otherwise specified, include both direct and indirect connections (linkages).

[0014] In the current field of atomizer power control technology, commonly used techniques include: 1) Temperature sensor-based power control technology, which integrates a temperature sensor into the atomizer to automatically adjust the power output according to temperature changes; 2) Resistance-based power adjustment technology, which detects the resistance value of the atomizer, automatically calculates and adjusts the power output to ensure a similar user experience on atomizers with different resistances; 3) Preset power levels combined with user manual selection adjustment technology, which provides several preset power levels and allows users to manually select the appropriate level according to their personal preferences; 4) Simple user preference memory technology, which records the user's last power setting and automatically applies that setting the next time the device is turned on.

[0015] However, due to the significant differences in power requirements between light and heavy users, the large differences in performance of different flavored atomizing substrates at the same power, the influence of environmental conditions on atomization effects, and the impact of user frequency and duration on optimal power selection, existing power control technologies based on temperature sensors only consider the single environmental factor of temperature. They lack comprehensive consideration of multi-dimensional information such as individual user differences, usage habits, and environmental factors, thus ignoring personalized user needs. Furthermore, temperature sensors increase equipment cost and complexity and are easily affected by atomizing substrate contamination, leading to inaccurate readings. In addition, this method can only respond passively and cannot actively predict the user's power requirements.

[0016] Power regulation technology based on resistance value detection ignores important variables such as atomization matrix characteristics and environmental factors, and controls power solely based on resistance value, failing to consider user preferences. Furthermore, this method cannot differentiate the needs of different usage scenarios for atomizers with the same resistance value; and it lacks learning ability, making it unable to optimize based on user feedback.

[0017] The adjustment technology that combines preset gears with manual user selection has limitations in terms of the number of gears it offers, making it unable to meet the needs of fine-tuning. Furthermore, this method relies entirely on user experience, making it difficult for novice users to quickly find suitable settings. It also lacks data accumulation and learning mechanisms, making it unable to automatically adjust according to different usage scenarios.

[0018] Simple user preference memory technology can only remember a single setting, cannot adapt to different usage scenarios, does not consider the impact of dynamic factors such as time and environment, lacks predictive ability, cannot proactively recommend the optimal settings, and has extremely low efficiency in utilizing user-related data, and cannot uncover user behavior patterns.

[0019] In this embodiment of the invention, multi-source heterogeneous data from one or more users are collected to comprehensively consider multi-dimensional information such as user habits, device characteristics, environmental factors, and time characteristics. Then, different basic models are intelligently combined through a multi-model fusion framework, and combined with hyperparameter optimization technology and a three-level model degradation mechanism, to achieve high-precision intelligent prediction and adaptive control of atomizer power.

[0020] Please refer to Figure 1 Some embodiments provide an intelligent prediction system for atomizer power based on multi-model fusion, including: The data acquisition module 100 is used to acquire multi-source heterogeneous data from one or more users; wherein the multi-source heterogeneous data includes one or more of the following: user behavior data of the atomizer, device characteristic data of the atomizer, environmental parameter data of the atomizer, and time characteristic data of the atomizer.

[0021] In this embodiment, user behavior data includes average user usage time (i.e., the average duration of each user's use of the atomizer per use, ranging from 1.5s to 5.0s) and number of uses within 24 hours (i.e., the number of times each user uses the atomizer within 24 hours, ranging from 5 to 50 times); device characteristic data includes one or more of the device model and flavor type, where the device model is used to distinguish devices with different functions, performance, and specifications, and it can generally be defined, identified, and distinguished by a number consisting of letters and / or numbers. For example, based on the differences between the appearance of different atomizer devices and / or the number of corresponding inhalation ports, the device model is identified as m01, m02, and m03—of course, this is just an example and does not mean that the actual device model identification can only be in this way; flavor type includes mint, fruit, and drink; environmental parameter data includes geographical location, referring to the country where the user is located when using the atomizer; time characteristic data represents Monday to Sunday, i.e., which day of the week the user used the atomizer, with a value ranging from 0 to 6.

[0022] The data preprocessing module 110 is used to preprocess the multi-source heterogeneous data of each user to obtain the feature vector corresponding to each user.

[0023] A feature engineering approach combining feature coding and standardization is used to preprocess the multi-source heterogeneous data corresponding to each user. For any type of data for each user, when the data represents a category, such as device model, taste type, or environmental parameter data, the data is feature-coded for subsequent use. When the data represents numerical values, such as average user usage time, number of uses within 24 hours, and time characteristic data, the data is standardized. In this embodiment, the average value and standard deviation of the corresponding values ​​of the data for all users are first obtained; the difference between the corresponding value of the data type for the user and the obtained average value is calculated; and then the ratio between the obtained difference and the obtained standard deviation is used as the standardization result of the data. Then, based on the encoding or standardization results of each type of data for that user, the corresponding feature vector is obtained.

[0024] The power prediction module 120 is used to input the feature vector corresponding to each user into the multi-model fusion framework to obtain the power prediction result for that user. The multi-model fusion framework includes several basic models. The LightGBM (Lightweight Gradient Boosting Machine) model can achieve efficient gradient boosting, has good ensemble performance, strong generalization ability, and strong handling capability for nonlinear relationships. The Random Forest model has excellent robustness and powerful feature importance analysis capabilities. The Ridge Regression model can quickly model linear relationships and has high computational efficiency. The Elastic Network model has strong feature selection capabilities and can effectively prevent overfitting. Therefore, this embodiment uses the LightGBM model, Random Forest model, Ridge Regression model, and Elastic Network model as basic models to construct an ensemble learning architecture based on multi-model fusion.

[0025] Furthermore, the model parameters of each base model and the reference weights between each base model in the multi-model fusion framework are obtained by searching in the preset search space of each base model in the hyperparameter optimization module 130 through an optimization engine (such as the Optuna optimization engine).

[0026] Among them, hyperparameters are a special type of parameter that needs to be manually set before model training begins, rather than being automatically learned during the model training process. They are used to control the model's structure, training process, or optimization strategy, and directly affect the model's performance, convergence speed, and generalization ability. In the hyperparameter optimization module 130, the hyperparameters of each base model are optimized. For any base model: the optimization engine samples the current hyperparameter combination in the preset search space of the base model to obtain the current hyperparameter combination, which includes the model parameters of the base model and the reference weights corresponding to the base model; then, the base model is trained according to the current hyperparameter combination; the performance of the base model is evaluated by cross-validation to obtain the average cross-validation score of the current hyperparameter combination, and it is determined whether the convergence condition is met; in this embodiment, the convergence condition means that the fluctuation of the average cross-validation score obtained in M ​​consecutive rounds of iterative training is less than a preset threshold or reaches the target performance, where M is set according to the actual situation.

[0027] If the convergence condition is met, the hyperparameter optimization ends and the optimal hyperparameter combination is obtained, that is, the hyperparameter combination that maximizes the average cross-validation score is taken as the optimal hyperparameter combination. If the convergence condition is not met and the maximum number of iterations has not been reached, sampling continues in the preset search space corresponding to the basic model to obtain a new combination of hyperparameters. The basic model is then trained for the next round based on the new combination of hyperparameters. If the maximum number of iterations is reached and the convergence condition is not met, hyperparameter optimization is stopped. Based on the reasons for non-convergence, such as improper definition of the search space, targeted adjustments are made before iterative training is restarted.

[0028] Since the multi-model fusion framework involves multiple base models, in the power prediction task, the reference weights of each base model can not only reflect their relative importance and actual contribution in the multi-model fusion framework, but also allow the advantages of different models to complement each other. At the same time, the distribution of reference weights will also reflect the bias of the integration strategy, which helps to improve the overall performance of the multi-model fusion framework.

[0029] It should be noted that the method used when sampling features and samples is TEP sampling, which is used to efficiently explore the optimal combination of hyperparameters in a predefined search space. Its core idea is based on Bayes' theorem and kernel density estimation, using historical search results to guide subsequent sampling processes, thereby finding the optimal solution faster.

[0030] Taking the LightGBM model as an example, the model parameters involved in the hyperparameter optimization of this basic model include the number of leaf nodes (its preset search space range is 10 to 300), learning rate (its preset search space range is 0.01 to 0.3), feature sampling ratio (its preset search space range is 0.4 to 1.0), sample sampling ratio (its preset search space range is 0.4 to 1.0), minimum number of samples per child node (its preset search space range is 5 to 100), and L1 and L2 regularization coefficients (their preset search space ranges are both 0 to 10). The preset search space range of the reference weights of this basic model is 0 to 1, and the sum of the reference weights of all basic models is 1. The current hyperparameter combination is obtained by searching the preset search space corresponding to each model parameter through TEP sampling. Then, the current hyperparameter combination is input into the created LightGBM regression model, and the performance of the base model is evaluated by K-fold cross-validation. In this embodiment, K is 5. In this process, K independent training sessions are performed based on the current hyperparameter combination. Each training session uses a different training subset / validation subset. Specifically, the entire training set is divided into K parts. In the i-th training session, the i-th part is used as the validation subset, and the remaining K-1 parts are used as the training subset. Then, the basic model is trained based on all training subsets. After the basic model is trained, the multi-source heterogeneous data corresponding to each user in the validation subset is input into the basic model to obtain the power prediction result for each user. Then, the root mean square error (RMSE) is calculated based on the power prediction result and the true value (i.e., the actual power of the user when using it). In this embodiment, the negative root mean square error (-RMSE) is used as the scoring index to evaluate the model performance. Finally, the average of the negative root mean square error corresponding to the K training iterations is calculated to obtain the average cross-validation score corresponding to the current hyperparameter combination. At this point, the obtained average cross-validation score is the model performance score corresponding to the current hyperparameter combination.

[0031] Then, it is determined whether the convergence condition is met. If the convergence condition is met, the hyperparameter optimization ends and the optimal hyperparameter combination is obtained. If the convergence condition is not met and the maximum number of iterations has not been reached, sampling continues in the preset search space corresponding to the basic model, and the basic model is retrained for the next round. If the maximum number of iterations is reached and the convergence condition is not met, the hyperparameter optimization stops, and targeted adjustments are made according to the reason for non-convergence before iterative training is carried out again.

[0032] In some embodiments, after reaching the maximum number of iterations, the hyperparameter combination with the highest average cross-validation score among all hyperparameter combinations is directly taken as the optimal hyperparameter combination.

[0033] The model management module 140 includes at least a three-level model degradation mechanism. In this embodiment, the three-level model degradation mechanism includes a high-precision training model, a cached lightweight model, and an on-the-spot training backup model to ensure the output of power prediction results.

[0034] Generally, power prediction results are obtained first through a multi-model fusion framework (i.e., a high-precision trained model). However, in actual use, factors such as network failures may occur, causing the high-precision trained model to fail to load and thus become unusable. To ensure stable power control services are provided under various abnormal conditions, this embodiment uses a three-level model degradation mechanism. By judging the availability status of models corresponding to different priorities, it ensures that a power prediction result will eventually be obtained. Specifically, if the model of the current priority is available, the output result of the model corresponding to the current priority is used as the power prediction result. If the model of the current priority is unavailable, the availability of the next priority model is then judged, and so on. In the three-level model degradation mechanism of this embodiment, the priorities from high to low are: high-precision trained model, cached lightweight model, and on-demand trained backup model.

[0035] It should be noted that the cached lightweight model is a pre-trained backup model, while the on-the-fly trained backup model is an empirical formula model. When other priority models are unavailable, the final power prediction result can be obtained through the on-the-fly trained model, thereby ensuring that a stable power control service can be provided.

[0036] In addition, the model management module 140 is also used to manage model training, model saving, model loading, and model version control (such as version updates).

[0037] The prediction service module 150 provides single prediction services, batch prediction services, and online learning services. The online learning service, as the number of users increases during actual use, collects multi-source heterogeneous data from new users and merges the new data with the existing data to generate a new, richer dataset. Based on this new dataset, the multi-model fusion framework is trained and optimized online, resulting in a new version of the multi-model fusion framework with timestamps. Subsequent power predictions are performed by automatically loading the latest version of the multi-model fusion framework.

[0038] Furthermore, when the number of test samples is 1000, the power prediction results output by the multi-model fusion framework in this embodiment are compared with the actual power as shown in the figure below. Figure 5As shown, where, Figure 5 The horizontal axis represents the actual power, and the vertical axis represents the predicted power. The root mean square error (RMSE) is 1.182, the mean absolute error (MAE) is 0.937, and the coefficient of determination (R²) is... 2 The value is 0.670; the residual distribution plot is as follows: Figure 6 As shown, the horizontal axis represents the residual and the vertical axis represents the frequency. The residual is the difference between the actual power and the predicted power. The mean residual is -0.112 and the median residual is -0.114.

[0039] API service module 160 adopts a RESTful architecture to provide RESTful interfaces, including POST / predict (single prediction interface), POST / predict / batch (batch prediction interface), POST / predict / upload (file upload prediction interface), GET / health (health check interface), and GET / model / info (model information interface). This embodiment further illustrates different prediction services using multi-source heterogeneous data from specific users and verifies the model performance; It should be noted that in the verification phase, this embodiment adopts a dynamic weight allocation strategy based on the performance of the verification set to assign corresponding reference weights to different base models. In this process, the root mean square error of each base model is obtained based on the verification set, the sum of the reciprocals A of the root mean square errors of all base models is calculated, and the ratio between the reciprocal B of the root mean square error of each base model and the sum of the reciprocals A is used as the reference weight of that base model.

[0040] In addition, system initialization and environment configuration are required before the prediction service can be provided. Environment configuration includes creating predictor instances and ensuring the model is ready, while system initialization includes defining features and configuring the encoder. For a single prediction service, multi-source heterogeneous data of a single user is acquired, the acquired multi-source user data is preprocessed, and then a feature vector is constructed. The obtained feature vector is then input into a multi-model fusion framework, and combined with the reference weights corresponding to each basic model, the power prediction result is obtained.

[0041] In a single prediction service, it is assumed that the average user usage time in the input multi-source heterogeneous data is 3.5s, the number of times it is used in 24 hours is 20, the device model is m01, the flavor type is fruit flavor, the code of the user's country is CN, and the corresponding time feature data is 1. Data preprocessing is performed based on the data type of each data in the input multi-source heterogeneous data. Specifically, when the data type of a data represents a category, feature encoding is performed on the data to convert it into a specific numerical value; when the data type represents a numerical value, the data is standardized. Feature vectors are constructed based on the preprocessed multi-source heterogeneous data. These feature vectors are then input into a multi-model fusion framework for power prediction, yielding prediction results from each base model. The prediction results from the LightGBM model, Random Forest model, Ridge Regression model, and Elastic Network model are 10.1, 10.3, 10.4, and 10.0, respectively. The prediction results from each base model are then weighted and fused to obtain the final power prediction result, based on the reference weights of each base model. If the reference weights for each base model are 0.34, 0.26, 0.2, and 0.2, respectively, then the final power prediction result for the input multi-source heterogeneous data is 10.2 (rounded to one decimal place), with a total processing time of 42.5 ms.

[0042] For batch prediction services, firstly, multi-source heterogeneous data containing multiple users with different user types and usage scenarios is generated. Taking 1000 samples as an example, the experimental results show that for these 1000 test samples, the processing time required by the multi-model fusion framework is 4.237 seconds, the average response time is 4.24 milliseconds / sample, and the throughput is 236.0 samples / second.

[0043] This embodiment further calculates the root mean square error (RMSE) and coefficient of determination (R²) through cross-validation (such as 5-fold cross-validation). 2 The model performance was evaluated using [a specific method / mechanism]. Validation results showed that the RMSE of the multi-model fusion framework was 1.347 ± 0.089, and R² was 0.892; while the RMSE of the LightGBM single model was 1.523, and R² was 0.851; the RMSE of the Random Forest single model was 1.634, and R² was 0.831; the RMSE of the Ridge Regression single model was 1.789, and R² was 0.798; and the RMSE of the Elastic Network single model was 1.712, and R² was 0.815. Therefore, the multi-model fusion framework improved the prediction accuracy by 13.7% compared to the single base model. Furthermore, experimental results show that the multi-model fusion framework improves prediction accuracy by 12% to 20% compared to a single base model, with the root mean square error controlled within the range of 1.2 to 1.5. Therefore, the multi-model fusion framework can significantly improve prediction accuracy. The single prediction response time of the multi-model fusion framework is less than 50 milliseconds, and the batch prediction processing efficiency is improved by more than 300%, greatly optimizing the response speed of the prediction service. The three-level model degradation mechanism can ensure service availability of more than 99.9%, and the accuracy of personalized power recommendation reaches more than 85%. Therefore, the multi-model fusion framework can significantly improve user experience. In other words, the multi-model fusion framework in this embodiment can achieve high-precision and fast-response personalized power prediction.

[0044] This embodiment collects multi-source heterogeneous data from one or more users, comprehensively considers multi-dimensional information such as user habits, device characteristics, environmental factors, and time characteristics, and uses a multi-model fusion framework to intelligently combine different basic models to construct an integrated learning architecture based on multi-model fusion. This establishes a power prediction model that is both general and reflects individual differences. Then, by combining the characteristics of different basic models and hyperparameter optimization techniques, a millisecond-level response speed is achieved while ensuring prediction accuracy. Online learning is performed based on user usage to continuously optimize prediction performance. In addition, this embodiment also uses a three-level model degradation mechanism to ensure stable power control services under various abnormal conditions, achieving high-precision intelligent prediction and adaptive control of atomizer power.

[0045] Please refer to Figure 2 Some embodiments provide a method for intelligent prediction of atomizer power based on multi-model fusion, including: Step S200: Obtain multi-source heterogeneous data from one or more users.

[0046] Among them, multi-source heterogeneous data includes one or more of the following: user behavior data of atomizers, device characteristic data of atomizers, environmental parameter data of atomizers, and time characteristic data of atomizers. Step S210: Perform data preprocessing on the multi-source heterogeneous data of each user to obtain the feature vector corresponding to each user.

[0047] Step S220: Input the feature vector corresponding to each user into the multi-model fusion framework to obtain the power prediction result corresponding to that user.

[0048] The multi-model fusion framework includes multiple base models. The model parameters of each base model in the multi-model fusion framework and the reference weights between each base model are obtained by the optimization engine searching in the preset search space of each base model.

[0049] It should be noted that the various method steps of the atomizer power intelligent prediction method based on multi-model fusion in this embodiment correspond to the data acquisition module, data preprocessing module, and power prediction module of the atomizer power intelligent prediction system based on multi-model fusion described above. Their specific implementation methods have been described in detail in the above embodiments and will not be repeated here.

[0050] Please refer to Figure 3 Some embodiments provide an atomizer system including one or more atomizers 300 and a cloud 310, each atomizer 300 and the cloud 310 being communicatively connected; Among them, the cloud 310 is used to receive multi-source heterogeneous data from each atomizer 300, and obtain the power prediction result of each atomizer 300 based on the multi-source heterogeneous data of each atomizer 300 and the aforementioned atomizer power intelligent prediction system based on multi-model fusion; and send the power prediction result of each atomizer 300 to the corresponding atomizer. Each atomizer 300 adjusts its power based on the power prediction result received from that atomizer.

[0051] Please refer to Figure 4 Some embodiments provide an atomizer, including: Memory 400 is used to store computer programs; The processor 410 is used to implement the above-mentioned intelligent prediction method for atomizer power based on multi-model fusion by executing the computer program stored in the memory 400.

[0052] Those skilled in the art will understand that all or part of the functions of the various methods in the above embodiments can be implemented by hardware or by computer programs. When all or part of the functions in the above embodiments are implemented by computer programs, the program can be stored in a computer-readable storage medium, which may include: read-only memory, random access memory, disk, optical disk, hard disk, etc., and the program is executed by a computer to achieve the above functions. For example, the program can be stored in the memory of a device, and when the program in the memory is executed by the processor, all or part of the above functions can be achieved. In addition, when all or part of the functions in the above embodiments are implemented by computer programs, the program can also be stored in a server, another computer, disk, optical disk, flash drive, or external hard drive, etc., and can be downloaded or copied to the memory of a local device, or the system of the local device can be updated. When the program in the memory is executed by the processor, all or part of the functions in the above embodiments can be achieved.

[0053] The above examples illustrate the present invention only to aid in understanding it and are not intended to limit the scope of the invention. Those skilled in the art can make various simple deductions, modifications, or substitutions based on the principles of this invention.

Claims

1. A multi-model fusion-based intelligent prediction system for atomizer power, characterized in that, include: The data acquisition module is used to acquire multi-source heterogeneous data from one or more users; the multi-source heterogeneous data includes one or more of the following: user behavior data of the atomizer, device characteristic data of the atomizer, environmental parameter data of the atomizer, and time characteristic data of the atomizer. The data preprocessing module is used to preprocess the multi-source heterogeneous data of each user to obtain the feature vector corresponding to each user; The power prediction module is used to input the feature vector corresponding to each user into the multi-model fusion framework to obtain the power prediction result corresponding to that user; wherein, the multi-model fusion framework includes multiple basic models; the model parameters and reference weights of each basic model in the multi-model fusion framework are obtained by searching in the preset search space of each basic model through the optimization engine in the hyperparameter optimization module.

2. The intelligent prediction system for atomizer power based on multi-model fusion as described in claim 1, characterized in that, In the hyperparameter optimization module, for any given base model: The current hyperparameter combination is obtained by sampling in the preset search space of the base model through the optimization engine; The basic model is trained based on the current hyperparameter combination, and the performance of the basic model is evaluated through cross-validation to obtain the average cross-validation score corresponding to the current hyperparameter combination, and to determine whether the convergence condition is met. If the convergence condition is met, the hyperparameter optimization ends and the optimal hyperparameter combination is obtained. If the convergence condition is not met and the maximum number of iterations has not been reached, sampling continues in the preset search space to obtain a new combination of hyperparameters, and the basic model is trained in the next round based on the new combination of hyperparameters. If the maximum number of iterations is reached and the convergence condition is not met, then the hyperparameter optimization is stopped.

3. The intelligent prediction system for atomizer power based on multi-model fusion as described in claim 1, characterized in that, It also includes a model management module, which includes at least a three-level model degradation mechanism, including a high-precision training model, a cached lightweight model, and an on-demand training backup model.

4. The intelligent prediction system for atomizer power based on multi-model fusion as described in claim 1, characterized in that, It also includes a prediction service module, which provides single prediction services, batch prediction services, and online learning services.

5. The intelligent power prediction system for atomizers based on multi-model fusion as described in claim 1, characterized in that, It also includes an API service module, which provides single prediction interface, batch prediction interface, file upload prediction interface, health check interface, and model information interface.

6. The intelligent prediction system for atomizer power based on multi-model fusion as described in claim 1, characterized in that, The user behavior data includes the average usage time and the number of times the user uses the device within 24 hours; the device characteristic data includes one or more of the device model and flavor type; and the environmental parameter data includes geographical location.

7. The intelligent prediction system for atomizer power based on multi-model fusion as described in claim 1, characterized in that, The various basic models include: LightGBM model, random forest model, ridge regression model, and elastic network model.

8. A method for intelligent prediction of atomizer power based on multi-model fusion, characterized in that, include: Acquire multi-source heterogeneous data from one or more users; the multi-source heterogeneous data includes one or more of the following: user behavior data of the atomizer, device characteristic data of the atomizer, environmental parameter data of the atomizer, and time characteristic data of the atomizer. The multi-source heterogeneous data for each user is preprocessed to obtain the feature vector corresponding to each user. The feature vector corresponding to each user is input into the multi-model fusion framework to obtain the power prediction result corresponding to that user; wherein, the multi-model fusion framework includes multiple basic models; the model parameters and reference weights of each basic model in the multi-model fusion framework are obtained by the optimization engine searching in the preset search space of each basic model.

9. An atomizer system, characterized in that, include: One or more atomizers and a cloud, with each atomizer and the cloud being communicatively connected; The cloud is used to receive multi-source heterogeneous data from each atomizer, obtain a power prediction result for each atomizer based on the multi-source heterogeneous data of each atomizer and the system as described in any one of claims 1-7, and send the power prediction result of each atomizer to the corresponding atomizer; Each atomizer adjusts its power based on the power prediction result received from that atomizer.

10. An atomizer, characterized in that, The atomizer includes: Memory, used to store computer programs; A processor is configured to implement the intelligent prediction method for atomizer power based on multi-model fusion as described in claim 8 by executing the computer program stored in the memory.

Images